Release v0.13.0 (#2093)

This is a bit of an edge case, but I noticed this while working on
something else.

PEFT allows mixed batch adapter inference, i.e. when predicting, the
same batch can use different adapters by passing the adapter_names
argument. However, this is not supported for DoRA (yet), so there is a
check that raises an error if DoRA is used.

Previously, this check would check all adapters for DoRA, even if those
adapters are not being used in adapter_names. This was unnecessarily
strict and with this PR, we only check the adapters that are actually
being used.

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,70 @@
+name: "\U0001F41B Bug Report"
+description: Submit a bug report to help us improve the library
+body:
+  - type: textarea
+    id: system-info
+    attributes:
+      label: System Info
+      description: Please share your relevant system information with us
+      placeholder: peft & accelerate & transformers version, platform, python version, ...
+    validations:
+      required: true
+
+  - type: textarea
+    id: who-can-help
+    attributes:
+      label: Who can help?
+      description: |
+        Your issue will be replied to more quickly if you can figure out the right person to tag with @.
+        If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**.
+
+        All issues are read by one of the core maintainers, so if you don't know who to tag, just leave this blank and
+        a core maintainer will ping the right person.
+
+        Please tag fewer than 3 people.
+
+        Library: @benjaminbossan @sayakpaul
+
+        Documentation: @stevhliu
+
+      placeholder: "@Username ..."
+
+  - type: checkboxes
+    id: information-scripts-examples
+    attributes:
+      label: Information
+      description: 'The problem arises when using:'
+      options:
+        - label: "The official example scripts"
+        - label: "My own modified scripts"
+
+  - type: checkboxes
+    id: information-tasks
+    attributes:
+      label: Tasks
+      description: "The tasks I am working on are:"
+      options:
+        - label: "An officially supported task in the `examples` folder"
+        - label: "My own task or dataset (give details below)"
+
+  - type: textarea
+    id: reproduction
+    validations:
+      required: true
+    attributes:
+      label: Reproduction
+      description: |
+        Please provide a code sample that reproduces the problem you ran into. It can be a Colab link or just a code snippet.
+        Please provide the simplest reproducer as possible so that we can quickly fix the issue. When you paste
+        the error message, please include the full traceback.
+
+      placeholder: |
+        Reproducer:
+
+  - type: textarea
+    id: expected-behavior
+    validations:
+      required: true
+    attributes:
+      label: Expected behavior
+      description: "A clear and concise description of what you would expect to happen."

.github/ISSUE_TEMPLATE/feature-request.yml

@@ -0,0 +1,30 @@
+name: "\U0001F680 Feature request"
+description: Submit a proposal/request for a new feature
+labels: [ "feature" ]
+body:
+  - type: textarea
+    id: feature-request
+    validations:
+      required: true
+    attributes:
+      label: Feature request
+      description: |
+        A clear and concise description of the feature proposal. Please provide a link to the paper and code in case they exist.
+
+  - type: textarea
+    id: motivation
+    validations:
+      required: true
+    attributes:
+      label: Motivation
+      description: |
+        Please outline the motivation for the proposal. Is your feature request related to a problem? 
+
+  - type: textarea
+    id: contribution
+    validations:
+      required: true
+    attributes:
+      label: Your contribution
+      description: |
+        Is there any way that you could help, e.g. by submitting a PR? 

.github/workflows/build_docker_images.yml

@@ -0,0 +1,172 @@
+name: Build Docker images (scheduled)
+
+on:
+  workflow_dispatch:
+  workflow_call:
+  schedule:
+    - cron: "0 1 * * *"
+
+concurrency:
+  group: docker-image-builds
+  cancel-in-progress: false
+
+env:
+  CI_SLACK_CHANNEL: ${{ secrets.CI_DOCKER_CHANNEL }}
+
+jobs:
+  latest-cpu:
+    name: "Latest Peft CPU [dev]"
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      - name: Check out code
+        uses: actions/checkout@v3
+      - name: Login to DockerHub
+        uses: docker/login-action@v2
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push CPU
+        uses: docker/build-push-action@v4
+        with:
+          context: ./docker/peft-cpu
+          push: true
+          tags: huggingface/peft-cpu
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ env.CI_SLACK_CHANNEL }}
+          title: 🤗 Results of the PEFT-CPU docker build
+          status: ${{ job.status }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+  latest-cuda:
+    name: "Latest Peft GPU [dev]"
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      - name: Check out code
+        uses: actions/checkout@v3
+      - name: Login to DockerHub
+        uses: docker/login-action@v1
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v4
+        with:
+          context: ./docker/peft-gpu
+          push: true
+          tags: huggingface/peft-gpu
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ env.CI_SLACK_CHANNEL }}
+          title: 🤗 Results of the PEFT-GPU docker build
+          status: ${{ job.status }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+  latest-cuda-bnb-source:
+    name: "Latest Peft GPU + bnb source [dev]"
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      - name: Check out code
+        uses: actions/checkout@v3
+      - name: Login to DockerHub
+        uses: docker/login-action@v1
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v4
+        with:
+          context: ./docker/peft-gpu-bnb-source
+          push: true
+          tags: huggingface/peft-gpu-bnb-source
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ env.CI_SLACK_CHANNEL }}
+          title: 🤗 Results of the PEFT-GPU (bnb source / HF latest) docker build
+          status: ${{ job.status }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+  latest-cuda-bnb-source-latest:
+    name: "Latest Peft GPU + bnb source [accelerate / peft / transformers latest]"
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      - name: Check out code
+        uses: actions/checkout@v3
+      - name: Login to DockerHub
+        uses: docker/login-action@v1
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v4
+        with:
+          context: ./docker/peft-gpu-bnb-latest
+          push: true
+          tags: huggingface/peft-gpu-bnb-latest
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ env.CI_SLACK_CHANNEL }}
+          title: 🤗 Results of the PEFT-GPU (bnb source / HF source) docker build
+          status: ${{ job.status }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+  latest-cuda-bnb-source-multi:
+    name: "Latest Peft GPU + bnb (multi-backend) source [accelerate / peft / transformers source]"
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      - name: Check out code
+        uses: actions/checkout@v3
+      - name: Login to DockerHub
+        uses: docker/login-action@v1
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v4
+        with:
+          context: ./docker/peft-gpu-bnb-multi-source
+          push: true
+          tags: huggingface/peft-gpu-bnb-multi-source
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ env.CI_SLACK_CHANNEL }}
+          title: 🤗 Results of the PEFT-GPU (bnb source multi-backend / HF latest) docker build
+          status: ${{ job.status }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+

.github/workflows/build_documentation.yml

@@ -0,0 +1,20 @@
+name: Build documentation
+
+on:
+  push:
+    branches:
+      - main
+      - doc-builder*
+      - v*-release
+
+jobs:
+   build:
+    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@main
+    with:
+      commit_sha: ${{ github.sha }}
+      package: peft
+      notebook_folder: peft_docs
+      custom_container: huggingface/transformers-doc-builder
+    secrets:
+      token: ${{ secrets.HUGGINGFACE_PUSH }}
+      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/build_pr_documentation.yml

@@ -0,0 +1,17 @@
+name: Build PR Documentation
+
+on:
+  pull_request:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  build:
+    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@main
+    with:
+      commit_sha: ${{ github.event.pull_request.head.sha }}
+      pr_number: ${{ github.event.number }}
+      package: peft
+      custom_container: huggingface/transformers-doc-builder

.github/workflows/integrations_tests.yml

@@ -0,0 +1,82 @@
+name: integration tests
+
+on:
+  workflow_dispatch:
+    inputs:
+      branch:
+        description: 'Branch to test on'
+        required: true
+
+jobs:
+  run_transformers_integration_tests:
+    strategy:
+      fail-fast: false
+      matrix:
+        transformers-version: ['main', 'latest']
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.inputs.branch }}
+          repository: ${{ github.event.pull_request.head.repo.full_name }}
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.10"
+          cache: "pip"
+          cache-dependency-path: "setup.py"
+      - name: print environment variables
+        run: |
+          echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
+          echo "env.CI_SHA = ${{ env.CI_SHA }}"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install .[test]
+          if [ "${{ matrix.transformers-version }}" == "main" ]; then
+              pip install -U git+https://github.com/huggingface/transformers.git
+          else
+              echo "Nothing to do as transformers latest already installed"
+          fi
+
+      - name: Test transformers integration
+        run: |
+          cd .. && git clone https://github.com/huggingface/transformers.git && cd transformers/ && git rev-parse HEAD
+          RUN_SLOW=1 pytest tests/peft_integration/test_peft_integration.py
+  run_diffusers_integration_tests:
+    strategy:
+      fail-fast: false
+      matrix:
+        # For now diffusers integration is not on PyPI
+        diffusers-version: ['main']
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.inputs.branch }}
+          repository: ${{ github.event.pull_request.head.repo.full_name }}
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.10"
+          cache: "pip"
+          cache-dependency-path: "setup.py"
+      - name: print environment variables
+        run: |
+          echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
+          echo "env.CI_SHA = ${{ env.CI_SHA }}"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install .[test]
+          
+          if [ "${{ matrix.diffusers-version }}" == "main" ]; then
+              pip install -U git+https://github.com/huggingface/diffusers.git
+          else
+              echo "Nothing to do as diffusers latest already installed"
+          fi
+
+      - name: Test diffusers integration
+        run: |
+          cd .. && git clone https://github.com/huggingface/diffusers.git && cd diffusers/ && git rev-parse HEAD
+          pytest tests/lora/test_lora_layers_peft.py

.github/workflows/nightly-bnb.yml

@@ -0,0 +1,249 @@
+name: BNB from source self-hosted runner with slow tests (scheduled)
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "0 2 * * *"
+
+env:
+  RUN_SLOW: "yes"
+  IS_GITHUB_CI: "1"
+  # To be able to run tests on CUDA 12.2
+  NVIDIA_DISABLE_REQUIRE: "1"
+  SLACK_API_TOKEN: ${{ secrets.SLACK_API_TOKEN }}
+
+
+jobs:
+  run_all_tests_single_gpu:
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+          docker-image-name: ["huggingface/peft-gpu-bnb-source:latest", "huggingface/peft-gpu-bnb-latest:latest", "huggingface/peft-gpu-bnb-multi-source:latest"]
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      CUDA_VISIBLE_DEVICES: "0"
+      TEST_TYPE: "single_gpu_${{ matrix.docker-image-name }}"
+    container:
+      image: ${{ matrix.docker-image-name }}
+      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - uses: actions/checkout@v3
+      - name: Pip install
+        run: |
+          source activate peft
+          pip install -e . --no-deps
+          pip install pytest-reportlog pytest-cov parameterized datasets scipy einops
+          pip install "pytest>=7.2.0,<8.0.0" # see: https://github.com/huggingface/transformers/blob/ce4fff0be7f6464d713f7ac3e0bbaafbc6959ae5/setup.py#L148C6-L148C26
+          mkdir transformers-clone && git clone https://github.com/huggingface/transformers.git transformers-clone # rename to transformers clone to avoid modules conflict
+          if [ "${{ matrix.docker-image-name }}" == "huggingface/peft-gpu-bnb-latest:latest" ]; then
+            cd transformers-clone
+            transformers_version=$(pip show transformers | grep '^Version:' | cut -d ' ' -f2 | sed 's/\.dev0//')
+            echo "Checking out tag for Transformers version: v$transformers_version"
+            git fetch --tags
+            git checkout tags/v$transformers_version
+            cd ..
+          fi
+
+      - name: Test bnb import
+        id: import
+        if: always()
+        run: |
+          source activate peft
+          python3 -m bitsandbytes
+          python3 -c "import bitsandbytes as bnb"
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ secrets.BNB_SLACK_CHANNEL_ID }}
+          title: 🤗 Results of bitsandbytes import
+          status: ${{ steps.import.outcome }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+      - name: Run examples on single GPU
+        id: examples_tests
+        if: always()
+        run: |
+          source activate peft
+          make tests_examples_single_gpu_bnb
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ secrets.BNB_SLACK_CHANNEL_ID }}
+          title: 🤗 Results of bitsandbytes examples tests - single GPU
+          status: ${{ steps.examples_tests.outcome }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+      - name: Run core tests on single GPU
+        id: core_tests
+        if: always()
+        run: |
+          source activate peft
+          make tests_core_single_gpu_bnb
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ secrets.BNB_SLACK_CHANNEL_ID }}
+          title: 🤗 Results of bitsandbytes core tests - single GPU
+          status: ${{ steps.core_tests.outcome }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+      # TODO: this is a test to see if BNB multi-backend single-GPU tests succeed w/o regression tests
+      # - name: Run BNB regression tests on single GPU
+      #   id: regression_tests
+      #   if: always()
+      #   run: |
+      #     source activate peft
+      #     make tests_gpu_bnb_regression
+
+      # - name: Post to Slack
+      #   if: always()
+      #   uses: huggingface/hf-workflows/.github/actions/post-slack@main
+      #   with:
+      #     slack_channel: ${{ secrets.BNB_SLACK_CHANNEL_ID }}
+      #     title: 🤗 Results of bitsandbytes regression tests - single GPU
+      #     status: ${{ steps.regression_tests.outcome }}
+      #     slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+      - name: Run transformers tests on single GPU
+        id: transformers_tests
+        if: always()
+        run: |
+          source activate peft
+          make transformers_tests
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ secrets.BNB_SLACK_CHANNEL_ID }}
+          title: 🤗 Results of bitsandbytes transformers tests - single GPU
+          status: ${{ steps.transformers_tests.outcome }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+      - name: Generate Report
+        if: always()
+        run: |
+          pip install slack_sdk tabulate
+          python scripts/log_reports.py --slack_channel_name bnb-daily-ci-collab >> $GITHUB_STEP_SUMMARY
+
+  run_all_tests_multi_gpu:
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        docker-image-name: ["huggingface/peft-gpu-bnb-source:latest", "huggingface/peft-gpu-bnb-latest:latest", "huggingface/peft-gpu-bnb-multi-source:latest"]
+    runs-on:
+      group: aws-g6-12xlarge-plus
+    env:
+      CUDA_VISIBLE_DEVICES: "0,1"
+      TEST_TYPE: "multi_gpu_${{ matrix.docker-image-name }}"
+    container:
+      image: ${{ matrix.docker-image-name }}
+      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - uses: actions/checkout@v3
+      - name: Pip install
+        run: |
+          source activate peft
+          pip install -e . --no-deps
+          pip install pytest-reportlog pytest-cov parameterized datasets scipy einops
+          pip install "pytest>=7.2.0,<8.0.0" # see: https://github.com/huggingface/transformers/blob/ce4fff0be7f6464d713f7ac3e0bbaafbc6959ae5/setup.py#L148C6-L148C26
+          mkdir transformers-clone && git clone https://github.com/huggingface/transformers.git transformers-clone
+          if [ "${{ matrix.docker-image-name }}" == "huggingface/peft-gpu-bnb-latest:latest" ]; then
+            cd transformers-clone
+            transformers_version=$(pip show transformers | grep '^Version:' | cut -d ' ' -f2 | sed 's/\.dev0//')
+            echo "Checking out tag for Transformers version: v$transformers_version"
+            git fetch --tags
+            git checkout tags/v$transformers_version
+            cd ..
+          fi
+
+      - name: Test bnb import
+        id: import
+        if: always()
+        run: |
+          source activate peft
+          python3 -m bitsandbytes
+          python3 -c "import bitsandbytes as bnb"
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ secrets.BNB_SLACK_CHANNEL_ID }}
+          title: 🤗 Results of bitsandbytes import
+          status: ${{ steps.import.outcome }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+      - name: Run core GPU tests on multi-gpu
+        if: always()
+        run: |
+          source activate peft
+
+      - name: Run examples on multi GPU
+        id: examples_tests
+        if: always()
+        run: |
+          source activate peft
+          make tests_examples_multi_gpu_bnb
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ secrets.BNB_SLACK_CHANNEL_ID }}
+          title: 🤗 Results of bitsandbytes examples tests - multi GPU
+          status: ${{ steps.examples_tests.outcome }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+      - name: Run core tests on multi GPU
+        id: core_tests
+        if: always()
+        run: |
+          source activate peft
+          make tests_core_multi_gpu_bnb
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ secrets.BNB_SLACK_CHANNEL_ID }}
+          title: 🤗 Results of bitsandbytes core tests - multi GPU
+          status: ${{ steps.core_tests.outcome }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+      - name: Run transformers tests on multi GPU
+        id: transformers_tests
+        if: always()
+        run: |
+          source activate peft
+          make transformers_tests
+
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ secrets.BNB_SLACK_CHANNEL_ID }}
+          title: 🤗 Results of bitsandbytes transformers tests - multi GPU
+          status: ${{ steps.transformers_tests.outcome }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+      - name: Generate Report
+        if: always()
+        run: |
+          pip install slack_sdk tabulate
+          python scripts/log_reports.py --slack_channel_name bnb-daily-ci-collab >> $GITHUB_STEP_SUMMARY

.github/workflows/nightly.yml

@@ -0,0 +1,110 @@
+name: Self-hosted runner with slow tests (scheduled)
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "0 2 * * *"
+
+env:
+  RUN_SLOW: "yes"
+  IS_GITHUB_CI: "1"
+  # To be able to run tests on CUDA 12.2
+  NVIDIA_DISABLE_REQUIRE: "1"
+  SLACK_API_TOKEN: ${{ secrets.SLACK_API_TOKEN }}
+
+
+jobs:
+  run_all_tests_single_gpu:
+    strategy:
+      fail-fast: false
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      CUDA_VISIBLE_DEVICES: "0"
+      TEST_TYPE: "single_gpu"
+    container:
+      image: huggingface/peft-gpu:latest
+      options: --gpus all --shm-size "16gb" -e NVIDIA_DISABLE_REQUIRE=true
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - uses: actions/checkout@v3
+      - name: Pip install
+        run: |
+          source activate peft
+          pip install -e . --no-deps
+          pip install pytest-reportlog
+
+      - name: Run common tests on single GPU
+        run: |
+          source activate peft
+          make tests_common_gpu
+
+      - name: Run examples on single GPU
+        run: |
+          source activate peft
+          make tests_examples_single_gpu
+
+      - name: Run core tests on single GPU
+        run: |
+          source activate peft
+          make tests_core_single_gpu
+
+      - name: Run regression tests on single GPU
+        run: |
+          source activate peft
+          make tests_regression
+
+      - name: Generate Report
+        if: always()
+        run: |
+          pip install slack_sdk tabulate
+          python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY
+
+  run_all_tests_multi_gpu:
+    strategy:
+      fail-fast: false
+    runs-on:
+      group: aws-g6-12xlarge-plus
+    env:
+      CUDA_VISIBLE_DEVICES: "0,1"
+      TEST_TYPE: "multi_gpu"
+    container:
+      image: huggingface/peft-gpu:latest
+      options: --gpus all --shm-size "16gb" -e NVIDIA_DISABLE_REQUIRE=true
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - uses: actions/checkout@v3
+      - name: Pip install
+        run: |
+          source activate peft
+          pip install -e . --no-deps
+          pip install pytest-reportlog
+
+      - name: Run core GPU tests on multi-gpu
+        run: |
+          source activate peft
+
+      - name: Run common tests on multi GPU
+        run: |
+          source activate peft
+          make tests_common_gpu
+
+      - name: Run examples on multi GPU
+        run: |
+          source activate peft
+          make tests_examples_multi_gpu
+
+      - name: Run core tests on multi GPU
+        run: |
+          source activate peft
+          make tests_core_multi_gpu
+
+      - name: Generate Report
+        if: always()
+        run: |
+          pip install slack_sdk tabulate
+          python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY

.github/workflows/stale.yml

@@ -0,0 +1,30 @@
+name: Stale Bot
+
+on:
+  schedule:
+    - cron: "0 15 * * *"
+
+jobs:
+  close_stale_issues:
+    name: Close Stale Issues
+    if: github.repository == 'huggingface/peft'
+    runs-on: ubuntu-latest
+    permissions:
+      issues: write
+      pull-requests: write
+    env:
+      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+    steps:
+    - uses: actions/checkout@v3
+
+    - name: Setup Python
+      uses: actions/setup-python@v4
+      with:
+        python-version: 3.8
+
+    - name: Install requirements
+      run: |
+        pip install PyGithub
+    - name: Close stale issues
+      run: |
+        python scripts/stale.py

.github/workflows/test-docker-build.yml

@@ -0,0 +1,59 @@
+name: Test Docker images (on PR)
+
+on:
+  pull_request:
+    paths:
+      # Run only when DockerFile files are modified
+      - "docker/**"
+jobs:
+  get_changed_files:
+    name: "Build all modified docker images"
+    runs-on: ubuntu-latest
+    outputs:
+      matrix: ${{ steps.set-matrix.outputs.matrix }}
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v3
+      - name: Get changed files
+        id: changed-files
+        uses: tj-actions/changed-files@1c8e6069583811afb28f97afeaf8e7da80c6be5c #v42
+        with:
+          files: docker/**
+          json: "true"
+      - name: Run step if only the files listed above change
+        if: steps.changed-files.outputs.any_changed == 'true'
+        id: set-matrix
+        env:
+          ALL_CHANGED_FILES: ${{ steps.changed-files.outputs.all_changed_files }}
+        run: |
+          echo "matrix=${{ steps.changed-files.outputs.all_changed_files}}" >> $GITHUB_OUTPUT
+  build_modified_files:
+    needs: get_changed_files
+    name: Build Docker images on modified files
+    runs-on: ubuntu-latest
+    if: ${{ needs.get_changed_files.outputs.matrix }} != ''
+    strategy:
+      fail-fast: false
+      matrix:
+        docker-file: ${{ fromJson(needs.get_changed_files.outputs.matrix) }}
+    steps:
+      - name: Cleanup disk
+        run: |
+          sudo ls -l /usr/local/lib/
+          sudo ls -l /usr/share/
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /usr/share/dotnet
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      - name: Check out code
+        uses: actions/checkout@v3
+      - name: Build Docker image
+        uses: docker/build-push-action@v4
+        with:
+          file: ${{ matrix.docker-file }}
+          context: .
+          push: False

.github/workflows/tests-main.yml

@@ -0,0 +1,36 @@
+name: tests on transformers main
+
+on:
+  push:
+    branches: [main]
+    paths-ignore:
+        - 'docs/**'
+
+jobs:
+  tests:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python 3.11
+        uses: actions/setup-python@v4
+        with:
+          python-version: 3.11
+          cache: "pip"
+          cache-dependency-path: "setup.py"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          # cpu version of pytorch
+          pip install -U git+https://github.com/huggingface/transformers.git
+          pip install -e .[test]
+      - name: Test with pytest
+        run: |
+          make test
+      - name: Post to Slack
+        if: always()
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: ${{ secrets.SLACK_CHANNEL_ID }}
+          title: 🤗 Results of transformers main tests
+          status: ${{ job.status }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}

.github/workflows/tests.yml

@@ -0,0 +1,61 @@
+name: tests
+
+on:
+  push:
+    branches: [main]
+    paths-ignore:
+      - 'docs/**'
+  pull_request:
+    paths-ignore:
+      - 'docs/**'
+
+jobs:
+  check_code_quality:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.8"
+          cache: "pip"
+          cache-dependency-path: "setup.py"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install .[dev]
+      - name: Check quality
+        run: |
+          make quality
+
+  tests:
+    needs: check_code_quality
+    strategy:
+      # TODO: remove 'fail-fast' line once timeout issue from the Hub is solved
+      fail-fast: false
+      matrix:
+        python-version: ["3.8", "3.9", "3.10", "3.11"]
+        os: ["ubuntu-latest", "macos-12", "windows-latest"]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v4
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: "pip"
+          cache-dependency-path: "setup.py"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          # cpu version of pytorch
+          pip install -e .[test]
+      - name: Downgrade numpy on MacOS and Windows
+        # TODO: remove numpy downgrade on MacOS & Windows once torch fixes numpy 2.0 issue
+        shell: bash
+        if: matrix.os == 'windows-latest' || matrix.os == 'macos-12'
+        run: |
+          pip install --force-reinstall -U "numpy<2.0.0"
+      - name: Test with pytest
+        run: |
+          make test

.github/workflows/torch_compile_tests.yml

@@ -0,0 +1,52 @@
+name: torch compile tests
+
+on:
+  workflow_dispatch:
+    inputs:
+      branch:
+        description: 'Branch to test on'
+        required: true
+      pytorch_nightly:
+        description: 'Whether to use PyTorch nightly (true/false)'
+        required: false
+        default: false
+
+env:
+  RUN_SLOW: "yes"
+  IS_GITHUB_CI: "1"
+  # To be able to run tests on CUDA 12.2
+  NVIDIA_DISABLE_REQUIRE: "1"
+
+jobs:
+  run_tests_with_compile:
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      PEFT_DEBUG_WITH_TORCH_COMPILE: 1
+      CUDA_VISIBLE_DEVICES: "0"
+      TEST_TYPE: "single_gpu_huggingface/peft-gpu-bnb-latest:latest"
+    container:
+      image: "huggingface/peft-gpu-bnb-latest:latest"
+      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.inputs.branch }}
+          repository: ${{ github.event.pull_request.head.repo.full_name }}
+      - name: Pip install
+        run: |
+          source activate peft
+          pip install -e . --no-deps
+          pip install pytest-cov pytest-reportlog parameterized datasets scipy einops
+          pip install "pytest>=7.2.0,<8.0.0" # see: https://github.com/huggingface/transformers/blob/ce4fff0be7f6464d713f7ac3e0bbaafbc6959ae5/setup.py#L148C6-L148C26
+          if [ "${{ github.event.inputs.pytorch_nightly }}" = "true" ]; then
+            python -m pip install --upgrade --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
+          fi
+      - name: Test compile with pytest
+        run: |
+          source activate peft
+          echo "PEFT_DEBUG_WITH_TORCH_COMPILE=$PEFT_DEBUG_WITH_TORCH_COMPILE"
+          make tests_torch_compile

.github/workflows/trufflehog.yml

@@ -0,0 +1,15 @@
+on:
+  push:
+
+name: Secret Leaks
+
+jobs:
+  trufflehog:
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+      with:
+        fetch-depth: 0
+    - name: Secret Scanning
+      uses: trufflesecurity/trufflehog@main

.github/workflows/upload_pr_documentation.yml

@@ -0,0 +1,16 @@
+name: Upload PR Documentation
+
+on:
+  workflow_run:
+    workflows: ["Build PR Documentation"]
+    types:
+      - completed
+
+jobs:
+  build:
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@main
+    with:
+      package_name: peft
+    secrets:
+      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}
+      comment_bot_token: ${{ secrets.COMMENT_BOT_TOKEN }}

.gitignore

@@ -138,4 +138,4 @@ dmypy.json
 .DS_Store
 
 # More test things
-wandb
+wandb

.pre-commit-config.yaml

@@ -0,0 +1,13 @@
+repos:
+  - repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.6.1
+    hooks:
+      - id: ruff
+        args:
+          - --fix
+      - id: ruff-format
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v4.6.0
+    hooks:
+      - id: check-merge-conflict
+      - id: check-yaml

MANIFEST.in

@@ -1 +0,0 @@
-include LICENSE

Makefile

@@ -1,19 +1,61 @@
 .PHONY: quality style test docs
 
-check_dirs := src examples
+check_dirs := src tests examples docs scripts docker
 
 # Check that source code meets quality standards
 
 # this target runs checks on all files
 quality:
-	black --check $(check_dirs)
-	isort --check-only $(check_dirs)
-	flake8 $(check_dirs)
-	doc-builder style src --max_len 119 --check_only
+	ruff check $(check_dirs)
+	ruff format --check $(check_dirs)
+	doc-builder style src/peft tests docs/source --max_len 119 --check_only
 
 # Format source code automatically and check is there are any problems left that need manual fixing
 style:
-	black $(check_dirs)
-	isort $(check_dirs)
-	doc-builder style src --max_len 119
-	
+	ruff check --fix $(check_dirs)
+	ruff format $(check_dirs)
+	doc-builder style src/peft tests docs/source --max_len 119
+
+test:
+	python -m pytest -n 3 tests/ $(if $(IS_GITHUB_CI),--report-log "ci_tests.log",)
+
+tests_examples_multi_gpu:
+	python -m pytest -m multi_gpu_tests tests/test_gpu_examples.py $(if $(IS_GITHUB_CI),--report-log "multi_gpu_examples.log",)
+
+tests_examples_single_gpu:
+	python -m pytest -m single_gpu_tests tests/test_gpu_examples.py $(if $(IS_GITHUB_CI),--report-log "single_gpu_examples.log",)
+
+tests_core_multi_gpu:
+	python -m pytest -m multi_gpu_tests tests/test_common_gpu.py $(if $(IS_GITHUB_CI),--report-log "core_multi_gpu.log",)
+
+tests_core_single_gpu:
+	python -m pytest -m single_gpu_tests tests/test_common_gpu.py $(if $(IS_GITHUB_CI),--report-log "core_single_gpu.log",)
+
+tests_common_gpu:
+	python -m pytest tests/test_decoder_models.py $(if $(IS_GITHUB_CI),--report-log "common_decoder.log",)
+	python -m pytest tests/test_encoder_decoder_models.py $(if $(IS_GITHUB_CI),--report-log "common_encoder_decoder.log",)
+
+tests_examples_multi_gpu_bnb:
+	python -m pytest -m "multi_gpu_tests and bitsandbytes" tests/test_gpu_examples.py $(if $(IS_GITHUB_CI),--report-log "multi_gpu_examples.log",)
+
+tests_examples_single_gpu_bnb:
+	python -m pytest -m "single_gpu_tests and bitsandbytes" tests/test_gpu_examples.py $(if $(IS_GITHUB_CI),--report-log "single_gpu_examples.log",)
+
+tests_core_multi_gpu_bnb:
+	python -m pytest -m "multi_gpu_tests and bitsandbytes" tests/test_common_gpu.py $(if $(IS_GITHUB_CI),--report-log "core_multi_gpu.log",)
+
+tests_core_single_gpu_bnb:
+	python -m pytest -m "single_gpu_tests and bitsandbytes" tests/test_common_gpu.py $(if $(IS_GITHUB_CI),--report-log "core_single_gpu.log",)
+
+tests_gpu_bnb_regression:
+	python -m pytest tests/bnb/test_bnb_regression.py $(if $(IS_GITHUB_CI),--report-log "bnb_regression_gpu.log",)
+
+# For testing transformers tests for bnb runners
+transformers_tests:
+	RUN_SLOW=1 python -m pytest transformers-clone/tests/quantization/bnb $(if $(IS_GITHUB_CI),--report-log "transformers_tests.log",)
+
+tests_regression:
+	python -m pytest -s --regression tests/regression/ $(if $(IS_GITHUB_CI),--report-log "regression_tests.log",)
+
+tests_torch_compile:
+	python -m pytest tests/test_torch_compile.py $(if $(IS_GITHUB_CI),--report-log "compile_tests.log",)

README.md

@@ -19,43 +19,67 @@ limitations under the License.
     <p>State-of-the-art Parameter-Efficient Fine-Tuning (PEFT) methods</p>
 </h3>
 
-Parameter-Efficient Fine-Tuning (PEFT) methods enable efficient adaptation of pre-trained language models (PLMs) to various downstream applications without fine-tuning all the model's parameters. Fine-tuning large-scale PLMs is often prohibitively costly. In this regard, PEFT methods only fine-tune a small number of (extra) model parameters, thereby greatly decreasing the computational and storage costs. Recent State-of-the-Art PEFT techniques achieve performance comparable to that of full fine-tuning. 
+Fine-tuning large pretrained models is often prohibitively costly due to their scale. Parameter-Efficient Fine-Tuning (PEFT) methods enable efficient adaptation of large pretrained models to various downstream applications by only fine-tuning a small number of (extra) model parameters instead of all the model's parameters. This significantly decreases the computational and storage costs. Recent state-of-the-art PEFT techniques achieve performance comparable to fully fine-tuned models.
 
-Seamlessly integrated with 🤗 Accelerate for large scale models leveraging PyTorch FSDP. 
+PEFT is integrated with Transformers for easy model training and inference, Diffusers for conveniently managing different adapters, and Accelerate for distributed training and inference for really big models.
 
-Supported methods:
+> [!TIP]
+> Visit the [PEFT](https://huggingface.co/PEFT) organization to read about the PEFT methods implemented in the library and to see notebooks demonstrating how to apply these methods to a variety of downstream tasks. Click the "Watch repos" button on the organization page to be notified of newly implemented methods and notebooks!
 
-1. LoRA: [LORA: LOW-RANK ADAPTATION OF LARGE LANGUAGE MODELS](https://arxiv.org/pdf/2106.09685.pdf)
-2. Prefix Tuning: [P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks](https://arxiv.org/pdf/2110.07602.pdf)
-3. P-Tuning: [GPT Understands, Too](https://arxiv.org/pdf/2103.10385.pdf)
-4. Prompt Tuning: [The Power of Scale for Parameter-Efficient Prompt Tuning](https://arxiv.org/pdf/2104.08691.pdf) 
+Check the PEFT Adapters API Reference section for a list of supported PEFT methods, and read the [Adapters](https://huggingface.co/docs/peft/en/conceptual_guides/adapter), [Soft prompts](https://huggingface.co/docs/peft/en/conceptual_guides/prompting), and [IA3](https://huggingface.co/docs/peft/en/conceptual_guides/ia3) conceptual guides to learn more about how these methods work.
 
-## Getting started
+## Quickstart
+
+Install PEFT from pip:
+
+```bash
+pip install peft
+```
+
+Prepare a model for training with a PEFT method such as LoRA by wrapping the base model and PEFT configuration with `get_peft_model`. For the bigscience/mt0-large model, you're only training 0.19% of the parameters!
 
 ```python
 from transformers import AutoModelForSeq2SeqLM
-from peft import get_peft_config, get_peft_model, LoRAConfig, TaskType
+from peft import get_peft_config, get_peft_model, LoraConfig, TaskType
 model_name_or_path = "bigscience/mt0-large"
 tokenizer_name_or_path = "bigscience/mt0-large"
 
-peft_config = LoRAConfig(
+peft_config = LoraConfig(
     task_type=TaskType.SEQ_2_SEQ_LM, inference_mode=False, r=8, lora_alpha=32, lora_dropout=0.1
 )
 
 model = AutoModelForSeq2SeqLM.from_pretrained(model_name_or_path)
 model = get_peft_model(model, peft_config)
 model.print_trainable_parameters()
-# output: trainable params: 2359296 || all params: 1231940608 || trainable%: 0.19151053100118282
+"trainable params: 2359296 || all params: 1231940608 || trainable%: 0.19151053100118282"
 ```
 
-## Use Cases
+To load a PEFT model for inference:
 
-### Get comparable performance to full finetuning by adapting LLMs to downstream tasks using consumer hardware
+```py
+from peft import AutoPeftModelForCausalLM
+from transformers import AutoTokenizer
+import torch
 
-GPU memory required for adapting LLMs on the few-shot dataset `ought/raft/twitter_complaints`. Here, settings considered
-are full finetuning, PEFT-LoRA using plain PyTorch and  PEFT-LoRA using DeepSpeed with CPU Offloading. 
+model = AutoPeftModelForCausalLM.from_pretrained("ybelkada/opt-350m-lora").to("cuda")
+tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
 
-Hardware: Single A100 80GB GPU with CPU RAM above 64GB
+model.eval()
+inputs = tokenizer("Preheat the oven to 350 degrees and place the cookie dough", return_tensors="pt")
+
+outputs = model.generate(input_ids=inputs["input_ids"].to("cuda"), max_new_tokens=50)
+print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
+
+"Preheat the oven to 350 degrees and place the cookie dough in the center of the oven. In a large bowl, combine the flour, baking powder, baking soda, salt, and cinnamon. In a separate bowl, combine the egg yolks, sugar, and vanilla."
+```
+
+## Why you should use PEFT
+
+There are many benefits of using PEFT but the main one is the huge savings in compute and storage, making PEFT applicable to many different use cases.
+
+### High performance on consumer hardware
+
+Consider the memory requirements for training the following models on the [ought/raft/twitter_complaints](https://huggingface.co/datasets/ought/raft/viewer/twitter_complaints) dataset with an A100 80GB GPU with more than 64GB of CPU RAM.
 
 |   Model         | Full Finetuning | PEFT-LoRA PyTorch  | PEFT-LoRA DeepSpeed with CPU Offloading |
 | --------- | ---- | ---- | ---- |
@@ -63,9 +87,7 @@ Hardware: Single A100 80GB GPU with CPU RAM above 64GB
 | bigscience/mt0-xxl (12B params) | OOM GPU | 56GB GPU / 3GB CPU | 22GB GPU / 52GB CPU |
 | bigscience/bloomz-7b1 (7B params) | OOM GPU | 32GB GPU / 3.8GB CPU | 18.1GB GPU / 35GB CPU |
 
-Performance of PEFT-LoRA tuned `bigscience/T0_3B` on `ought/raft/twitter_complaints` leaderboard. 
-A point to note is that we didn't try to sequeeze performance by playing around with input instruction templates, LoRA hyperparams and other training related hyperparams. Also, we didn't use the larger 13B mt0-xxl model.
-So, we are already seeing comparable performance to SoTA with parameter effcient tuning. Also, the final checkpoint size is just `19MB` in comparison to `11GB` size of the backbone `bigscience/T0_3B` model.
+With LoRA you can fully finetune a 12B parameter model that would've otherwise run out of memory on the 80GB GPU, and comfortably fit and train a 3B parameter model. When you look at the 3B parameter model's performance, it is comparable to a fully finetuned model at a fraction of the GPU memory.
 
 |   Submission Name        | Accuracy |
 | --------- | ---- |
@@ -73,258 +95,64 @@ So, we are already seeing comparable performance to SoTA with parameter effcient
 | Flan-T5 | 0.892 |
 | lora-t0-3b | 0.863 |
 
-**Therefore, we can see that performance comparable to SoTA is achievable by PEFT methods with consumer hardware such as 16GB and 24GB GPUs.**
+> [!TIP]
+> The bigscience/T0_3B model performance isn't optimized in the table above. You can squeeze even more performance out of it by playing around with the input instruction templates, LoRA hyperparameters, and other training related hyperparameters. The final checkpoint size of this model is just 19MB compared to 11GB of the full bigscience/T0_3B model. Learn more about the advantages of finetuning with PEFT in this [blog post](https://www.philschmid.de/fine-tune-flan-t5-peft).
 
-### Parameter Efficient Tuning of Diffusion Models
+### Quantization
 
-GPU memory required by different settings during training are given below. The final checkpoint size being `8.8 MB`.
+Quantization is another method for reducing the memory requirements of a model by representing the data in a lower precision. It can be combined with PEFT methods to make it even easier to train and load LLMs for inference.
 
-Hardware: Single A100 80GB GPU with CPU RAM above 64G
+* Learn how to finetune [meta-llama/Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf) with QLoRA and the [TRL](https://huggingface.co/docs/trl/index) library on a 16GB GPU in the [Finetune LLMs on your own consumer hardware using tools from PyTorch and Hugging Face ecosystem](https://pytorch.org/blog/finetune-llms/) blog post.
+* Learn how to finetune a [openai/whisper-large-v2](https://huggingface.co/openai/whisper-large-v2) model for multilingual automatic speech recognition with LoRA and 8-bit quantization in this [notebook](https://colab.research.google.com/drive/1DOkD_5OUjFa0r5Ik3SgywJLJtEo2qLxO?usp=sharing) (see this [notebook](https://colab.research.google.com/drive/1vhF8yueFqha3Y3CpTHN6q9EVcII9EYzs?usp=sharing) instead for an example of streaming a dataset).
 
-|   Model         | Full Finetuning | PEFT-LoRA  | PEFT-LoRA with Gradient Checkpoitning  |
+### Save compute and storage
+
+PEFT can help you save storage by avoiding full finetuning of models on each of downstream task or dataset. In many cases, you're only finetuning a very small fraction of a model's parameters and each checkpoint is only a few MBs in size (instead of GBs). These smaller PEFT adapters demonstrate performance comparable to a fully finetuned model. If you have many datasets, you can save a lot of storage with a PEFT model and not have to worry about catastrophic forgetting or overfitting the backbone or base model.
+
+## PEFT integrations
+
+PEFT is widely supported across the Hugging Face ecosystem because of the massive efficiency it brings to training and inference.
+
+### Diffusers
+
+The iterative diffusion process consumes a lot of memory which can make it difficult to train. PEFT can help reduce the memory requirements and reduce the storage size of the final model checkpoint. For example, consider the memory required for training a Stable Diffusion model with LoRA on an A100 80GB GPU with more than 64GB of CPU RAM. The final model checkpoint size is only 8.8MB!
+
+|   Model         | Full Finetuning | PEFT-LoRA  | PEFT-LoRA with Gradient Checkpointing  |
 | --------- | ---- | ---- | ---- |
 | CompVis/stable-diffusion-v1-4 | 27.5GB GPU / 3.97GB CPU | 15.5GB GPU / 3.84GB CPU | 8.12GB GPU / 3.77GB CPU | 
 
+> [!TIP]
+> Take a look at the [examples/lora_dreambooth/train_dreambooth.py](examples/lora_dreambooth/train_dreambooth.py) training script to try training your own Stable Diffusion model with LoRA, and play around with the [smangrul/peft-lora-sd-dreambooth](https://huggingface.co/spaces/smangrul/peft-lora-sd-dreambooth) Space which is running on a T4 instance. Learn more about the PEFT integration in Diffusers in this [tutorial](https://huggingface.co/docs/peft/main/en/tutorial/peft_integrations#diffusers).
 
-**Training**
-An example of using LoRA for parameter efficient dreambooth training is given in `~examples/lora_dreambooth/train_dreambooth.py`
+### Accelerate
 
-```bash
-export MODEL_NAME= "CompVis/stable-diffusion-v1-4" #"stabilityai/stable-diffusion-2-1"
-export INSTANCE_DIR="path-to-instance-images"
-export CLASS_DIR="path-to-class-images"
-export OUTPUT_DIR="path-to-save-model"
+[Accelerate](https://huggingface.co/docs/accelerate/index) is a library for distributed training and inference on various training setups and hardware (GPUs, TPUs, Apple Silicon, etc.). PEFT models work with Accelerate out of the box, making it really convenient to train really large models or use them for inference on consumer hardware with limited resources.
 
-accelerate launch train_dreambooth.py \
-  --pretrained_model_name_or_path=$MODEL_NAME  \
-  --instance_data_dir=$INSTANCE_DIR \
-  --class_data_dir=$CLASS_DIR \
-  --output_dir=$OUTPUT_DIR \
-  --train_text_encoder \
-  --with_prior_preservation --prior_loss_weight=1.0 \
-  --instance_prompt="a photo of sks dog" \
-  --class_prompt="a photo of dog" \
-  --resolution=512 \
-  --train_batch_size=1 \
-  --lr_scheduler="constant" \
-  --lr_warmup_steps=0 \
-  --num_class_images=200 \
-  --use_lora \
-  --lora_r 16 \
-  --lora_alpha 27 \
-  --lora_text_encoder_r 16 \
-  --lora_text_encoder_alpha 17 \
-  --learning_rate=1e-4 \
-  --gradient_accumulation_steps=1 \
-  --gradient_checkpointing \
-  --max_train_steps=800
-```
+### TRL
 
-Try out the 🤗 Gradio Space which should run seamlessly on a T4 instance:
-[smangrul/peft-lora-sd-dreambooth](https://huggingface.co/spaces/smangrul/peft-lora-sd-dreambooth).
+PEFT can also be applied to training LLMs with RLHF components such as the ranker and policy. Get started by reading:
 
-![peft lora dreambooth gradio space](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/peft_lora_dreambooth_gradio_space.png)
+* [Fine-tune a Mistral-7b model with Direct Preference Optimization](https://towardsdatascience.com/fine-tune-a-mistral-7b-model-with-direct-preference-optimization-708042745aac) with PEFT and the [TRL](https://huggingface.co/docs/trl/index) library to learn more about the Direct Preference Optimization (DPO) method and how to apply it to a LLM.
+* [Fine-tuning 20B LLMs with RLHF on a 24GB consumer GPU](https://huggingface.co/blog/trl-peft) with PEFT and the [TRL](https://huggingface.co/docs/trl/index) library, and then try out the [gpt2-sentiment_peft.ipynb](https://github.com/huggingface/trl/blob/main/examples/notebooks/gpt2-sentiment.ipynb) notebook to optimize GPT2 to generate positive movie reviews.
+* [StackLLaMA: A hands-on guide to train LLaMA with RLHF](https://huggingface.co/blog/stackllama) with PEFT, and then try out the [stack_llama/scripts](https://github.com/huggingface/trl/tree/main/examples/research_projects/stack_llama/scripts) for supervised finetuning, reward modeling, and RL finetuning.
 
-### Parameter Efficient Tuning of LLMs for RLHF components such as Ranker and Policy [ToDo]
+## Model support
 
-### Save compute and storage even for medium and small models
+Use this [Space](https://stevhliu-peft-methods.hf.space) or check out the [docs](https://huggingface.co/docs/peft/main/en/index) to find which models officially support a PEFT method out of the box. Even if you don't see a model listed below, you can manually configure the model config to enable PEFT for a model. Read the [New transformers architecture](https://huggingface.co/docs/peft/main/en/developer_guides/custom_models#new-transformers-architectures) guide to learn how.
 
-Save storage by avoiding full finetuning of models on each of the downstream tasks/datasets,
-With PEFT methods, users only need to store tiny checkpoints in the order of `MBs` all the while retaining 
-performance comparable to full finetuning.
+## Contribute
 
-An example of using LoRA for the task of adaping `LayoutLMForTokenClassification` on `FUNSD` dataset is given in `~examples/token_classification/PEFT_LoRA_LayoutLMForTokenClassification_on_FUNSD.py`. We can observe that with only `0.62 %` of parameters being trainable, we achieve performance (F1 0.777) comparable to full finetuning (F1 0.786) (without any hyerparam tuning runs for extracting more performance), and the checkpoint of this is only `2.8MB`. Now, if there are `N` such datasets, just have these PEFT models one for each dataset and save a lot of storage without having to worry about the problem of catastrophic forgetting or overfitting of backbone/base model.
-
-Another example is fine-tuning `roberta-large` on `MRPC` GLUE dataset suing differenct PEFT methods. The notebooks are given in `~examples/sequence_classification`. 
-
-
-## PEFT + 🤗 Accelerate
-
-PEFT models work with 🤗 Accelerate out of the box. Use 🤗 Accelerate for Distributed training on various hardware such as GPUs, Apple Silicon devices etc during training.
-Use 🤗 Accelerate for inferencing on consumer hardware with small resources.
-
-### Example of PEFT model training using 🤗 Accelerate's DeepSpeed integation
-
- Currently DeepSpeed requires PR [ZeRO3 handling frozen weights](https://github.com/microsoft/DeepSpeed/pull/2653) to fix [[REQUEST] efficiently deal with frozen weights during training](https://github.com/microsoft/DeepSpeed/issues/2615) issue. Example is provided in `~examples/conditional_generation/peft_lora_seq2seq_accelerate_ds_zero3_offload.py`. 
-  a. First run `accelerate config --config_file ds_zero3_cpu.yaml` and answer the questionaire. 
-  Below are the contents of the config file.
-  ```
-  compute_environment: LOCAL_MACHINE
-  deepspeed_config:
-    gradient_accumulation_steps: 1
-    gradient_clipping: 1.0
-    offload_optimizer_device: cpu
-    offload_param_device: cpu
-    zero3_init_flag: true
-    zero3_save_16bit_model: true
-    zero_stage: 3
-  distributed_type: DEEPSPEED
-  downcast_bf16: 'no'
-  dynamo_backend: 'NO'
-  fsdp_config: {}
-  machine_rank: 0
-  main_training_function: main
-  megatron_lm_config: {}
-  mixed_precision: 'no'
-  num_machines: 1
-  num_processes: 1
-  rdzv_backend: static
-  same_network: true
-  use_cpu: false
-  ```
-  b. run the below command to launch example script
-  ```
-  accelerate launch --config_file ds_zero3_cpu.yaml examples/peft_lora_seq2seq_accelerate_ds_zero3_offload.py
-  ```
-
-  c. output logs:
-  ```bash
-  GPU Memory before entering the train : 1916
-  GPU Memory consumed at the end of the train (end-begin): 66
-  GPU Peak Memory consumed during the train (max-begin): 7488
-  GPU Total Peak Memory consumed during the train (max): 9404
-  CPU Memory before entering the train : 19411
-  CPU Memory consumed at the end of the train (end-begin): 0
-  CPU Peak Memory consumed during the train (max-begin): 0
-  CPU Total Peak Memory consumed during the train (max): 19411
-  epoch=4: train_ppl=tensor(1.0705, device='cuda:0') train_epoch_loss=tensor(0.0681, device='cuda:0')
-  100%|████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:27<00:00,  3.92s/it]
-  GPU Memory before entering the eval : 1982
-  GPU Memory consumed at the end of the eval (end-begin): -66
-  GPU Peak Memory consumed during the eval (max-begin): 672
-  GPU Total Peak Memory consumed during the eval (max): 2654
-  CPU Memory before entering the eval : 19411
-  CPU Memory consumed at the end of the eval (end-begin): 0
-  CPU Peak Memory consumed during the eval (max-begin): 0
-  CPU Total Peak Memory consumed during the eval (max): 19411
-  accuracy=100.0
-  eval_preds[:10]=['no complaint', 'no complaint', 'complaint', 'complaint', 'no complaint', 'no complaint', 'no complaint', 'complaint', 'complaint', 'no complaint']
-  dataset['train'][label_column][:10]=['no complaint', 'no complaint', 'complaint', 'complaint', 'no complaint', 'no complaint', 'no complaint', 'complaint', 'complaint', 'no complaint']
-  ```
-
-### Example of PEFT model inference using 🤗 Accelerate's Big Model Inferencing capabilities
-
-Example is provided in `~examples/causal_language_modeling/peft_lora_clm_accelerate_big_model_inference.ipynb`. 
-
-
-## Models support matrix
-
-### Causal Language Modeling
-|   Model         | LoRA | Prefix Tuning  | P-Tuning | Prompt Tuning  |
-| --------- | ---- | ---- | ---- | ----  |
-| GPT-2          | ✅  | ✅  | ✅  | ✅  |
-| Bloom          | ✅  | ✅  | ✅  | ✅  |
-| OPT            | ✅  | ✅  | ✅  | ✅  |
-| GPT-Neo        | ✅  | ✅  | ✅  | ✅  |
-| GPT-J          | ✅  | ✅  | ✅  | ✅  |
-
-### Conditional Generation
-|   Model         | LoRA | Prefix Tuning  | P-Tuning | Prompt Tuning  | 
-| --------- | ---- | ---- | ---- | ---- |
-| T5        | ✅   | ✅   | ✅   | ✅   |
-| BART      | ✅   | ✅   | ✅   | ✅   |
-
-### Sequence Classification
-|   Model         | LoRA | Prefix Tuning  | P-Tuning | Prompt Tuning  | 
-| --------- | ---- | ---- | ---- | ----  |
-| BERT           | ✅  | ✅  | ✅  | ✅  |  
-| RoBERTa        | ✅  | ✅  | ✅  | ✅  |
-| GPT-2          | ✅  | ✅  | ✅  | ✅  | 
-| Bloom          | ✅  | ✅  | ✅  | ✅  |   
-| OPT            | ✅  | ✅  | ✅  | ✅  |
-| GPT-Neo        | ✅  | ✅  | ✅  | ✅  |
-| GPT-J          | ✅  | ✅  | ✅  | ✅  |
-| Deberta        | ✅  |     | ✅  | ✅  |     
-| Deberta-v2     | ✅  |     | ✅  | ✅  |    
-
-### Token Classification
-|   Model         | LoRA | Prefix Tuning  | P-Tuning | Prompt Tuning  | 
-| --------- | ---- | ---- | ---- | ----  |
-| BERT           | ✅  | ✅  |   |   |  
-| RoBERTa        | ✅  | ✅  |   |   |
-| GPT-2          | ✅  | ✅  |   |   | 
-| Bloom          | ✅  | ✅  |   |   |   
-| OPT            | ✅  | ✅  |   |   |
-| GPT-Neo        | ✅  | ✅  |   |   |
-| GPT-J          | ✅  | ✅  |   |   |
-| Deberta        | ✅  |     |   |   | 
-| Deberta-v2     | ✅  |     |   |   |
-
-
-## Caveats:
-
-1. Below is an example of using PyTorch FSDP for training. However, it doesn't lead to 
-any GPU memory savings. Please refer issue [[FSDP] FSDP with CPU offload consumes 1.65X more GPU memory when training models with most of the params frozen](https://github.com/pytorch/pytorch/issues/91165). 
-
-  ```python
-  from peft.utils.other import fsdp_auto_wrap_policy
-
-  ...
-
-  if os.environ.get("ACCELERATE_USE_FSDP", None) is not None:
-      accelerator.state.fsdp_plugin.auto_wrap_policy = fsdp_auto_wrap_policy(model)
-
-  model = accelerator.prepare(model)
-  ```
-
-  Example of parameter efficient tuning with `mt0-xxl` base model using 🤗 Accelerate is provided in `~examples/conditional_generation/peft_lora_seq2seq_accelerate_fsdp.py`. 
-  a. First run `accelerate config --config_file fsdp_config.yaml` and answer the questionaire. 
-  Below are the contents of the config file.
-  ```
-  command_file: null
-  commands: null
-  compute_environment: LOCAL_MACHINE
-  deepspeed_config: {}
-  distributed_type: FSDP
-  downcast_bf16: 'no'
-  dynamo_backend: 'NO'
-  fsdp_config:
-    fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-    fsdp_backward_prefetch_policy: BACKWARD_PRE
-    fsdp_offload_params: true
-    fsdp_sharding_strategy: 1
-    fsdp_state_dict_type: FULL_STATE_DICT
-    fsdp_transformer_layer_cls_to_wrap: T5Block
-  gpu_ids: null
-  machine_rank: 0
-  main_process_ip: null
-  main_process_port: null
-  main_training_function: main
-  megatron_lm_config: {}
-  mixed_precision: 'no'
-  num_machines: 1
-  num_processes: 2
-  rdzv_backend: static
-  same_network: true
-  tpu_name: null
-  tpu_zone: null
-  use_cpu: false
-  ```
-  b. run the below command to launch example script
-  ```
-  accelerate launch --config_file fsdp_config.yaml examples/peft_lora_seq2seq_accelerate_fsdp.py
-  ```
-
-2. When using `P_TUNING` or `PROMPT_TUNING` with `SEQ_2_SEQ` task, remember to remove the `num_virtual_token` virtual prompt predictions from the left side of the model outputs during evaluations. 
-
-3. `P_TUNING` or `PROMPT_TUNING` doesn't support `generate` functionality of transformers bcause `generate` strictly requires `input_ids`/`decoder_input_ids` but 
-`P_TUNING`/`PROMPT_TUNING` appends soft prompt embeddings to `input_embeds` to create
-new `input_embeds` to be given to the model. Therefore, `generate` doesn't support this yet.
-
-## Backlog:
-1. Explore and possibly integrate `(IA)^3` and `UniPELT`
-2. Add tests
-3. Add more use cases and examples
+If you would like to contribute to PEFT, please check out our [contribution guide](https://huggingface.co/docs/peft/developer_guides/contributing).
 
 ## Citing 🤗 PEFT
 
-If you use 🤗 PEFT in your publication, please cite it by using the following BibTeX entry.
+To use 🤗 PEFT in your publication, please cite it by using the following BibTeX entry.
 
 ```bibtex
 @Misc{peft,
   title =        {PEFT: State-of-the-art Parameter-Efficient Fine-Tuning methods},
-  author =       {Sourab Mangrulkar, Sylvain Gugger},
+  author =       {Sourab Mangrulkar and Sylvain Gugger and Lysandre Debut and Younes Belkada and Sayak Paul and Benjamin Bossan},
   howpublished = {\url{https://github.com/huggingface/peft}},
   year =         {2022}
 }
-```
+```

docker/README.md

@@ -0,0 +1,11 @@
+# PEFT Docker images
+
+Here we store all PEFT Docker images used in our testing infrastructure. We use python 3.8 for now on all our images.
+
+- `peft-cpu`: PEFT compiled on CPU with all other HF libraries installed on main branch
+- `peft-gpu`: PEFT complied for NVIDIA GPUs wih all other HF libraries installed on main branch
+- `peft-gpu-bnb-source`: PEFT complied for NVIDIA GPUs with `bitsandbytes` and all other HF libraries installed from main branch
+- `peft-gpu-bnb-latest`: PEFT complied for NVIDIA GPUs with `bitsandbytes` complied from main and all other HF libraries installed from latest PyPi
+- `peft-gpu-bnb-multi-source`: PEFT complied for NVIDIA GPUs with `bitsandbytes` complied from `multi-backend` branch and all other HF libraries installed from main branch
+
+`peft-gpu-bnb-source` and `peft-gpu-bnb-multi-source` are essentially the same, with the only difference being `bitsandbytes` compiled on another branch. Make sure to propagate the changes you applied on one file to the other!

docker/peft-cpu/Dockerfile

@@ -0,0 +1,52 @@
+# Builds GPU docker image of PyTorch
+# Uses multi-staged approach to reduce size
+# Stage 1
+# Use base conda image to reduce time
+FROM continuumio/miniconda3:latest AS compile-image
+# Specify py version
+ENV PYTHON_VERSION=3.8
+# Install apt libs - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+RUN apt-get update && \
+    apt-get install -y curl git wget software-properties-common git-lfs && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+
+# Install audio-related libraries 
+RUN apt-get update && \
+    apt install -y ffmpeg
+
+RUN apt install -y libsndfile1-dev
+RUN git lfs install
+
+# Create our conda env - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+RUN conda create --name peft python=${PYTHON_VERSION} ipython jupyter pip
+RUN python3 -m pip install --no-cache-dir --upgrade pip
+
+# Below is copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+# We don't install pytorch here yet since CUDA isn't available
+# instead we use the direct torch wheel
+ENV PATH /opt/conda/envs/peft/bin:$PATH
+# Activate our bash shell
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+# Activate the conda env and install transformers + accelerate from source
+RUN source activate peft && \
+    python3 -m pip install --no-cache-dir \
+    librosa \
+    "soundfile>=0.12.1" \
+    scipy \
+    git+https://github.com/huggingface/transformers \
+    git+https://github.com/huggingface/accelerate \
+    peft[test]@git+https://github.com/huggingface/peft
+
+# Install apt libs
+RUN apt-get update && \
+    apt-get install -y curl git wget && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+RUN echo "source activate peft" >> ~/.profile
+
+# Activate the virtualenv
+CMD ["/bin/bash"]

docker/peft-gpu-bnb-latest/Dockerfile

@@ -0,0 +1,68 @@
+# Builds GPU docker image of PyTorch
+# Uses multi-staged approach to reduce size
+# Stage 1
+# Use base conda image to reduce time
+FROM continuumio/miniconda3:latest AS compile-image
+# Specify py version
+ENV PYTHON_VERSION=3.8
+# Install apt libs - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+RUN apt-get update && \
+    apt-get install -y curl git wget software-properties-common git-lfs && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+# Install audio-related libraries 
+RUN apt-get update && \
+    apt install -y ffmpeg
+
+RUN apt install -y libsndfile1-dev
+RUN git lfs install
+
+# Create our conda env - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+RUN conda create --name peft python=${PYTHON_VERSION} ipython jupyter pip
+RUN python3 -m pip install --no-cache-dir --upgrade pip
+
+# Below is copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+# We don't install pytorch here yet since CUDA isn't available
+# instead we use the direct torch wheel
+ENV PATH /opt/conda/envs/peft/bin:$PATH
+# Activate our bash shell
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+
+# Stage 2
+FROM nvidia/cuda:12.1.0-devel-ubuntu22.04 AS build-image
+COPY --from=compile-image /opt/conda /opt/conda
+ENV PATH /opt/conda/bin:$PATH
+
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+
+# Install apt libs
+RUN apt-get update && \
+    apt-get install -y curl git wget cmake && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+# Activate the conda env and install transformers + accelerate from latest pypi
+# Also clone BNB and build it from source.
+RUN source activate peft && \
+    python3 -m pip install -U --no-cache-dir \
+    librosa \
+    "soundfile>=0.12.1" \
+    scipy \
+    transformers \
+    accelerate \
+    peft \
+    optimum \
+    auto-gptq && \
+    git clone https://github.com/TimDettmers/bitsandbytes && cd bitsandbytes && \
+    cmake -B . -DCOMPUTE_BACKEND=cuda -S . && \
+    cmake --build . && \
+    pip install -e . && \ 
+    pip freeze | grep bitsandbytes
+
+RUN echo "source activate peft" >> ~/.profile
+
+# Activate the virtualenv
+CMD ["/bin/bash"]

docker/peft-gpu-bnb-multi-source/Dockerfile

@@ -0,0 +1,68 @@
+# Builds GPU docker image of PyTorch
+# Uses multi-staged approach to reduce size
+# Stage 1
+# Use base conda image to reduce time
+FROM continuumio/miniconda3:latest AS compile-image
+# Specify py version
+ENV PYTHON_VERSION=3.8
+# Install apt libs - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+RUN apt-get update && \
+    apt-get install -y curl git wget software-properties-common git-lfs && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+# Install audio-related libraries 
+RUN apt-get update && \
+    apt install -y ffmpeg
+
+RUN apt install -y libsndfile1-dev
+RUN git lfs install
+
+# Create our conda env - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+RUN conda create --name peft python=${PYTHON_VERSION} ipython jupyter pip
+RUN python3 -m pip install --no-cache-dir --upgrade pip
+
+# Below is copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+# We don't install pytorch here yet since CUDA isn't available
+# instead we use the direct torch wheel
+ENV PATH /opt/conda/envs/peft/bin:$PATH
+# Activate our bash shell
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+
+# Stage 2
+FROM nvidia/cuda:12.1.0-devel-ubuntu22.04 AS build-image
+COPY --from=compile-image /opt/conda /opt/conda
+ENV PATH /opt/conda/bin:$PATH
+
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+
+# Install apt libs
+RUN apt-get update && \
+    apt-get install -y curl git wget cmake && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+# Activate the conda env and install transformers + accelerate from source
+# Also clone BNB and build it from source.
+RUN source activate peft && \
+    python3 -m pip install -U --no-cache-dir \
+    librosa \
+    "soundfile>=0.12.1" \
+    scipy \
+    git+https://github.com/huggingface/transformers \
+    git+https://github.com/huggingface/accelerate \
+    peft[test]@git+https://github.com/huggingface/peft \
+    optimum \
+    auto-gptq && \
+    git clone https://github.com/TimDettmers/bitsandbytes && cd bitsandbytes && git checkout multi-backend-refactor && \
+    cmake -B . -DCOMPUTE_BACKEND=cuda -S . && \
+    cmake --build . && \
+    pip install -e . && \ 
+    pip freeze | grep bitsandbytes
+
+RUN echo "source activate peft" >> ~/.profile
+
+# Activate the virtualenv
+CMD ["/bin/bash"]

docker/peft-gpu-bnb-source/Dockerfile

@@ -0,0 +1,68 @@
+# Builds GPU docker image of PyTorch
+# Uses multi-staged approach to reduce size
+# Stage 1
+# Use base conda image to reduce time
+FROM continuumio/miniconda3:latest AS compile-image
+# Specify py version
+ENV PYTHON_VERSION=3.8
+# Install apt libs - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+RUN apt-get update && \
+    apt-get install -y curl git wget software-properties-common git-lfs && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+# Install audio-related libraries 
+RUN apt-get update && \
+    apt install -y ffmpeg
+
+RUN apt install -y libsndfile1-dev
+RUN git lfs install
+
+# Create our conda env - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+RUN conda create --name peft python=${PYTHON_VERSION} ipython jupyter pip
+RUN python3 -m pip install --no-cache-dir --upgrade pip
+
+# Below is copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+# We don't install pytorch here yet since CUDA isn't available
+# instead we use the direct torch wheel
+ENV PATH /opt/conda/envs/peft/bin:$PATH
+# Activate our bash shell
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+
+# Stage 2
+FROM nvidia/cuda:12.1.0-devel-ubuntu22.04 AS build-image
+COPY --from=compile-image /opt/conda /opt/conda
+ENV PATH /opt/conda/bin:$PATH
+
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+
+# Install apt libs
+RUN apt-get update && \
+    apt-get install -y curl git wget cmake && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+# Activate the conda env and install transformers + accelerate from source
+# Also clone BNB and build it from source.
+RUN source activate peft && \
+    python3 -m pip install -U --no-cache-dir \
+    librosa \
+    "soundfile>=0.12.1" \
+    scipy \
+    git+https://github.com/huggingface/transformers \
+    git+https://github.com/huggingface/accelerate \
+    peft[test]@git+https://github.com/huggingface/peft \
+    optimum \
+    auto-gptq && \
+    git clone https://github.com/TimDettmers/bitsandbytes && cd bitsandbytes && \
+    cmake -B . -DCOMPUTE_BACKEND=cuda -S . && \
+    cmake --build . && \
+    pip install -e . && \ 
+    pip freeze | grep bitsandbytes
+
+RUN echo "source activate peft" >> ~/.profile
+
+# Activate the virtualenv
+CMD ["/bin/bash"]

docker/peft-gpu/Dockerfile

@@ -0,0 +1,83 @@
+# Builds GPU docker image of PyTorch
+# Uses multi-staged approach to reduce size
+# Stage 1
+# Use base conda image to reduce time
+FROM continuumio/miniconda3:latest AS compile-image
+# Specify py version
+ENV PYTHON_VERSION=3.8
+# Install apt libs - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+RUN apt-get update && \
+    apt-get install -y curl git wget software-properties-common git-lfs && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+# Install audio-related libraries 
+RUN apt-get update && \
+    apt install -y ffmpeg
+
+RUN apt install -y libsndfile1-dev
+RUN git lfs install
+
+# Create our conda env - copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+RUN conda create --name peft python=${PYTHON_VERSION} ipython jupyter pip
+RUN python3 -m pip install --no-cache-dir --upgrade pip
+
+# Below is copied from https://github.com/huggingface/accelerate/blob/main/docker/accelerate-gpu/Dockerfile
+# We don't install pytorch here yet since CUDA isn't available
+# instead we use the direct torch wheel
+ENV PATH /opt/conda/envs/peft/bin:$PATH
+# Activate our bash shell
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+
+# Stage 2
+FROM nvidia/cuda:12.2.2-devel-ubuntu22.04 AS build-image
+COPY --from=compile-image /opt/conda /opt/conda
+ENV PATH /opt/conda/bin:$PATH
+
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+RUN source activate peft && \ 
+    python3 -m pip install --no-cache-dir bitsandbytes optimum auto-gptq
+
+# Add autoawq for quantization testing
+RUN source activate peft && \
+    python3 -m pip install --no-cache-dir https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.4/autoawq-0.2.4-cp38-cp38-linux_x86_64.whl
+RUN source activate peft && \
+    python3 -m pip install --no-cache-dir https://github.com/casper-hansen/AutoAWQ_kernels/releases/download/v0.0.6/autoawq_kernels-0.0.6-cp38-cp38-linux_x86_64.whl
+
+# Install apt libs
+RUN apt-get update && \
+    apt-get install -y curl git wget && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+# Add eetq for quantization testing
+RUN source activate peft && \
+    python3 -m pip install git+https://github.com/NetEase-FuXi/EETQ.git
+
+# Activate the conda env and install transformers + accelerate from source
+RUN source activate peft && \
+    python3 -m pip install -U --no-cache-dir \
+    librosa \
+    "soundfile>=0.12.1" \
+    scipy \
+    git+https://github.com/huggingface/transformers \
+    git+https://github.com/huggingface/accelerate \
+    peft[test]@git+https://github.com/huggingface/peft
+
+# Add aqlm for quantization testing
+RUN source activate peft && \
+    pip install aqlm[gpu]>=1.0.2
+
+# Add HQQ for quantization testing
+RUN source activate peft && \
+pip install hqq
+
+RUN source activate peft && \ 
+    pip freeze | grep transformers
+
+RUN echo "source activate peft" >> ~/.profile
+
+# Activate the virtualenv
+CMD ["/bin/bash"]

docs/Makefile

@@ -0,0 +1,19 @@
+# Minimal makefile for Sphinx documentation
+#
+
+# You can set these variables from the command line.
+SPHINXOPTS    =
+SPHINXBUILD   = sphinx-build
+SOURCEDIR     = source
+BUILDDIR      = _build
+
+# Put it first so that "make" without argument is like "make help".
+help:
+	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+
+.PHONY: help Makefile
+
+# Catch-all target: route all unknown targets to Sphinx using the new
+# "make mode" option.  $(O) is meant as a shortcut for $(SPHINXOPTS).
+%: Makefile
+	@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)

docs/README.md

@@ -0,0 +1,267 @@
+<!---
+Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
+# Generating the documentation
+
+To generate the documentation, you first have to build it. Several packages are necessary to build the doc, 
+you can install them with the following command, at the root of the code repository:
+
+```bash
+pip install -e ".[docs]"
+```
+
+Then you need to install our special tool that builds the documentation:
+
+```bash
+pip install git+https://github.com/huggingface/doc-builder
+```
+
+---
+**NOTE**
+
+You only need to generate the documentation to inspect it locally (if you're planning changes and want to
+check how they look before committing for instance). You don't have to commit to the built documentation.
+
+---
+
+## Building the documentation
+
+Once you have setup the `doc-builder` and additional packages, you can generate the documentation by 
+typing the following command:
+
+```bash
+doc-builder build peft docs/source/ --build_dir ~/tmp/test-build
+```
+
+You can adapt the `--build_dir` to set any temporary folder you prefer. This command will create it and generate
+the MDX files that will be rendered as the documentation on the main website. You can inspect them in your favorite
+Markdown editor.
+
+## Previewing the documentation
+
+To preview the docs, first install the `watchdog` module with:
+
+```bash
+pip install watchdog
+```
+
+Then run the following command:
+
+```bash
+doc-builder preview {package_name} {path_to_docs}
+```
+
+For example:
+
+```bash
+doc-builder preview peft docs/source
+```
+
+The docs will be viewable at [http://localhost:3000](http://localhost:3000). You can also preview the docs once you have opened a PR. You will see a bot add a comment to a link where the documentation with your changes lives.
+
+---
+**NOTE**
+
+The `preview` command only works with existing doc files. When you add a completely new file, you need to update `_toctree.yml` & restart `preview` command (`ctrl-c` to stop it & call `doc-builder preview ...` again).
+
+---
+
+## Adding a new element to the navigation bar
+
+Accepted files are Markdown (.md or .mdx).
+
+Create a file with its extension and put it in the source directory. You can then link it to the toc-tree by putting
+the filename without the extension in the [`_toctree.yml`](https://github.com/huggingface/peft/blob/main/docs/source/_toctree.yml) file.
+
+## Renaming section headers and moving sections
+
+It helps to keep the old links working when renaming the section header and/or moving sections from one document to another. This is because the old links are likely to be used in Issues, Forums, and Social media and it'd make for a much more superior user experience if users reading those months later could still easily navigate to the originally intended information.
+
+Therefore, we simply keep a little map of moved sections at the end of the document where the original section was. The key is to preserve the original anchor.
+
+So if you renamed a section from: "Section A" to "Section B", then you can add at the end of the file:
+
+```
+Sections that were moved:
+
+[ <a href="#section-b">Section A</a><a id="section-a"></a> ]
+```
+and of course, if you moved it to another file, then:
+
+```
+Sections that were moved:
+
+[ <a href="../new-file#section-b">Section A</a><a id="section-a"></a> ]
+```
+
+Use the relative style to link to the new file so that the versioned docs continue to work.
+
+
+## Writing Documentation - Specification
+
+The `huggingface/peft` documentation follows the
+[Google documentation](https://sphinxcontrib-napoleon.readthedocs.io/en/latest/example_google.html) style for docstrings,
+although we can write them directly in Markdown.
+
+### Adding a new tutorial
+
+Adding a new tutorial or section is done in two steps:
+
+- Add a new file under `./source`. This file can either be ReStructuredText (.rst) or Markdown (.md).
+- Link that file in `./source/_toctree.yml` on the correct toc-tree.
+
+Make sure to put your new file under the proper section. It's unlikely to go in the first section (*Get Started*), so
+depending on the intended targets (beginners, more advanced users, or researchers) it should go into sections two, three, or
+four.
+
+### Writing source documentation
+
+Values that should be put in `code` should either be surrounded by backticks: \`like so\`. Note that argument names
+and objects like True, None, or any strings should usually be put in `code`.
+
+When mentioning a class, function, or method, it is recommended to use our syntax for internal links so that our tool
+adds a link to its documentation with this syntax: \[\`XXXClass\`\] or \[\`function\`\]. This requires the class or 
+function to be in the main package.
+
+If you want to create a link to some internal class or function, you need to
+provide its path. For instance: \[\`utils.gather\`\]. This will be converted into a link with
+`utils.gather` in the description. To get rid of the path and only keep the name of the object you are
+linking to in the description, add a ~: \[\`~utils.gather\`\] will generate a link with `gather` in the description.
+
+The same works for methods so you can either use \[\`XXXClass.method\`\] or \[~\`XXXClass.method\`\].
+
+#### Defining arguments in a method
+
+Arguments should be defined with the `Args:` (or `Arguments:` or `Parameters:`) prefix, followed by a line return and
+an indentation. The argument should be followed by its type, with its shape if it is a tensor, a colon, and its
+description:
+
+```
+    Args:
+        n_layers (`int`): The number of layers of the model.
+```
+
+If the description is too long to fit in one line (more than 119 characters in total), another indentation is necessary 
+before writing the description after the argument.
+
+Finally, to maintain uniformity if any *one* description is too long to fit on one line, the 
+rest of the parameters should follow suit and have an indention before their description.
+
+Here's an example showcasing everything so far:
+
+```
+    Args:
+        gradient_accumulation_steps (`int`, *optional*, default to 1):
+            The number of steps that should pass before gradients are accumulated. A number > 1 should be combined with `Accelerator.accumulate`.
+        cpu (`bool`, *optional*):
+            Whether or not to force the script to execute on CPU. Will ignore GPU available if set to `True` and force the execution on one process only.
+```
+
+For optional arguments or arguments with defaults we follow the following syntax: imagine we have a function with the
+following signature:
+
+```
+def my_function(x: str = None, a: float = 1):
+```
+
+then its documentation should look like this:
+
+```
+    Args:
+        x (`str`, *optional*):
+            This argument controls ... and has a description longer than 119 chars.
+        a (`float`, *optional*, defaults to 1):
+            This argument is used to ... and has a description longer than 119 chars.
+```
+
+Note that we always omit the "defaults to \`None\`" when None is the default for any argument. Also note that even
+if the first line describing your argument type and its default gets long, you can't break it into several lines. You can
+however write as many lines as you want in the indented description (see the example above with `input_ids`).
+
+#### Writing a multi-line code block
+
+Multi-line code blocks can be useful for displaying examples. They are done between two lines of three backticks as usual in Markdown:
+
+
+````
+```python
+# first line of code
+# second line
+# etc
+```
+````
+
+#### Writing a return block
+
+The return block should be introduced with the `Returns:` prefix, followed by a line return and an indentation.
+The first line should be the type of the return, followed by a line return. No need to indent further for the elements
+building the return.
+
+Here's an example of a single value return:
+
+```
+    Returns:
+        `List[int]`: A list of integers in the range [0, 1] --- 1 for a special token, 0 for a sequence token.
+```
+
+Here's an example of a tuple return, comprising several objects:
+
+```
+    Returns:
+        `tuple(torch.FloatTensor)` comprising various elements depending on the configuration ([`BertConfig`]) and inputs:
+        - ** loss** (*optional*, returned when `masked_lm_labels` is provided) `torch.FloatTensor` of shape `(1,)` --
+          Total loss is the sum of the masked language modeling loss and the next sequence prediction (classification) loss.
+        - **prediction_scores** (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`) --
+          Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+```
+
+## Styling the docstring
+
+We have an automatic script running with the `make style` comment that will make sure that:
+- the docstrings fully take advantage of the line width
+- all code examples are formatted using black, like the code of the Transformers library
+
+This script may have some weird failures if you make a syntax mistake or if you uncover a bug. Therefore, it's
+recommended to commit your changes before running `make style`, so you can revert the changes done by that script
+easily.
+
+## Writing documentation examples
+
+The syntax, for example, docstrings can look as follows:
+
+```
+    Example:
+
+    ```python
+    >>> import time
+    >>> from accelerate import Accelerator
+    >>> accelerator = Accelerator()
+    >>> if accelerator.is_main_process:
+    ...     time.sleep(2)
+    >>> else:
+    ...     print("I'm waiting for the main process to finish its sleep...")
+    >>> accelerator.wait_for_everyone()
+    >>> # Should print on every process at the same time
+    >>> print("Everyone is here")
+    ```
+```
+
+The docstring should give a minimal, clear example of how the respective function 
+is to be used in inference and also include the expected (ideally sensible)
+output.
+Often, readers will try out the example before even going through the function 
+or class definitions. Therefore, it is of utmost importance that the example 
+works as expected.

docs/source/_config.py

@@ -0,0 +1,7 @@
+# docstyle-ignore
+INSTALL_CONTENT = """
+# PEFT installation
+! pip install peft accelerate transformers
+# To install from source instead of the last release, comment the command above and uncomment the following one.
+# ! pip install git+https://github.com/huggingface/peft.git
+"""

docs/source/_toctree.yml

@@ -0,0 +1,127 @@
+- title: Get started
+  sections:
+  - local: index
+    title: 🤗 PEFT
+  - local: quicktour
+    title: Quicktour
+  - local: install
+    title: Installation
+
+- title: Tutorial
+  sections:
+  - local: tutorial/peft_model_config
+    title: Configurations and models
+  - local: tutorial/peft_integrations
+    title: Integrations
+
+- title: PEFT method guides
+  sections:
+  - local: task_guides/prompt_based_methods
+    title: Prompt-based methods
+  - local: task_guides/lora_based_methods
+    title: LoRA methods
+  - local: task_guides/ia3
+    title: IA3
+
+- title: Developer guides
+  sections:
+  - local: developer_guides/model_merging
+    title: Model merging
+  - local: developer_guides/quantization
+    title: Quantization
+  - local: developer_guides/lora
+    title: LoRA
+  - local: developer_guides/custom_models
+    title: Custom models
+  - local: developer_guides/low_level_api
+    title: Adapter injection
+  - local: developer_guides/mixed_models
+    title: Mixed adapter types
+  - local: developer_guides/torch_compile
+    title: torch.compile
+  - local: developer_guides/contributing
+    title: Contribute to PEFT
+  - local: developer_guides/troubleshooting
+    title: Troubleshooting
+  - local: developer_guides/checkpoint
+    title: PEFT checkpoint format
+
+- title: 🤗 Accelerate integrations
+  sections:
+  - local: accelerate/deepspeed
+    title: DeepSpeed
+  - local: accelerate/fsdp
+    title: Fully Sharded Data Parallel
+
+- title: Conceptual guides
+  sections:
+  - local: conceptual_guides/adapter
+    title: Adapters
+  - local: conceptual_guides/prompting
+    title: Soft prompts
+  - local: conceptual_guides/ia3
+    title: IA3
+  - local: conceptual_guides/oft
+    title: OFT/BOFT
+
+- sections:
+  - sections:
+    - local: package_reference/auto_class
+      title: AutoPeftModel
+    - local: package_reference/peft_model
+      title: PEFT model
+    - local: package_reference/peft_types
+      title: PEFT types
+    - local: package_reference/config
+      title: Configuration
+    - local: package_reference/tuners
+      title: Tuner
+    title: Main classes
+  - sections:
+    - local: package_reference/adalora
+      title: AdaLoRA
+    - local: package_reference/ia3
+      title: IA3
+    - local: package_reference/llama_adapter
+      title: Llama-Adapter
+    - local: package_reference/loha
+      title: LoHa
+    - local: package_reference/lokr
+      title: LoKr
+    - local: package_reference/lora
+      title: LoRA
+    - local: package_reference/xlora
+      title: X-LoRA
+    - local: package_reference/adapter_utils
+      title: LyCORIS
+    - local: package_reference/multitask_prompt_tuning
+      title: Multitask Prompt Tuning
+    - local: package_reference/oft
+      title: OFT
+    - local: package_reference/boft
+      title: BOFT
+    - local: package_reference/poly
+      title: Polytropon
+    - local: package_reference/p_tuning
+      title: P-tuning
+    - local: package_reference/prefix_tuning
+      title: Prefix tuning
+    - local: package_reference/prompt_tuning
+      title: Prompt tuning
+    - local: package_reference/layernorm_tuning
+      title: Layernorm tuning
+    - local: package_reference/vera
+      title: VeRA
+    - local: package_reference/fourierft
+      title: FourierFT
+    - local: package_reference/vblora
+      title: VB-LoRA
+
+    title: Adapters
+  - sections:
+    - local: package_reference/merge_utils
+      title: Model merge
+    - local: package_reference/helpers
+      title: Helpers
+    title: Utilities
+  title: API reference

docs/source/accelerate/deepspeed.md

@@ -0,0 +1,447 @@
+<!--⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# DeepSpeed
+
+[DeepSpeed](https://www.deepspeed.ai/) is a library designed for speed and scale for distributed training of large models with billions of parameters. At its core is the Zero Redundancy Optimizer (ZeRO) that shards optimizer states (ZeRO-1), gradients (ZeRO-2), and parameters (ZeRO-3) across data parallel processes. This drastically reduces memory usage, allowing you to scale your training to billion parameter models. To unlock even more memory efficiency, ZeRO-Offload reduces GPU compute and memory by leveraging CPU resources during optimization.
+
+Both of these features are supported in 🤗 Accelerate, and you can use them with 🤗 PEFT. 
+
+## Compatibility with `bitsandbytes` quantization + LoRA
+
+Below is a table that summarizes the compatibility between PEFT's LoRA, [`bitsandbytes`](https://github.com/TimDettmers/bitsandbytes) library and DeepSpeed Zero stages with respect to fine-tuning. DeepSpeed Zero-1 and 2 will have no effect at inference as stage 1 shards the optimizer states and stage 2 shards the optimizer states and gradients:
+
+| DeepSpeed stage   | Is compatible? |
+|---|---|
+| Zero-1 |  🟢 |
+| Zero-2   |  🟢 |
+| Zero-3  |  🟢 |
+
+For DeepSpeed Stage 3 + QLoRA, please refer to the section [Use PEFT QLoRA and DeepSpeed with ZeRO3 for finetuning large models on multiple GPUs](#use-peft-qlora-and-deepspeed-with-zero3-for-finetuning-large-models-on-multiple-gpus) below.
+
+For confirming these observations, we ran the SFT (Supervised Fine-tuning) [offical example scripts](https://github.com/huggingface/trl/tree/main/examples) of the [Transformers Reinforcement Learning (TRL) library](https://github.com/huggingface/trl) using QLoRA + PEFT and the accelerate configs available [here](https://github.com/huggingface/trl/tree/main/examples/accelerate_configs). We ran these experiments on a 2x NVIDIA T4 GPU.
+
+# Use PEFT and DeepSpeed with ZeRO3 for finetuning large models on multiple devices and multiple nodes
+
+This section of guide will help you learn how to use our DeepSpeed [training script](https://github.com/huggingface/peft/blob/main/examples/sft/train.py) for performing SFT. You'll configure the script to do SFT (supervised fine-tuning) of Llama-70B model with LoRA and ZeRO-3 on 8xH100 80GB GPUs on a single machine. You can configure it to scale to multiple machines by changing the accelerate config.
+
+## Configuration
+
+Start by running the following command to [create a DeepSpeed configuration file](https://huggingface.co/docs/accelerate/quicktour#launching-your-distributed-script) with 🤗 Accelerate. The `--config_file` flag allows you to save the configuration file to a specific location, otherwise it is saved as a `default_config.yaml` file in the 🤗 Accelerate cache.
+
+The configuration file is used to set the default options when you launch the training script.
+
+```bash
+accelerate config --config_file deepspeed_config.yaml
+```
+
+You'll be asked a few questions about your setup, and configure the following arguments. In this example, you'll use ZeRO-3 so make sure you pick those options.
+
+```bash
+`zero_stage`: [0] Disabled, [1] optimizer state partitioning, [2] optimizer+gradient state partitioning and [3] optimizer+gradient+parameter partitioning
+`gradient_accumulation_steps`: Number of training steps to accumulate gradients before averaging and applying them. Pass the same value as you would pass via cmd argument else you will encounter mismatch error.
+`gradient_clipping`: Enable gradient clipping with value. Don't set this as you will be passing it via cmd arguments.
+`offload_optimizer_device`: [none] Disable optimizer offloading, [cpu] offload optimizer to CPU, [nvme] offload optimizer to NVMe SSD. Only applicable with ZeRO >= Stage-2. Set this as `none` as don't want to enable offloading.
+`offload_param_device`: [none] Disable parameter offloading, [cpu] offload parameters to CPU, [nvme] offload parameters to NVMe SSD. Only applicable with ZeRO Stage-3. Set this as `none` as don't want to enable offloading.
+`zero3_init_flag`: Decides whether to enable `deepspeed.zero.Init` for constructing massive models. Only applicable with ZeRO Stage-3. Set this to `True`.
+`zero3_save_16bit_model`: Decides whether to save 16-bit model weights when using ZeRO Stage-3. Set this to `True`.
+`mixed_precision`: `no` for FP32 training, `fp16` for FP16 mixed-precision training and `bf16` for BF16 mixed-precision training. Set this to `True`.
+```
+
+Once this is done, the corresponding config should look like below and you can find it in config folder at [deepspeed_config.yaml](https://github.com/huggingface/peft/blob/main/examples/sft/configs/deepspeed_config.yaml):
+
+```yml
+compute_environment: LOCAL_MACHINE                                                                                                                                           
+debug: false
+deepspeed_config:
+  deepspeed_multinode_launcher: standard
+  gradient_accumulation_steps: 4
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: true
+  zero3_save_16bit_model: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+## Launch command
+
+The launch command is available at [run_peft_deepspeed.sh](https://github.com/huggingface/peft/blob/main/examples/sft/run_peft_deepspeed.sh) and it is also shown below:
+```bash
+accelerate launch --config_file "configs/deepspeed_config.yaml"  train.py \
+--seed 100 \
+--model_name_or_path "meta-llama/Llama-2-70b-hf" \
+--dataset_name "smangrul/ultrachat-10k-chatml" \
+--chat_template_format "chatml" \
+--add_special_tokens False \
+--append_concat_token False \
+--splits "train,test" \
+--max_seq_len 2048 \
+--num_train_epochs 1 \
+--logging_steps 5 \
+--log_level "info" \
+--logging_strategy "steps" \
+--eval_strategy "epoch" \
+--save_strategy "epoch" \
+--push_to_hub \
+--hub_private_repo True \
+--hub_strategy "every_save" \
+--bf16 True \
+--packing True \
+--learning_rate 1e-4 \
+--lr_scheduler_type "cosine" \
+--weight_decay 1e-4 \
+--warmup_ratio 0.0 \
+--max_grad_norm 1.0 \
+--output_dir "llama-sft-lora-deepspeed" \
+--per_device_train_batch_size 8 \
+--per_device_eval_batch_size 8 \
+--gradient_accumulation_steps 4 \
+--gradient_checkpointing True \
+--use_reentrant False \
+--dataset_text_field "content" \
+--use_flash_attn True \
+--use_peft_lora True \
+--lora_r 8 \
+--lora_alpha 16 \
+--lora_dropout 0.1 \
+--lora_target_modules "all-linear" \
+--use_4bit_quantization False
+```
+
+Notice that we are using LoRA with  rank=8, alpha=16 and targeting all linear layers. We are passing the deepspeed config file and finetuning 70B Llama model on a subset of the ultrachat dataset.
+
+## The important parts
+
+Let's dive a little deeper into the script so you can see what's going on, and understand how it works.
+
+The first thing to know is that the script uses DeepSpeed for distributed training as the DeepSpeed config has been passed. The `SFTTrainer` class handles all the heavy lifting of creating the PEFT model using the peft config that is passed. After that, when you call `trainer.train()`, `SFTTrainer` internally uses 🤗 Accelerate to prepare the model, optimizer and trainer using the DeepSpeed config to create DeepSpeed engine which is then trained. The main code snippet is below:
+
+```python
+# trainer
+trainer = SFTTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    args=training_args,
+    train_dataset=train_dataset,
+    eval_dataset=eval_dataset,
+    peft_config=peft_config,
+    packing=data_args.packing,
+    dataset_kwargs={
+        "append_concat_token": data_args.append_concat_token,
+        "add_special_tokens": data_args.add_special_tokens,
+    },
+    dataset_text_field=data_args.dataset_text_field,
+    max_seq_length=data_args.max_seq_length,
+)
+trainer.accelerator.print(f"{trainer.model}")
+
+# train
+checkpoint = None
+if training_args.resume_from_checkpoint is not None:
+    checkpoint = training_args.resume_from_checkpoint
+trainer.train(resume_from_checkpoint=checkpoint)
+
+# saving final model
+trainer.save_model()
+```
+
+## Memory usage
+
+In the above example, the memory consumed per GPU is 64 GB (80%) as seen in the screenshot below:
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/peft_deepspeed_mem_usage.png"/>
+</div>
+<small>GPU memory usage for the training run</small>
+
+## More resources
+You can also refer this blog post [Falcon 180B Finetuning using 🤗 PEFT and DeepSpeed](https://medium.com/@sourabmangrulkar/falcon-180b-finetuning-using-peft-and-deepspeed-b92643091d99) on how to finetune 180B Falcon model on 16 A100 GPUs on 2 machines.
+
+
+# Use PEFT QLoRA and DeepSpeed with ZeRO3 for finetuning large models on multiple GPUs
+
+In this section, we will look at how to use QLoRA and DeepSpeed Stage-3 for finetuning 70B llama model on 2X40GB GPUs.
+For this, we first need `bitsandbytes>=0.43.0`, `accelerate>=0.28.0`, `transformers>4.38.2`, `trl>0.7.11` and `peft>0.9.0`. We need to set `zero3_init_flag` to true when using Accelerate config. Below is the config which can be found at [deepspeed_config_z3_qlora.yaml](https://github.com/huggingface/peft/blob/main/examples/sft/configs/deepspeed_config_z3_qlora.yaml):
+
+```yml
+compute_environment: LOCAL_MACHINE                                                                                                                                           
+debug: false
+deepspeed_config:
+  deepspeed_multinode_launcher: standard
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: true
+  zero3_save_16bit_model: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 2
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+Launch command is given below which is available at [run_peft_qlora_deepspeed_stage3.sh](https://github.com/huggingface/peft/blob/main/examples/sft/run_peft_deepspeed.sh):
+```
+accelerate launch --config_file "configs/deepspeed_config_z3_qlora.yaml"  train.py \
+--seed 100 \
+--model_name_or_path "meta-llama/Llama-2-70b-hf" \
+--dataset_name "smangrul/ultrachat-10k-chatml" \
+--chat_template_format "chatml" \
+--add_special_tokens False \
+--append_concat_token False \
+--splits "train,test" \
+--max_seq_len 2048 \
+--num_train_epochs 1 \
+--logging_steps 5 \
+--log_level "info" \
+--logging_strategy "steps" \
+--eval_strategy "epoch" \
+--save_strategy "epoch" \
+--push_to_hub \
+--hub_private_repo True \
+--hub_strategy "every_save" \
+--bf16 True \
+--packing True \
+--learning_rate 1e-4 \
+--lr_scheduler_type "cosine" \
+--weight_decay 1e-4 \
+--warmup_ratio 0.0 \
+--max_grad_norm 1.0 \
+--output_dir "llama-sft-qlora-dsz3" \
+--per_device_train_batch_size 2 \
+--per_device_eval_batch_size 2 \
+--gradient_accumulation_steps 2 \
+--gradient_checkpointing True \
+--use_reentrant True \
+--dataset_text_field "content" \
+--use_flash_attn True \
+--use_peft_lora True \
+--lora_r 8 \
+--lora_alpha 16 \
+--lora_dropout 0.1 \
+--lora_target_modules "all-linear" \
+--use_4bit_quantization True \
+--use_nested_quant True \
+--bnb_4bit_compute_dtype "bfloat16" \
+--bnb_4bit_quant_storage_dtype "bfloat16"
+```
+
+Notice the new argument being passed `bnb_4bit_quant_storage_dtype` which denotes the data type for packing the 4-bit parameters. For example, when it is set to `bfloat16`, **32/4 = 8** 4-bit params are packed together post quantization.
+
+In terms of training code, the important code changes are: 
+
+```diff
+...
+
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=args.use_4bit_quantization,
+    bnb_4bit_quant_type=args.bnb_4bit_quant_type,
+    bnb_4bit_compute_dtype=compute_dtype,
+    bnb_4bit_use_double_quant=args.use_nested_quant,
++   bnb_4bit_quant_storage=quant_storage_dtype,
+)
+
+...
+
+model = AutoModelForCausalLM.from_pretrained(
+    args.model_name_or_path,
+    quantization_config=bnb_config,
+    trust_remote_code=True,
+    attn_implementation="flash_attention_2" if args.use_flash_attn else "eager",
++   torch_dtype=quant_storage_dtype or torch.float32,
+)
+```
+
+Notice that `torch_dtype` for `AutoModelForCausalLM` is same as the `bnb_4bit_quant_storage` data type. That's it. Everything else is handled by Trainer and TRL.
+
+## Memory usage
+
+In the above example, the memory consumed per GPU is **36.6 GB**. Therefore, what took 8X80GB GPUs with DeepSpeed Stage 3+LoRA and a couple of 80GB GPUs with DDP+QLoRA now requires 2X40GB GPUs. This makes finetuning of large models more accessible.
+
+# Use PEFT and DeepSpeed with ZeRO3 and CPU Offloading for finetuning large models on a single GPU
+This section of guide will help you learn how to use our DeepSpeed [training script](https://github.com/huggingface/peft/blob/main/examples/conditional_generation/peft_lora_seq2seq_accelerate_ds_zero3_offload.py). You'll configure the script to train a large model for conditional generation with ZeRO-3 and CPU Offload.
+
+<Tip>
+
+💡 To help you get started, check out our example training scripts for [causal language modeling](https://github.com/huggingface/peft/blob/main/examples/causal_language_modeling/peft_lora_clm_accelerate_ds_zero3_offload.py) and [conditional generation](https://github.com/huggingface/peft/blob/main/examples/conditional_generation/peft_lora_seq2seq_accelerate_ds_zero3_offload.py). You can adapt these scripts for your own applications or even use them out of the box if your task is similar to the one in the scripts.
+
+</Tip>
+
+## Configuration
+
+Start by running the following command to [create a DeepSpeed configuration file](https://huggingface.co/docs/accelerate/quicktour#launching-your-distributed-script) with 🤗 Accelerate. The `--config_file` flag allows you to save the configuration file to a specific location, otherwise it is saved as a `default_config.yaml` file in the 🤗 Accelerate cache.
+
+The configuration file is used to set the default options when you launch the training script.
+
+```bash
+accelerate config --config_file ds_zero3_cpu.yaml
+```
+
+You'll be asked a few questions about your setup, and configure the following arguments. In this example, you'll use ZeRO-3 along with CPU-Offload so make sure you pick those options.
+
+```bash
+`zero_stage`: [0] Disabled, [1] optimizer state partitioning, [2] optimizer+gradient state partitioning and [3] optimizer+gradient+parameter partitioning
+`gradient_accumulation_steps`: Number of training steps to accumulate gradients before averaging and applying them.
+`gradient_clipping`: Enable gradient clipping with value.
+`offload_optimizer_device`: [none] Disable optimizer offloading, [cpu] offload optimizer to CPU, [nvme] offload optimizer to NVMe SSD. Only applicable with ZeRO >= Stage-2.
+`offload_param_device`: [none] Disable parameter offloading, [cpu] offload parameters to CPU, [nvme] offload parameters to NVMe SSD. Only applicable with ZeRO Stage-3.
+`zero3_init_flag`: Decides whether to enable `deepspeed.zero.Init` for constructing massive models. Only applicable with ZeRO Stage-3.
+`zero3_save_16bit_model`: Decides whether to save 16-bit model weights when using ZeRO Stage-3.
+`mixed_precision`: `no` for FP32 training, `fp16` for FP16 mixed-precision training and `bf16` for BF16 mixed-precision training. 
+```
+
+An example [configuration file](https://github.com/huggingface/peft/blob/main/examples/conditional_generation/accelerate_ds_zero3_cpu_offload_config.yaml) might look like the following. The most important thing to notice is that `zero_stage` is set to `3`, and `offload_optimizer_device` and `offload_param_device` are set to the `cpu`.
+
+```yml
+compute_environment: LOCAL_MACHINE
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  gradient_clipping: 1.0
+  offload_optimizer_device: cpu
+  offload_param_device: cpu
+  zero3_init_flag: true
+  zero3_save_16bit_model: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+dynamo_backend: 'NO'
+fsdp_config: {}
+machine_rank: 0
+main_training_function: main
+megatron_lm_config: {}
+mixed_precision: 'no'
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+use_cpu: false
+```
+
+## The important parts
+
+Let's dive a little deeper into the script so you can see what's going on, and understand how it works.
+
+Within the [`main`](https://github.com/huggingface/peft/blob/2822398fbe896f25d4dac5e468624dc5fd65a51b/examples/conditional_generation/peft_lora_seq2seq_accelerate_ds_zero3_offload.py#L103) function, the script creates an [`~accelerate.Accelerator`] class to initialize all the necessary requirements for distributed training.
+
+<Tip>
+
+💡 Feel free to change the model and dataset inside the `main` function. If your dataset format is different from the one in the script, you may also need to write your own preprocessing function. 
+
+</Tip>
+
+The script also creates a configuration for the 🤗 PEFT method you're using, which in this case, is LoRA. The [`LoraConfig`] specifies the task type and important parameters such as the dimension of the low-rank matrices, the matrices scaling factor, and the dropout probability of the LoRA layers. If you want to use a different 🤗 PEFT method, make sure you replace `LoraConfig` with the appropriate [class](../package_reference/tuners).
+
+```diff
+ def main():
++    accelerator = Accelerator()
+     model_name_or_path = "facebook/bart-large"
+     dataset_name = "twitter_complaints"
++    peft_config = LoraConfig(
+         task_type=TaskType.SEQ_2_SEQ_LM, inference_mode=False, r=8, lora_alpha=32, lora_dropout=0.1
+     )
+```
+
+Throughout the script, you'll see the [`~accelerate.Accelerator.main_process_first`] and [`~accelerate.Accelerator.wait_for_everyone`] functions which help control and synchronize when processes are executed.
+
+The [`get_peft_model`] function takes a base model and the [`peft_config`] you prepared earlier to create a [`PeftModel`]:
+
+```diff
+  model = AutoModelForSeq2SeqLM.from_pretrained(model_name_or_path)
++ model = get_peft_model(model, peft_config)
+```
+
+Pass all the relevant training objects to 🤗 Accelerate's [`~accelerate.Accelerator.prepare`] which makes sure everything is ready for training:
+
+```py
+model, train_dataloader, eval_dataloader, test_dataloader, optimizer, lr_scheduler = accelerator.prepare(
+    model, train_dataloader, eval_dataloader, test_dataloader, optimizer, lr_scheduler
+)
+```
+
+The next bit of code checks whether the DeepSpeed plugin is used in the `Accelerator`, and if the plugin exists, then we check if we are using ZeRO-3. This conditional flag is used when calling `generate` function call during inference for syncing GPUs when the model parameters are sharded:
+
+```py
+is_ds_zero_3 = False
+if getattr(accelerator.state, "deepspeed_plugin", None):
+    is_ds_zero_3 = accelerator.state.deepspeed_plugin.zero_stage == 3
+```
+
+Inside the training loop, the usual `loss.backward()` is replaced by 🤗 Accelerate's [`~accelerate.Accelerator.backward`] which uses the correct `backward()` method based on your configuration:
+
+```diff
+  for epoch in range(num_epochs):
+      with TorchTracemalloc() as tracemalloc:
+          model.train()
+          total_loss = 0
+          for step, batch in enumerate(tqdm(train_dataloader)):
+              outputs = model(**batch)
+              loss = outputs.loss
+              total_loss += loss.detach().float()
++             accelerator.backward(loss)
+              optimizer.step()
+              lr_scheduler.step()
+              optimizer.zero_grad()
+```
+
+That is all! The rest of the script handles the training loop, evaluation, and even pushes it to the Hub for you.
+
+## Train
+
+Run the following command to launch the training script. Earlier, you saved the configuration file to `ds_zero3_cpu.yaml`, so you'll need to pass the path to the launcher with the `--config_file` argument like this:
+
+```bash
+accelerate launch --config_file ds_zero3_cpu.yaml examples/peft_lora_seq2seq_accelerate_ds_zero3_offload.py
+```
+
+You'll see some output logs that track memory usage during training, and once it's completed, the script returns the accuracy and compares the predictions to the labels:
+
+```bash
+GPU Memory before entering the train : 1916
+GPU Memory consumed at the end of the train (end-begin): 66
+GPU Peak Memory consumed during the train (max-begin): 7488
+GPU Total Peak Memory consumed during the train (max): 9404
+CPU Memory before entering the train : 19411
+CPU Memory consumed at the end of the train (end-begin): 0
+CPU Peak Memory consumed during the train (max-begin): 0
+CPU Total Peak Memory consumed during the train (max): 19411
+epoch=4: train_ppl=tensor(1.0705, device='cuda:0') train_epoch_loss=tensor(0.0681, device='cuda:0')
+100%|████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:27<00:00,  3.92s/it]
+GPU Memory before entering the eval : 1982
+GPU Memory consumed at the end of the eval (end-begin): -66
+GPU Peak Memory consumed during the eval (max-begin): 672
+GPU Total Peak Memory consumed during the eval (max): 2654
+CPU Memory before entering the eval : 19411
+CPU Memory consumed at the end of the eval (end-begin): 0
+CPU Peak Memory consumed during the eval (max-begin): 0
+CPU Total Peak Memory consumed during the eval (max): 19411
+accuracy=100.0
+eval_preds[:10]=['no complaint', 'no complaint', 'complaint', 'complaint', 'no complaint', 'no complaint', 'no complaint', 'complaint', 'complaint', 'no complaint']
+dataset['train'][label_column][:10]=['no complaint', 'no complaint', 'complaint', 'complaint', 'no complaint', 'no complaint', 'no complaint', 'complaint', 'complaint', 'no complaint']
+```
+
+# Caveats
+1. Merging when using PEFT and DeepSpeed is currently unsupported and will raise error.
+2. When using CPU offloading, the major gains from using PEFT to shrink the optimizer states and gradients to that of the adapter weights would be realized on CPU RAM and there won't be savings with respect to GPU memory.
+3. DeepSpeed Stage 3 and qlora when used with CPU offloading leads to more GPU memory usage when compared to disabling CPU offloading. 

docs/source/accelerate/fsdp.md

@@ -0,0 +1,292 @@
+<!--⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# Fully Sharded Data Parallel
+
+[Fully sharded data parallel](https://pytorch.org/docs/stable/fsdp.html) (FSDP) is developed for distributed training of large pretrained models up to 1T parameters. FSDP achieves this by sharding the model parameters, gradients, and optimizer states across data parallel processes and it can also offload sharded model parameters to a CPU. The memory efficiency afforded by FSDP allows you to scale training to larger batch or model sizes.
+
+Both of these features are supported in 🤗 Accelerate, and you can use them with 🤗 PEFT. 
+
+# Use PEFT and FSDP
+This section of guide will help you learn how to use our DeepSpeed [training script](https://github.com/huggingface/peft/blob/main/examples/sft/train.py) for performing SFT. You'll configure the script to do SFT (supervised fine-tuning) of Llama-70B model with LoRA and FSDP on 8xH100 80GB GPUs on a single machine. You can configure it to scale to multiple machines by changing the accelerate config.
+
+## Configuration
+
+Start by running the following command to [create a FSDP configuration file](https://huggingface.co/docs/accelerate/quicktour#launching-your-distributed-script) with 🤗 Accelerate. The `--config_file` flag allows you to save the configuration file to a specific location, otherwise it is saved as a `default_config.yaml` file in the 🤗 Accelerate cache.
+
+The configuration file is used to set the default options when you launch the training script.
+
+```bash
+accelerate config --config_file fsdp_config.yaml
+```
+
+You'll be asked a few questions about your setup, and configure the following arguments. In this example, you'll answer the questionnaire as shown in the image below.
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/fsdp-peft-config.png"/>
+</div>
+<small>Creating Accelerate's config to use FSDP</small>
+
+Once this is done, the corresponding config should look like below and you can find it in config folder at [fsdp_config.yaml](https://github.com/huggingface/peft/blob/main/examples/sft/configs/fsdp_config.yaml):
+
+```yml
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_backward_prefetch: BACKWARD_PRE
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: false
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+## Launch command
+
+The launch command is available at [run_peft_fsdp.sh](https://github.com/huggingface/peft/blob/main/examples/sft/run_peft_fsdp.sh) and it is also shown below:
+```bash
+accelerate launch --config_file "configs/fsdp_config.yaml"  train.py \
+--seed 100 \
+--model_name_or_path "meta-llama/Llama-2-70b-hf" \
+--dataset_name "smangrul/ultrachat-10k-chatml" \
+--chat_template_format "chatml" \
+--add_special_tokens False \
+--append_concat_token False \
+--splits "train,test" \
+--max_seq_len 2048 \
+--num_train_epochs 1 \
+--logging_steps 5 \
+--log_level "info" \
+--logging_strategy "steps" \
+--eval_strategy "epoch" \
+--save_strategy "epoch" \
+--push_to_hub \
+--hub_private_repo True \
+--hub_strategy "every_save" \
+--bf16 True \
+--packing True \
+--learning_rate 1e-4 \
+--lr_scheduler_type "cosine" \
+--weight_decay 1e-4 \
+--warmup_ratio 0.0 \
+--max_grad_norm 1.0 \
+--output_dir "llama-sft-lora-fsdp" \
+--per_device_train_batch_size 8 \
+--per_device_eval_batch_size 8 \
+--gradient_accumulation_steps 4 \
+--gradient_checkpointing True \
+--use_reentrant False \
+--dataset_text_field "content" \
+--use_flash_attn True \
+--use_peft_lora True \
+--lora_r 8 \
+--lora_alpha 16 \
+--lora_dropout 0.1 \
+--lora_target_modules "all-linear" \
+--use_4bit_quantization False
+```
+
+Notice that we are using LoRA with  rank=8, alpha=16 and targeting all linear layers. We are passing the FSDP config file and finetuning the 70B Llama model on a subset of the [ultrachat dataset](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k).
+
+## The important parts
+
+Let's dive a little deeper into the script so you can see what's going on, and understand how it works.
+
+The first thing to know is that the script uses FSDP for distributed training as the FSDP config has been passed. The `SFTTrainer` class handles all the heavy lifting of creating PEFT model using the peft config that is passed. After that when you call `trainer.train()`, Trainer internally uses 🤗 Accelerate to prepare model, optimizer and trainer using the FSDP config to create FSDP wrapped model which is then trained. The main code snippet is below:
+
+```python
+# trainer
+trainer = SFTTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    args=training_args,
+    train_dataset=train_dataset,
+    eval_dataset=eval_dataset,
+    peft_config=peft_config,
+    packing=data_args.packing,
+    dataset_kwargs={
+        "append_concat_token": data_args.append_concat_token,
+        "add_special_tokens": data_args.add_special_tokens,
+    },
+    dataset_text_field=data_args.dataset_text_field,
+    max_seq_length=data_args.max_seq_length,
+)
+trainer.accelerator.print(f"{trainer.model}")
+if model_args.use_peft_lora:
+    # handle PEFT+FSDP case
+    trainer.model.print_trainable_parameters()
+    if getattr(trainer.accelerator.state, "fsdp_plugin", None):
+        from peft.utils.other import fsdp_auto_wrap_policy
+
+        fsdp_plugin = trainer.accelerator.state.fsdp_plugin
+        fsdp_plugin.auto_wrap_policy = fsdp_auto_wrap_policy(trainer.model)
+
+# train
+checkpoint = None
+if training_args.resume_from_checkpoint is not None:
+    checkpoint = training_args.resume_from_checkpoint
+trainer.train(resume_from_checkpoint=checkpoint)
+
+# saving final model
+if trainer.is_fsdp_enabled:
+    trainer.accelerator.state.fsdp_plugin.set_state_dict_type("FULL_STATE_DICT")
+trainer.save_model()
+```
+
+
+Here, one main thing to note currently when using FSDP with PEFT is that `use_orig_params` needs to be `False` to realize GPU memory savings. Due to `use_orig_params=False`, the auto wrap policy for FSDP needs to change so that trainable and non-trainable parameters are wrapped separately. This is done by the code snippt below which uses the util function `fsdp_auto_wrap_policy` from PEFT:
+
+```
+if getattr(trainer.accelerator.state, "fsdp_plugin", None):
+    from peft.utils.other import fsdp_auto_wrap_policy
+
+    fsdp_plugin = trainer.accelerator.state.fsdp_plugin
+    fsdp_plugin.auto_wrap_policy = fsdp_auto_wrap_policy(trainer.model)
+```
+
+## Memory usage
+
+In the above example, the memory consumed per GPU is  72-80 GB (90-98%) as seen in the screenshot below. The slight increase in GPU memory at the end is when saving the model using `FULL_STATE_DICT` state dict type instead of the `SHARDED_STATE_DICT` so that the model has adapter weights that can be loaded normally with `from_pretrained` method during inference:
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/peft_fsdp_mem_usage.png"/>
+</div>
+<small>GPU memory usage for the training run</small>
+
+# Use PEFT QLoRA and FSDP for finetuning large models on multiple GPUs
+
+In this section, we will look at how to use QLoRA and FSDP for finetuning 70B llama model on 2X24GB GPUs. [Answer.AI](https://www.answer.ai/) in collaboration with bitsandbytes and Hugging Face 🤗 open sourced code enabling the usage of FSDP+QLoRA and explained the whole process in their insightful blogpost [You can now train a 70b language model at home](https://www.answer.ai/posts/2024-03-06-fsdp-qlora.html). This is now integrated in Hugging Face ecosystem. 
+
+For this, we first need `bitsandbytes>=0.43.0`, `accelerate>=0.28.0`, `transformers>4.38.2`, `trl>0.7.11` and `peft>0.9.0`. We need to set `fsdp_cpu_ram_efficient_loading=true`, `fsdp_use_orig_params=false` and `fsdp_offload_params=true`(cpu offloading) when using Accelerate config. When not using accelerate launcher, you can alternately set the environment variable `export FSDP_CPU_RAM_EFFICIENT_LOADING=true`.  Here, we will be using accelerate config and below is the config which can be found at [fsdp_config_qlora.yaml](https://github.com/huggingface/peft/blob/main/examples/sft/configs/fsdp_config_qlora.yaml):
+
+```yml
+compute_environment: LOCAL_MACHINE                                                                                                                                           
+debug: false                                                                                                                                                                 
+distributed_type: FSDP
+downcast_bf16: 'no'
+fsdp_config:
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_backward_prefetch: BACKWARD_PRE
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_forward_prefetch: false
+  fsdp_offload_params: true
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
+  fsdp_use_orig_params: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'no'
+num_machines: 1
+num_processes: 2
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+Launch command is given below which is available at [run_peft_qlora_fsdp.sh](https://github.com/huggingface/peft/blob/main/examples/sft/run_peft_qlora_fsdp.sh):
+```
+accelerate launch --config_file "configs/fsdp_config_qlora.yaml"  train.py \
+--seed 100 \
+--model_name_or_path "meta-llama/Llama-2-70b-hf" \
+--dataset_name "smangrul/ultrachat-10k-chatml" \
+--chat_template_format "chatml" \
+--add_special_tokens False \
+--append_concat_token False \
+--splits "train,test" \
+--max_seq_len 2048 \
+--num_train_epochs 1 \
+--logging_steps 5 \
+--log_level "info" \
+--logging_strategy "steps" \
+--eval_strategy "epoch" \
+--save_strategy "epoch" \
+--push_to_hub \
+--hub_private_repo True \
+--hub_strategy "every_save" \
+--bf16 True \
+--packing True \
+--learning_rate 1e-4 \
+--lr_scheduler_type "cosine" \
+--weight_decay 1e-4 \
+--warmup_ratio 0.0 \
+--max_grad_norm 1.0 \
+--output_dir "llama-sft-qlora-fsdp" \
+--per_device_train_batch_size 2 \
+--per_device_eval_batch_size 2 \
+--gradient_accumulation_steps 2 \
+--gradient_checkpointing True \
+--use_reentrant True \
+--dataset_text_field "content" \
+--use_flash_attn True \
+--use_peft_lora True \
+--lora_r 8 \
+--lora_alpha 16 \
+--lora_dropout 0.1 \
+--lora_target_modules "all-linear" \
+--use_4bit_quantization True \
+--use_nested_quant True \
+--bnb_4bit_compute_dtype "bfloat16" \
+--bnb_4bit_quant_storage_dtype "bfloat16"
+```
+
+Notice the new argument being passed, `bnb_4bit_quant_storage_dtype`, which denotes the data type for packing the 4-bit parameters. For example, when it is set to `bfloat16`, **16/4 = 4** 4-bit params are packed together post quantization. When using mixed precision training with `bfloat16`, `bnb_4bit_quant_storage_dtype` can be either `bfloat16` for pure `bfloat16` finetuning, or `float32` for automatic mixed precision (this consumes more GPU memory). When using mixed precision training with `float16`, `bnb_4bit_quant_storage_dtype` should be set to `float32` for stable automatic mixed precision training.
+
+In terms of training code, the important code changes are: 
+
+```diff
+...
+
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=args.use_4bit_quantization,
+    bnb_4bit_quant_type=args.bnb_4bit_quant_type,
+    bnb_4bit_compute_dtype=compute_dtype,
+    bnb_4bit_use_double_quant=args.use_nested_quant,
++   bnb_4bit_quant_storage=quant_storage_dtype,
+)
+
+...
+
+model = AutoModelForCausalLM.from_pretrained(
+    args.model_name_or_path,
+    quantization_config=bnb_config,
+    trust_remote_code=True,
+    attn_implementation="flash_attention_2" if args.use_flash_attn else "eager",
++   torch_dtype=quant_storage_dtype or torch.float32,
+)
+```
+
+Notice that `torch_dtype` for `AutoModelForCausalLM` is same as the `bnb_4bit_quant_storage` data type. That's it. Everything else is handled by Trainer and TRL.
+
+## Memory usage
+
+In the above example, the memory consumed per GPU is **19.6 GB** while CPU RAM usage is around **107 GB**. When disabling CPU offloading, the GPU memory usage is  **35.6 GB/ GPU**. Therefore, what took 16X80GB GPUs for full finetuning, 8X80GB GPUs with FSDP+LoRA, and a couple of 80GB GPUs with DDP+QLoRA, now requires 2X24GB GPUs. This makes finetuning of large models more accessible.
+
+## More resources
+You can also refer the [llama-recipes](https://github.com/facebookresearch/llama-recipes/?tab=readme-ov-file#fine-tuning) repo and [Getting started with Llama](https://llama.meta.com/get-started/#fine-tuning) guide on how to finetune using FSDP and PEFT.
+
+## Caveats
+1. Merging when using PEFT and FSDP is currently unsupported and will raise error.
+2. Passing `modules_to_save` config parameter to is untested at present.
+3. GPU Memory saving when using CPU Offloading is untested at present.
+4. When using FSDP+QLoRA, `paged_adamw_8bit` currently results in an error when saving a checkpoint.
+5. DoRA training with FSDP should work (albeit at lower speed than LoRA). If combined with bitsandbytes (QDoRA), 4-bit quantization should also work, but 8-bit quantization has known issues and is not recommended.

docs/source/conceptual_guides/adapter.md

@@ -0,0 +1,107 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Adapters
+
+Adapter-based methods add extra trainable parameters after the attention and fully-connected layers of a frozen pretrained model to reduce memory-usage and speed up training. The method varies depending on the adapter, it could simply be an extra added layer or it could be expressing the weight updates ∆W as a low-rank decomposition of the weight matrix. Either way, the adapters are typically small but demonstrate comparable performance to a fully finetuned model and enable training larger models with fewer resources.
+
+This guide will give you a brief overview of the adapter methods supported by PEFT (if you're interested in learning more details about a specific method, take a look at the linked paper).
+
+## Low-Rank Adaptation (LoRA)
+
+<Tip>
+
+LoRA is one of the most popular PEFT methods and a good starting point if you're just getting started with PEFT. It was originally developed for large language models but it is a tremendously popular training method for diffusion models because of its efficiency and effectiveness.
+
+</Tip>
+
+As mentioned briefly earlier, [LoRA](https://hf.co/papers/2106.09685) is a technique that accelerates finetuning large models while consuming less memory.
+
+LoRA represents the weight updates ∆W with two smaller matrices (called *update matrices*) through low-rank decomposition. These new matrices can be trained to adapt to the new data while keeping the overall number of parameters low. The original weight matrix remains frozen and doesn't receive any further updates. To produce the final results, the original and extra adapted weights are combined. You could also merge the adapter weights with the base model to eliminate inference latency.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/lora_animated.gif"/>
+</div>
+
+This approach has a number of advantages:
+
+* LoRA makes finetuning more efficient by drastically reducing the number of trainable parameters.
+* The original pretrained weights are kept frozen, which means you can have multiple lightweight and portable LoRA models for various downstream tasks built on top of them.
+* LoRA is orthogonal to other parameter-efficient methods and can be combined with many of them.
+* Performance of models finetuned using LoRA is comparable to the performance of fully finetuned models.
+
+In principle, LoRA can be applied to any subset of weight matrices in a neural network to reduce the number of trainable parameters. However, for simplicity and further parameter efficiency, LoRA is typically only applied to the attention blocks in Transformer models. The resulting number of trainable parameters in a LoRA model depends on the size of the update matrices, which is determined mainly by the rank `r` and the shape of the original weight matrix.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/lora.png"/>
+</div>
+<small><a href="https://hf.co/papers/2103.10385">Navigating Text-To-Image Customization: From LyCORIS Fine-Tuning to Model Evaluation</a></small>
+
+## Mixture of LoRA Experts (X-LoRA)
+
+[X-LoRA](https://arxiv.org/abs/2402.07148) is a mixture of experts method for LoRA which works by using dense or sparse gating to dynamically activate LoRA experts. The LoRA experts as well as the base model are frozen during training, resulting in a low parameter count as only the gating layers must be trained. In particular, the gating layers output scalings which (depending on config) are granular on the layer and token level. Additionally, during inference, X-LoRA dynamically activates LoRA adapters to recall knowledge and effectively mix them:
+
+The below graphic demonstrates how the scalings change for different prompts for each token. This highlights the activation of different adapters as the generation progresses and the sequence creates new context.
+
+![Token-by-token scalings](https://github.com/EricLBuehler/xlora/raw/master/res/token_by_token_scalings.gif)
+
+For each step, X-LoRA requires the base model to be run twice: first, to get hidden states without any LoRA adapters, and secondly, the hidden states are used to calculate scalings which are applied to the LoRA adapters and the model is run a second time. The output of the second run is the result of the model step.
+
+Ultimately, X-LoRA allows the model to reflect upon it's knowledge because of the dual forward pass scheme, and dynamically reconfigure the architecture.
+
+## Low-Rank Hadamard Product (LoHa)
+
+Low-rank decomposition can impact performance because the weight updates are limited to the low-rank space, which can constrain a model's expressiveness. However, you don't necessarily want to use a larger rank because it increases the number of trainable parameters. To address this, [LoHa](https://huggingface.co/papers/2108.06098) (a method originally developed for computer vision) was applied to diffusion models where the ability to generate diverse images is an important consideration. LoHa should also work with general model types, but the embedding layers aren't currently implemented in PEFT.
+
+LoHa uses the [Hadamard product](https://en.wikipedia.org/wiki/Hadamard_product_(matrices)) (element-wise product) instead of the matrix product. ∆W is represented by four smaller matrices instead of two - like in LoRA - and each pair of these low-rank matrices are combined with the Hadamard product. As a result, ∆W can have the same number of trainable parameters but a higher rank and expressivity.
+
+## Low-Rank Kronecker Product (LoKr)
+
+[LoKr](https://hf.co/papers/2309.14859) is very similar to LoRA and LoHa, and it is also mainly applied to diffusion models, though you could also use it with other model types. LoKr replaces the matrix product with the [Kronecker product](https://en.wikipedia.org/wiki/Kronecker_product) instead. The Kronecker product decomposition creates a block matrix which preserves the rank of the original weight matrix. Another benefit of the Kronecker product is that it can be vectorized by stacking the matrix columns. This can speed up the process because you're avoiding fully reconstructing ∆W.
+
+## Orthogonal Finetuning (OFT)
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/oft.png"/>
+</div>
+<small><a href="https://hf.co/papers/2306.07280">Controlling Text-to-Image Diffusion by Orthogonal Finetuning</a></small>
+
+[OFT](https://hf.co/papers/2306.07280) is a method that primarily focuses on preserving a pretrained model's generative performance in the finetuned model. It tries to maintain the same cosine similarity (hyperspherical energy) between all pairwise neurons in a layer because this better captures the semantic information among neurons. This means OFT is more capable at preserving the subject and it is better for controllable generation (similar to [ControlNet](https://huggingface.co/docs/diffusers/using-diffusers/controlnet)).
+
+OFT preserves the hyperspherical energy by learning an orthogonal transformation for neurons to keep the cosine similarity between them unchanged. In practice, this means taking the matrix product of an orthogonal matrix with the pretrained weight matrix. However, to be parameter-efficient, the orthogonal matrix is represented as a block-diagonal matrix with rank `r` blocks. Whereas LoRA reduces the number of trainable parameters with low-rank structures, OFT reduces the number of trainable parameters with a sparse block-diagonal matrix structure.
+
+## Orthogonal Butterfly (BOFT)
+
+[BOFT](https://hf.co/papers/2311.06243) is a method that primarily focuses on preserving a pretrained model's generative performance in the finetuned model. It tries to maintain the same cosine similarity (hyperspherical energy) between all pairwise neurons in a layer because this better captures the semantic information among neurons. This means OFT is more capable at preserving the subject and it is better for controllable generation (similar to [ControlNet](https://huggingface.co/docs/diffusers/using-diffusers/controlnet)).
+
+OFT preserves the hyperspherical energy by learning an orthogonal transformation for neurons to keep the cosine similarity between them unchanged. In practice, this means taking the matrix product of an orthogonal matrix with the pretrained weight matrix. However, to be parameter-efficient, the orthogonal matrix is represented as a block-diagonal matrix with rank `r` blocks. Whereas LoRA reduces the number of trainable parameters with low-rank structures, OFT reduces the number of trainable parameters with a sparse block-diagonal matrix structure.
+
+## Adaptive Low-Rank Adaptation (AdaLoRA)
+
+[AdaLoRA](https://hf.co/papers/2303.10512) manages the parameter budget introduced from LoRA by allocating more parameters - in other words, a higher rank `r` - for important weight matrices that are better adapted for a task and pruning less important ones. The rank is controlled by a method similar to singular value decomposition (SVD). The ∆W is parameterized with two orthogonal matrices and a diagonal matrix which contains singular values. This parametrization method avoids iteratively applying SVD which is computationally expensive. Based on this method, the rank of ∆W is adjusted according to an importance score. ∆W is divided into triplets and each triplet is scored according to its contribution to model performance. Triplets with low importance scores are pruned and triplets with high importance scores are kept for finetuning.
+
+## Llama-Adapter
+
+[Llama-Adapter](https://hf.co/papers/2303.16199) is a method for adapting Llama into a instruction-following model. To help adapt the model for instruction-following, the adapter is trained with a 52K instruction-output dataset.
+
+A set of of learnable adaption prompts are prefixed to the input instruction tokens. These are inserted into the upper layers of the model because it is better to learn with the higher-level semantics of the pretrained model. The instruction-output tokens prefixed to the input guide the adaption prompt to generate a contextual response.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/llama-adapter.png"/>
+</div>
+<small><a href="https://hf.co/papers/2303.16199">LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention</a></small>
+
+To avoid adding noise to the tokens, the adapter uses zero-initialized attention. On top of this, the adapter adds a learnable gating factor (initialized with zeros) to progressively add information to the model during training. This prevents overwhelming the model's pretrained knowledge with the newly learned instructions.

docs/source/conceptual_guides/ia3.md

@@ -0,0 +1,68 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# IA3 
+
+This conceptual guide gives a brief overview of [IA3](https://arxiv.org/abs/2205.05638), a parameter-efficient fine tuning technique that is 
+intended to improve over [LoRA](./lora).
+
+To make fine-tuning more efficient, IA3 (Infused Adapter by Inhibiting and Amplifying Inner Activations) 
+rescales inner activations with learned vectors. These learned vectors are injected in the attention and feedforward modules 
+in a typical transformer-based architecture. These learned vectors are the only trainable parameters during fine-tuning, and thus the original 
+weights remain frozen. Dealing with learned vectors (as opposed to learned low-rank updates to a weight matrix like LoRA)
+keeps the number of trainable parameters much smaller. 
+
+Being similar to LoRA, IA3 carries many of the same advantages: 
+
+* IA3 makes fine-tuning more efficient by drastically reducing the number of trainable parameters. (For T0, an IA3 model only has about 0.01% trainable parameters, while even LoRA has > 0.1%)
+* The original pre-trained weights are kept frozen, which means you can have multiple lightweight and portable IA3 models for various downstream tasks built on top of them.
+* Performance of models fine-tuned using IA3 is comparable to the performance of fully fine-tuned models.
+* IA3 does not add any inference latency because adapter weights can be merged with the base model.
+
+In principle, IA3 can be applied to any subset of weight matrices in a neural network to reduce the number of trainable
+parameters. Following the authors' implementation, IA3 weights are added to the key, value and feedforward layers
+of a Transformer model. To be specific, for transformer models, IA3 weights are added to the outputs of key and value layers, and to the input of the second feedforward layer
+in each transformer block.
+
+Given the target layers for injecting IA3 parameters, the number of trainable parameters
+can be determined based on the size of the weight matrices.
+
+
+## Common IA3 parameters in PEFT
+
+As with other methods supported by PEFT, to fine-tune a model using IA3, you need to:
+
+1. Instantiate a base model.
+2. Create a configuration (`IA3Config`) where you define IA3-specific parameters.
+3. Wrap the base model with `get_peft_model()` to get a trainable `PeftModel`.
+4. Train the `PeftModel` as you normally would train the base model.
+
+`IA3Config` allows you to control how IA3 is applied to the base model through the following parameters:
+
+- `target_modules`: The modules (for example, attention blocks) to apply the IA3 vectors.
+- `feedforward_modules`: The list of modules to be treated as feedforward layers in `target_modules`. While learned vectors are multiplied with
+the output activation for attention blocks, the vectors are multiplied with the input for classic feedforward layers. Note that `feedforward_modules` must be a subset of `target_modules`.
+- `modules_to_save`: List of modules apart from IA3 layers to be set as trainable and saved in the final checkpoint. These typically include model's custom head that is randomly initialized for the fine-tuning task.
+
+## Example Usage
+
+For the task of sequence classification, one can initialize the IA3 config for a Llama model as follows:
+
+```py
+peft_config = IA3Config(
+    task_type=TaskType.SEQ_CLS, target_modules=["k_proj", "v_proj", "down_proj"], feedforward_modules=["down_proj"]
+)
+```

docs/source/conceptual_guides/oft.md

@@ -0,0 +1,107 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Orthogonal Finetuning (OFT and BOFT) 
+
+This conceptual guide gives a brief overview of [OFT](https://arxiv.org/abs/2306.07280) and [BOFT](https://arxiv.org/abs/2311.06243), a parameter-efficient fine-tuning technique that utilizes orthogonal matrix to multiplicatively transform the pretrained weight matrices.
+
+To achieve efficient fine-tuning, OFT represents the weight updates with an orthogonal transformation. The orthogonal transformation is parameterized by an orthogonal matrix multiplied to the pretrained weight matrix. These new matrices can be trained to adapt to the new data while keeping the overall number of changes low. The original weight matrix remains frozen and doesn’t receive any further adjustments. To produce the final results, both the original and the adapted weights are multiplied togethor.
+
+Orthogonal Butterfly (BOFT) generalizes OFT with Butterfly factorization and further improves its parameter efficiency and finetuning flexibility. In short, OFT can be viewed as a special case of BOFT. Different from LoRA that uses additive low-rank weight updates, BOFT uses multiplicative orthogonal weight updates. The comparison is shown below.
+
+<div class="flex justify-center">
+    <img src="https://raw.githubusercontent.com/wy1iu/butterfly-oft/main/assets/BOFT_comparison.png"/>
+</div>
+
+
+BOFT has some advantages compared to LoRA: 
+
+* BOFT proposes a simple yet generic way to finetune pretrained models to downstream tasks, yielding a better preservation of pretraining knowledge and a better parameter efficiency.
+* Through the orthogonality, BOFT introduces a structural constraint, i.e., keeping the [hyperspherical energy](https://arxiv.org/abs/1805.09298) unchanged during finetuning. This can effectively reduce the forgetting of pretraining knowledge.
+* BOFT uses the butterfly factorization to efficiently parameterize the orthogonal matrix, which yields a compact yet expressive learning space (i.e., hypothesis class).
+* The sparse matrix decomposition in BOFT brings in additional inductive biases that are beneficial to generalization.
+
+In principle, BOFT can be applied to any subset of weight matrices in a neural network to reduce the number of trainable parameters. Given the target layers for injecting BOFT parameters, the number of trainable parameters can be determined based on the size of the weight matrices.
+
+## Merge OFT/BOFT weights into the base model
+
+Similar to LoRA, the weights learned by OFT/BOFT can be integrated into the pretrained weight matrices using the merge_and_unload() function. This function merges the adapter weights with the base model which allows you to effectively use the newly merged model as a standalone model.
+
+<div class="flex justify-center">
+    <img src="https://raw.githubusercontent.com/wy1iu/butterfly-oft/main/assets/boft_merge.png"/>
+</div>
+
+This works because during training, the orthogonal weight matrix (R in the diagram above) and the pretrained weight matrices are separate. But once training is complete, these weights can actually be merged (multiplied) into a new weight matrix that is equivalent.
+
+## Utils for OFT / BOFT
+
+### Common OFT / BOFT parameters in PEFT
+
+As with other methods supported by PEFT, to fine-tune a model using OFT or BOFT, you need to:
+
+1. Instantiate a base model.
+2. Create a configuration (`OFTConfig` or `BOFTConfig`) where you define OFT/BOFT-specific parameters.
+3. Wrap the base model with `get_peft_model()` to get a trainable `PeftModel`.
+4. Train the `PeftModel` as you normally would train the base model.
+
+
+### BOFT-specific paramters
+
+`BOFTConfig` allows you to control how OFT/BOFT is applied to the base model through the following parameters:
+
+- `boft_block_size`: the BOFT matrix block size across different layers, expressed in `int`. Smaller block size results in sparser update matrices with fewer trainable paramters. **Note**, please choose `boft_block_size` to be divisible by most layer's input dimension (`in_features`), e.g., 4, 8, 16. Also, please only 
+specify either `boft_block_size` or `boft_block_num`, but not both simultaneously or leaving both to 0, because `boft_block_size` x `boft_block_num` must equal the layer's input dimension.
+- `boft_block_num`: the number of BOFT matrix blocks across different layers, expressed in `int`. Fewer blocks result in sparser update matrices with fewer trainable paramters. **Note**, please choose `boft_block_num` to be divisible by most layer's input dimension (`in_features`), e.g., 4, 8, 16. Also, please only 
+specify either `boft_block_size` or `boft_block_num`, but not both simultaneously or leaving both to 0, because `boft_block_size` x `boft_block_num` must equal the layer's input dimension.
+- `boft_n_butterfly_factor`: the number of butterfly factors. **Note**, for `boft_n_butterfly_factor=1`, BOFT is the same as vanilla OFT, for `boft_n_butterfly_factor=2`, the effective block size of OFT becomes twice as big and the number of blocks become half.
+- `bias`: specify if the `bias` parameters should be trained. Can be `"none"`, `"all"` or `"boft_only"`.
+- `boft_dropout`: specify the probability of multiplicative dropout.
+- `target_modules`: The modules (for example, attention blocks) to inject the OFT/BOFT matrices.
+- `modules_to_save`: List of modules apart from OFT/BOFT matrices to be set as trainable and saved in the final checkpoint. These typically include model's custom head that is randomly initialized for the fine-tuning task.
+
+
+
+## BOFT Example Usage
+
+For an example of the BOFT method application to various downstream tasks, please refer to the following guides:
+
+Take a look at the following step-by-step guides on how to finetune a model with BOFT:
+- [Dreambooth finetuning with BOFT](../task_guides/boft_dreambooth) 
+- [Controllable generation finetuning with BOFT (ControlNet)](../task_guides/boft_controlnet) 
+
+For the task of image classification, one can initialize the BOFT config for a DinoV2 model as follows:
+
+```py
+import transformers
+from transformers import AutoModelForSeq2SeqLM, BOFTConfig
+from peft import BOFTConfig, get_peft_model
+
+config = BOFTConfig(
+    boft_block_size=4,
+    boft_n_butterfly_factor=2,
+    target_modules=["query", "value", "key", "output.dense", "mlp.fc1", "mlp.fc2"],
+    boft_dropout=0.1,
+    bias="boft_only",
+    modules_to_save=["classifier"],
+)
+
+model = transformers.Dinov2ForImageClassification.from_pretrained(
+    "facebook/dinov2-large",
+    num_labels=100,
+)
+
+boft_model = get_peft_model(model, config)
+```

docs/source/conceptual_guides/prompting.md

@@ -0,0 +1,77 @@
+<!--⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# Soft prompts
+
+Training large pretrained language models is very time-consuming and compute-intensive. As they continue to grow in size, there is increasing interest in more efficient training methods such as *prompting*. Prompting primes a frozen pretrained model for a specific downstream task by including a text prompt that describes the task or even demonstrates an example of the task. With prompting, you can avoid fully training a separate model for each downstream task, and use the same frozen pretrained model instead. This is a lot easier because you can use the same model for several different tasks, and it is significantly more efficient to train and store a smaller set of prompt parameters than to train all the model's parameters.
+
+There are two categories of prompting methods:
+
+- hard prompts are manually handcrafted text prompts with discrete input tokens; the downside is that it requires a lot of effort to create a good prompt
+- soft prompts are learnable tensors concatenated with the input embeddings that can be optimized to a dataset; the downside is that they aren't human readable because you aren't matching these "virtual tokens" to the embeddings of a real word
+
+This conceptual guide provides a brief overview of the soft prompt methods included in 🤗 PEFT: prompt tuning, prefix tuning, P-tuning, and multitask prompt tuning.
+
+## Prompt tuning
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/prompt-tuning.png"/>
+</div>
+<small>Only train and store a significantly smaller set of task-specific prompt parameters <a href="https://hf.co/papers/2104.08691">(image source)</a>.</small>
+
+[Prompt tuning](https://hf.co/papers/2104.08691) was developed for text classification tasks on T5 models, and all downstream tasks are cast as a text generation task. For example, sequence classification usually assigns a single class label to a sequence of text. By casting it as a text generation task, the tokens that make up the class label are *generated*. Prompts are added to the input as a series of tokens. Typically, the model parameters are fixed which means the prompt tokens are also fixed by the model parameters.
+
+The key idea behind prompt tuning is that prompt tokens have their own parameters that are updated independently. This means you can keep the pretrained model's parameters frozen, and only update the gradients of the prompt token embeddings. The results are comparable to the traditional method of training the entire model, and prompt tuning performance scales as model size increases.
+
+Take a look at [Prompt tuning for causal language modeling](../task_guides/clm-prompt-tuning) for a step-by-step guide on how to train a model with prompt tuning.
+
+## Prefix tuning
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/prefix-tuning.png"/>
+</div>
+<small>Optimize the prefix parameters for each task <a href="https://hf.co/papers/2101.00190">(image source)</a>.</small>
+
+[Prefix tuning](https://hf.co/papers/2101.00190) was designed for natural language generation (NLG) tasks on GPT models. It is very similar to prompt tuning; prefix tuning also prepends a sequence of task-specific vectors to the input that can be trained and updated while keeping the rest of the pretrained model's parameters frozen. 
+
+The main difference is that the prefix parameters are inserted in **all** of the model layers, whereas prompt tuning only adds the prompt parameters to the model input embeddings. The prefix parameters are also optimized by a separate feed-forward network (FFN) instead of training directly on the soft prompts because it causes instability and hurts performance. The FFN is discarded after updating the soft prompts.
+
+As a result, the authors found that prefix tuning demonstrates comparable performance to fully finetuning a model, despite having 1000x fewer parameters, and it performs even better in low-data settings.
+
+Take a look at [Prefix tuning for conditional generation](../task_guides/seq2seq-prefix-tuning) for a step-by-step guide on how to train a model with prefix tuning.
+
+## P-tuning
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/p-tuning.png"/>
+</div>
+<small>Prompt tokens can be inserted anywhere in the input sequence, and they are optimized by a prompt encoder <a href="https://hf.co/papers/2103.10385">(image source)</a>.</small>
+
+[P-tuning](https://hf.co/papers/2103.10385) is designed for natural language understanding (NLU) tasks and all language models. 
+It is another variation of a soft prompt method; P-tuning also adds a trainable embedding tensor that can be optimized to find better prompts, and it uses a prompt encoder (a bidirectional long-short term memory network or LSTM) to optimize the prompt parameters. Unlike prefix tuning though:
+
+- the prompt tokens can be inserted anywhere in the input sequence, and it isn't restricted to only the beginning
+- the prompt tokens are only added to the input instead of adding them to every layer of the model
+- introducing *anchor* tokens can improve performance because they indicate characteristics of a component in the input sequence
+
+The results suggest that P-tuning is more efficient than manually crafting prompts, and it enables GPT-like models to compete with BERT-like models on NLU tasks.
+
+Take a look at [P-tuning for sequence classification](../task_guides/ptuning-seq-classification) for a step-by-step guide on how to train a model with P-tuning.
+
+## Multitask prompt tuning
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/mpt.png"/>
+</div>
+<small><a href="https://hf.co/papers/2303.02861">Multitask prompt tuning enables parameter-efficient transfer learning</a>.</small>
+
+[Multitask prompt tuning (MPT)](https://hf.co/papers/2303.02861) learns a single prompt from data for multiple task types that can be shared for different target tasks. Other existing approaches learn a separate soft prompt for each task that need to be retrieved or aggregated for adaptation to target tasks. MPT consists of two stages:
+
+1. source training - for each task, its soft prompt is decomposed into task-specific vectors. The task-specific vectors are multiplied together to form another matrix W, and the Hadamard product is used between W and a shared prompt matrix P to generate a task-specific prompt matrix. The task-specific prompts are distilled into a single prompt matrix that is shared across all tasks. This prompt is trained with multitask training.
+2. target adaptation - to adapt the single prompt for a target task, a target prompt is initialized and expressed as the Hadamard product of the shared prompt matrix and the task-specific low-rank prompt matrix.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/mpt-decomposition.png"/>
+</div>
+<small><a href="https://hf.co/papers/2103.10385">Prompt decomposition</a>.</small>

docs/source/developer_guides/checkpoint.md

@@ -0,0 +1,250 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# PEFT checkpoint format
+
+This document describes how PEFT's checkpoint files are structured and how to convert between the PEFT format and other formats.
+
+## PEFT files
+
+PEFT (parameter-efficient fine-tuning) methods only update a small subset of a model's parameters rather than all of them. This is nice because checkpoint files can generally be much smaller than the original model files and are easier to store and share. However, this also means that to load a PEFT model, you need to have the original model available as well.
+
+When you call [`~PeftModel.save_pretrained`] on a PEFT model, the PEFT model saves three files, described below:
+
+1. `adapter_model.safetensors` or `adapter_model.bin`
+
+By default, the model is saved in the `safetensors` format, a secure alternative to the `bin` format, which is known to be susceptible to [security vulnerabilities](https://huggingface.co/docs/hub/security-pickle) because it uses the pickle utility under the hood. Both formats store the same `state_dict` though, and are interchangeable.
+
+The `state_dict` only contains the parameters of the adapter module, not the base model. To illustrate the difference in size, a normal BERT model requires ~420MB of disk space, whereas an IA³ adapter on top of this BERT model only requires ~260KB.
+
+2. `adapter_config.json`
+
+The `adapter_config.json` file contains the configuration of the adapter module, which is necessary to load the model. Below is an example of an `adapter_config.json` for an IA³ adapter with standard settings applied to a BERT model:
+
+```json
+{
+  "auto_mapping": {
+    "base_model_class": "BertModel",
+    "parent_library": "transformers.models.bert.modeling_bert"
+  },
+  "base_model_name_or_path": "bert-base-uncased",
+  "fan_in_fan_out": false,
+  "feedforward_modules": [
+    "output.dense"
+  ],
+  "inference_mode": true,
+  "init_ia3_weights": true,
+  "modules_to_save": null,
+  "peft_type": "IA3",
+  "revision": null,
+  "target_modules": [
+    "key",
+    "value",
+    "output.dense"
+  ],
+  "task_type": null
+}
+```
+
+The configuration file contains:
+
+- the adapter module type stored, `"peft_type": "IA3"`
+- information about the base model like `"base_model_name_or_path": "bert-base-uncased"`
+- the revision of the model (if any), `"revision": null`
+
+If the base model is not a pretrained Transformers model, the latter two entries will be `null`. Other than that, the settings are all related to the specific IA³ adapter that was used to fine-tune the model.
+
+3. `README.md`
+
+The generated `README.md` is the model card of a PEFT model and contains a few pre-filled entries. The intent of this is to make it easier to share the model with others and to provide some basic information about the model. This file is not needed to load the model.
+
+## Convert to PEFT format
+
+When converting from another format to the PEFT format, we require both the `adapter_model.safetensors` (or `adapter_model.bin`) file and the `adapter_config.json` file.
+
+### adapter_model
+
+For the model weights, it is important to use the correct mapping from parameter name to value for PEFT to load the file. Getting this mapping right is an exercise in checking the implementation details, as there is no generally agreed upon format for PEFT adapters.
+
+Fortunately, figuring out this mapping is not overly complicated for common base cases. Let's look at a concrete example, the [`LoraLayer`](https://github.com/huggingface/peft/blob/main/src/peft/tuners/lora/layer.py):
+
+```python
+# showing only part of the code
+
+class LoraLayer(BaseTunerLayer):
+    # All names of layers that may contain (trainable) adapter weights
+    adapter_layer_names = ("lora_A", "lora_B", "lora_embedding_A", "lora_embedding_B")
+    # All names of other parameters that may contain adapter-related parameters
+    other_param_names = ("r", "lora_alpha", "scaling", "lora_dropout")
+
+    def __init__(self, base_layer: nn.Module, **kwargs) -> None:
+        self.base_layer = base_layer
+        self.r = {}
+        self.lora_alpha = {}
+        self.scaling = {}
+        self.lora_dropout = nn.ModuleDict({})
+        self.lora_A = nn.ModuleDict({})
+        self.lora_B = nn.ModuleDict({})
+        # For Embedding layer
+        self.lora_embedding_A = nn.ParameterDict({})
+        self.lora_embedding_B = nn.ParameterDict({})
+        # Mark the weight as unmerged
+        self._disable_adapters = False
+        self.merged_adapters = []
+        self.use_dora: dict[str, bool] = {}
+        self.lora_magnitude_vector: Optional[torch.nn.ParameterDict] = None  # for DoRA
+        self._caches: dict[str, Any] = {}
+        self.kwargs = kwargs
+```
+
+In the `__init__` code used by all `LoraLayer` classes in PEFT, there are a bunch of parameters used to initialize the model, but only a few are relevant for the checkpoint file: `lora_A`, `lora_B`, `lora_embedding_A`, and `lora_embedding_B`. These parameters are listed in the class attribute `adapter_layer_names` and contain the learnable parameters, so they must be included in the checkpoint file. All the other parameters, like the rank `r`, are derived from the `adapter_config.json` and must be included there (unless the default value is used).
+
+Let's check the `state_dict` of a PEFT LoRA model applied to BERT. When printing the first five keys using the default LoRA settings (the remaining keys are the same, just with different layer numbers), we get:
+
+- `base_model.model.encoder.layer.0.attention.self.query.lora_A.weight` 
+- `base_model.model.encoder.layer.0.attention.self.query.lora_B.weight` 
+- `base_model.model.encoder.layer.0.attention.self.value.lora_A.weight` 
+- `base_model.model.encoder.layer.0.attention.self.value.lora_B.weight` 
+- `base_model.model.encoder.layer.1.attention.self.query.lora_A.weight`
+- etc.
+
+Let's break this down:
+
+- By default, for BERT models, LoRA is applied to the `query` and `value` layers of the attention module. This is why you see `attention.self.query` and `attention.self.value` in the key names for each layer.
+- LoRA decomposes the weights into two low-rank matrices, `lora_A` and `lora_B`. This is where `lora_A` and `lora_B` come from in the key names.
+- These LoRA matrices are implemented as `nn.Linear` layers, so the parameters are stored in the `.weight` attribute (`lora_A.weight`, `lora_B.weight`).
+- By default, LoRA isn't applied to BERT's embedding layer, so there are _no entries_ for `lora_A_embedding` and `lora_B_embedding`.
+- The keys of the `state_dict` always start with `"base_model.model."`. The reason is that, in PEFT, we wrap the base model inside a tuner-specific model (`LoraModel` in this case), which itself is wrapped in a general PEFT model (`PeftModel`). For this reason, these two prefixes are added to the keys. When converting to the PEFT format, it is required to add these prefixes.
+
+<Tip>
+
+This last point is not true for prefix tuning techniques like prompt tuning. There, the extra embeddings are directly stored in the `state_dict` without any prefixes added to the keys.
+
+</Tip>
+
+When inspecting the parameter names in the loaded model, you might be surprised to find that they look a bit different, e.g. `base_model.model.encoder.layer.0.attention.self.query.lora_A.default.weight`. The difference is the *`.default`* part in the second to last segment. This part exists because PEFT generally allows the addition of multiple adapters at once (using an `nn.ModuleDict` or `nn.ParameterDict` to store them). For example, if you add another adapter called "other", the key for that adapter would be `base_model.model.encoder.layer.0.attention.self.query.lora_A.other.weight`.
+
+When you call [`~PeftModel.save_pretrained`], the adapter name is stripped from the keys. The reason is that the adapter name is not an important part of the model architecture; it is just an arbitrary name. When loading the adapter, you could choose a totally different name, and the model would still work the same way. This is why the adapter name is not stored in the checkpoint file.
+
+<Tip>
+
+If you call `save_pretrained("some/path")` and the adapter name is not `"default"`, the adapter is stored in a sub-directory with the same name as the adapter. So if the name is "other", it would be stored inside of `some/path/other`.
+
+</Tip>
+
+In some circumstances, deciding which values to add to the checkpoint file can become a bit more complicated. For example, in PEFT, DoRA is implemented as a special case of LoRA. If you want to convert a DoRA model to PEFT, you should create a LoRA checkpoint with extra entries for DoRA. You can see this in the `__init__` of the previous `LoraLayer` code:
+
+```python
+self.lora_magnitude_vector: Optional[torch.nn.ParameterDict] = None  # for DoRA
+```
+
+This indicates that there is an optional extra parameter per layer for DoRA.
+
+### adapter_config
+
+All the other information needed to load a PEFT model is contained in the `adapter_config.json` file. Let's check this file for a LoRA model applied to BERT:
+
+```json
+{
+  "alpha_pattern": {},
+  "auto_mapping": {
+    "base_model_class": "BertModel",
+    "parent_library": "transformers.models.bert.modeling_bert"
+  },
+  "base_model_name_or_path": "bert-base-uncased",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 8,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "query",
+    "value"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}
+```
+
+This contains a lot of entries, and at first glance, it could feel overwhelming to figure out all the right values to put in there. However, most of the entries are not necessary to load the model. This is either because they use the default values and don't need to be added or because they only affect the initialization of the LoRA weights, which is irrelevant when it comes to loading the model. If you find that you don't know what a specific parameter does, e.g., `"use_rslora",` don't add it, and you should be fine. Also note that as more options are added, this file will get more entries in the future, but it should be backward compatible.
+
+At the minimum, you should include the following entries:
+
+```json
+{
+  "target_modules": ["query", "value"],
+  "peft_type": "LORA"
+}
+```
+
+However, adding as many entries as possible, like the rank `r` or the `base_model_name_or_path` (if it's a Transformers model) is recommended. This information can help others understand the model better and share it more easily. To check which keys and values are expected, check out the [config.py](https://github.com/huggingface/peft/blob/main/src/peft/tuners/lora/config.py) file (as an example, this is the config file for LoRA) in the PEFT source code.
+
+## Model storage
+
+In some circumstances, you might want to store the whole PEFT model, including the base weights. This can be necessary if, for instance, the base model is not available to the users trying to load the PEFT model. You can merge the weights first or convert it into a Transformer model.
+
+### Merge the weights
+
+The most straightforward way to store the whole PEFT model is to merge the adapter weights into the base weights:
+
+```python
+merged_model = model.merge_and_unload()
+merged_model.save_pretrained(...)
+```
+
+There are some disadvantages to this approach, though:
+
+- Once [`~LoraModel.merge_and_unload`] is called, you get a basic model without any PEFT-specific functionality. This means you can't use any of the PEFT-specific methods anymore.
+- You cannot unmerge the weights, load multiple adapters at once, disable the adapter, etc.
+- Not all PEFT methods support merging weights.
+- Some PEFT methods may generally allow merging, but not with specific settings (e.g. when using certain quantization techniques).
+- The whole model will be much larger than the PEFT model, as it will contain all the base weights as well.
+
+But inference with a merged model should be a bit faster.
+
+### Convert to a Transformers model
+
+Another way to save the whole model, assuming the base model is a Transformers model, is to use this hacky approach to directly insert the PEFT weights into the base model and save it, which only works if you "trick" Transformers into believing the PEFT model is not a PEFT model. This only works with LoRA because other adapters are not implemented in Transformers.
+
+```python
+model = ...  # the PEFT model
+...
+# after you finish training the model, save it in a temporary location
+model.save_pretrained(<temp_location>)
+# now load this model directly into a transformers model, without the PEFT wrapper
+# the PEFT weights are directly injected into the base model
+model_loaded = AutoModel.from_pretrained(<temp_location>)
+# now make the loaded model believe that it is _not_ a PEFT model
+model_loaded._hf_peft_config_loaded = False
+# now when we save it, it will save the whole model
+model_loaded.save_pretrained(<final_location>)
+# or upload to Hugging Face Hub
+model_loaded.push_to_hub(<final_location>)
+```
+

docs/source/developer_guides/contributing.md

@@ -0,0 +1,92 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Contribute to PEFT
+
+We are happy to accept contributions to PEFT. If you plan to contribute, please read this to make the process as smooth as possible.
+
+## Installation
+
+For code contributions to PEFT, you should choose the ["source"](../install#source) installation method.
+
+If you are new to creating a pull request, follow the [Creating a pull request](https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/proposing-changes-to-your-work-with-pull-requests/creating-a-pull-request) guide by GitHub.
+
+## Tests and code quality checks
+
+Regardless of the contribution type (unless it’s only about the docs), you should run tests and code quality checks before creating a PR to ensure your contribution doesn’t break anything and follows the project standards.
+
+We provide a Makefile to execute the necessary tests. Run the code below for the unit test:
+
+```sh
+make test
+```
+
+Run one of the following to either only check or check and fix code quality and style:
+
+```sh
+make quality  # just check
+make style  # check and fix
+```
+
+You can also set up [`pre-commit`](https://pre-commit.com/) to run these fixes
+automatically as Git commit hooks.
+
+```bash
+$ pip install pre-commit
+$ pre-commit install
+```
+
+Running all the tests can take a couple of minutes, so during development it can be more efficient to only run tests specific to your change:
+
+```sh
+pytest tests/ -k <name-of-test>
+```
+
+This should finish much quicker and allow for faster iteration. However, you should still run the whole test suite before creating a PR because your change can inadvertently break tests that at first glance are unrelated.
+
+If your change is specific to a hardware setting (e.g., it requires CUDA), take a look at [tests/test_gpu_examples.py](https://github.com/huggingface/peft/blob/1c1c7fdaa6e6abaa53939b865dee1eded82ad032/tests/test_gpu_examples.py) and [tests/test_common_gpu.py](https://github.com/huggingface/peft/blob/1c1c7fdaa6e6abaa53939b865dee1eded82ad032/tests/test_common_gpu.py) to see if it makes sense to add tests there. If your change could have an effect on saving and loading models, please run the tests with the `--regression` flag to trigger regression tests.
+
+It can happen that while you’re working on your PR, the underlying code base changes due to other changes being merged. If that happens – especially when there is a merge conflict – please update your branch with the latest changes. This can be a merge or a rebase, and we'll squash and merge the PR once it’s ready.
+
+## PR description
+
+When opening a PR, please provide a nice description of the change you're proposing. If it relates to other issues or PRs, please reference them. Providing a good description not only helps the reviewers review your code better and faster, it can also be used later (as a basis) for the commit message which helps with long term maintenance of the project.
+
+If your code makes some non-trivial changes, it may also be a good idea to add comments to the code to explain those changes. For example, if you had to iterate on your implementation multiple times because the most obvious way didn’t work, it’s a good indication that a code comment is needed.
+
+## Bugfixes
+
+Please give a description of the circumstances that led to the bug. If there is an existing issue, please link to it (e.g., “Resolves #12345”).
+
+Ideally when a bugfix is provided, it should be accompanied by a test for the bug. The test should fail with the current code and pass with the bugfix. Add a comment to the test that references the issue or PR. Without a test, it is more difficult to prevent regressions in the future.
+
+## Add a new fine-tuning method
+
+New parameter-efficient fine-tuning methods are developed all the time. If you would like to add a new and promising method to PEFT, please follow these steps.
+
+1. Before you start to implement the new method, please open a GitHub issue with your proposal. This way, the maintainers can give you some early feedback.
+2. Please add a link to the source (usually a paper) of the method. Some evidence should be provided there is general interest in using the method. We will not add new methods that are freshly published, but there is no evidence of demand for it.
+3. When implementing the method, it makes sense to look for existing implementations that already exist as a guide. Moreover, when you structure your code, please take inspiration from the other PEFT methods. For example, if your method is similar to LoRA, it makes sense to structure your code similarly or even reuse some functions or classes where it makes sense (some code duplication is okay, but don’t overdo it).
+4. Ideally, in addition to the implementation of the new method, there should also be examples (notebooks, scripts), documentation, and an extensive test suite that proves the method works with a variety of tasks. However, this can be more challenging so it is acceptable to only provide the implementation and at least one working example. Documentation and tests can be added in follow up PRs.
+5. Once you have something that seems to be working, don’t hesitate to create a draft PR even if it’s not in a mergeable state yet. The maintainers are happy to give you feedback and guidance along the way.
+
+## Add other features
+
+It is best if you first open an issue on GitHub with a proposal to add the new feature. This way, you can discuss with the maintainers if it makes sense to add the feature before spending too much time on implementing it.
+
+New features should generally be accompanied by tests and documentation or examples. Without the latter, users will have a hard time discovering your cool new feature.
+
+Changes to the code should be implemented in a backward-compatible way. For example, existing code should continue to work the same way after the feature is merged.

docs/source/developer_guides/custom_models.md

@@ -0,0 +1,310 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Custom models
+
+Some fine-tuning techniques, such as prompt tuning, are specific to language models. That means in 🤗 PEFT, it is
+assumed a 🤗 Transformers model is being used. However, other fine-tuning techniques - like
+[LoRA](../conceptual_guides/lora) - are not restricted to specific model types.
+
+In this guide, we will see how LoRA can be applied to a multilayer perceptron, a computer vision model from the [timm](https://huggingface.co/docs/timm/index) library, or a new 🤗 Transformers architecture.
+
+## Multilayer perceptron
+
+Let's assume that we want to fine-tune a multilayer perceptron with LoRA. Here is the definition:
+
+```python
+from torch import nn
+
+
+class MLP(nn.Module):
+    def __init__(self, num_units_hidden=2000):
+        super().__init__()
+        self.seq = nn.Sequential(
+            nn.Linear(20, num_units_hidden),
+            nn.ReLU(),
+            nn.Linear(num_units_hidden, num_units_hidden),
+            nn.ReLU(),
+            nn.Linear(num_units_hidden, 2),
+            nn.LogSoftmax(dim=-1),
+        )
+
+    def forward(self, X):
+        return self.seq(X)
+```
+
+This is a straightforward multilayer perceptron with an input layer, a hidden layer, and an output layer.
+
+<Tip>
+
+For this toy example, we choose an exceedingly large number of hidden units to highlight the efficiency gains
+from PEFT, but those gains are in line with more realistic examples.
+
+</Tip>
+
+There are a few linear layers in this model that could be tuned with LoRA. When working with common 🤗 Transformers
+models, PEFT will know which layers to apply LoRA to, but in this case, it is up to us as a user to choose the layers.
+To determine the names of the layers to tune:
+
+```python
+print([(n, type(m)) for n, m in MLP().named_modules()])
+```
+
+This should print:
+
+```
+[('', __main__.MLP),
+ ('seq', torch.nn.modules.container.Sequential),
+ ('seq.0', torch.nn.modules.linear.Linear),
+ ('seq.1', torch.nn.modules.activation.ReLU),
+ ('seq.2', torch.nn.modules.linear.Linear),
+ ('seq.3', torch.nn.modules.activation.ReLU),
+ ('seq.4', torch.nn.modules.linear.Linear),
+ ('seq.5', torch.nn.modules.activation.LogSoftmax)]
+```
+
+Let's say we want to apply LoRA to the input layer and to the hidden layer, those are `'seq.0'` and `'seq.2'`. Moreover,
+let's assume we want to update the output layer without LoRA, that would be `'seq.4'`. The corresponding config would
+be:
+
+```python
+from peft import LoraConfig
+
+config = LoraConfig(
+    target_modules=["seq.0", "seq.2"],
+    modules_to_save=["seq.4"],
+)
+```
+
+With that, we can create our PEFT model and check the fraction of parameters trained:
+
+```python
+from peft import get_peft_model
+
+model = MLP()
+peft_model = get_peft_model(model, config)
+peft_model.print_trainable_parameters()
+# prints trainable params: 56,164 || all params: 4,100,164 || trainable%: 1.369798866581922
+```
+
+Finally, we can use any training framework we like, or write our own fit loop, to train the `peft_model`.
+
+For a complete example, check out [this notebook](https://github.com/huggingface/peft/blob/main/examples/multilayer_perceptron/multilayer_perceptron_lora.ipynb).
+
+## timm models
+
+The [timm](https://huggingface.co/docs/timm/index) library contains a large number of pretrained computer vision models.
+Those can also be fine-tuned with PEFT. Let's check out how this works in practice.
+
+To start, ensure that timm is installed in the Python environment:
+
+```bash
+python -m pip install -U timm
+```
+
+Next we load a timm model for an image classification task:
+
+```python
+import timm
+
+num_classes = ...
+model_id = "timm/poolformer_m36.sail_in1k"
+model = timm.create_model(model_id, pretrained=True, num_classes=num_classes)
+```
+
+Again, we need to make a decision about what layers to apply LoRA to. Since LoRA supports 2D conv layers, and since
+those are a major building block of this model, we should apply LoRA to the 2D conv layers. To identify the names of
+those layers, let's look at all the layer names:
+
+```python
+print([(n, type(m)) for n, m in model.named_modules()])
+```
+
+This will print a very long list, we'll only show the first few:
+
+```
+[('', timm.models.metaformer.MetaFormer),
+ ('stem', timm.models.metaformer.Stem),
+ ('stem.conv', torch.nn.modules.conv.Conv2d),
+ ('stem.norm', torch.nn.modules.linear.Identity),
+ ('stages', torch.nn.modules.container.Sequential),
+ ('stages.0', timm.models.metaformer.MetaFormerStage),
+ ('stages.0.downsample', torch.nn.modules.linear.Identity),
+ ('stages.0.blocks', torch.nn.modules.container.Sequential),
+ ('stages.0.blocks.0', timm.models.metaformer.MetaFormerBlock),
+ ('stages.0.blocks.0.norm1', timm.layers.norm.GroupNorm1),
+ ('stages.0.blocks.0.token_mixer', timm.models.metaformer.Pooling),
+ ('stages.0.blocks.0.token_mixer.pool', torch.nn.modules.pooling.AvgPool2d),
+ ('stages.0.blocks.0.drop_path1', torch.nn.modules.linear.Identity),
+ ('stages.0.blocks.0.layer_scale1', timm.models.metaformer.Scale),
+ ('stages.0.blocks.0.res_scale1', torch.nn.modules.linear.Identity),
+ ('stages.0.blocks.0.norm2', timm.layers.norm.GroupNorm1),
+ ('stages.0.blocks.0.mlp', timm.layers.mlp.Mlp),
+ ('stages.0.blocks.0.mlp.fc1', torch.nn.modules.conv.Conv2d),
+ ('stages.0.blocks.0.mlp.act', torch.nn.modules.activation.GELU),
+ ('stages.0.blocks.0.mlp.drop1', torch.nn.modules.dropout.Dropout),
+ ('stages.0.blocks.0.mlp.norm', torch.nn.modules.linear.Identity),
+ ('stages.0.blocks.0.mlp.fc2', torch.nn.modules.conv.Conv2d),
+ ('stages.0.blocks.0.mlp.drop2', torch.nn.modules.dropout.Dropout),
+ ('stages.0.blocks.0.drop_path2', torch.nn.modules.linear.Identity),
+ ('stages.0.blocks.0.layer_scale2', timm.models.metaformer.Scale),
+ ('stages.0.blocks.0.res_scale2', torch.nn.modules.linear.Identity),
+ ('stages.0.blocks.1', timm.models.metaformer.MetaFormerBlock),
+ ('stages.0.blocks.1.norm1', timm.layers.norm.GroupNorm1),
+ ('stages.0.blocks.1.token_mixer', timm.models.metaformer.Pooling),
+ ('stages.0.blocks.1.token_mixer.pool', torch.nn.modules.pooling.AvgPool2d),
+ ...
+ ('head.global_pool.flatten', torch.nn.modules.linear.Identity),
+ ('head.norm', timm.layers.norm.LayerNorm2d),
+ ('head.flatten', torch.nn.modules.flatten.Flatten),
+ ('head.drop', torch.nn.modules.linear.Identity),
+ ('head.fc', torch.nn.modules.linear.Linear)]
+ ]
+```
+
+Upon closer inspection, we see that the 2D conv layers have names such as `"stages.0.blocks.0.mlp.fc1"` and
+`"stages.0.blocks.0.mlp.fc2"`. How can we match those layer names specifically? You can write a [regular
+expressions](https://docs.python.org/3/library/re.html) to match the layer names. For our case, the regex
+`r".*\.mlp\.fc\d"` should do the job.
+
+Furthermore, as in the first example, we should ensure that the output layer, in this case the classification head, is
+also updated. Looking at the end of the list printed above, we can see that it's named `'head.fc'`. With that in mind,
+here is our LoRA config:
+
+```python
+config = LoraConfig(target_modules=r".*\.mlp\.fc\d", modules_to_save=["head.fc"])
+```
+
+Then we only need to create the PEFT model by passing our base model and the config to `get_peft_model`:
+
+```python
+peft_model = get_peft_model(model, config)
+peft_model.print_trainable_parameters()
+# prints trainable params: 1,064,454 || all params: 56,467,974 || trainable%: 1.88505789139876
+```
+
+This shows us that we only need to train less than 2% of all parameters, which is a huge efficiency gain.
+
+For a complete example, check out [this notebook](https://github.com/huggingface/peft/blob/main/examples/image_classification/image_classification_timm_peft_lora.ipynb).
+
+## New transformers architectures
+
+When new popular transformers architectures are released, we do our best to quickly add them to PEFT. If you come across a transformers model that is not supported out of the box, don't worry, it will most likely still work if the config is set correctly. Specifically, you have to identify the layers that should be adapted and set them correctly when initializing the corresponding config class, e.g. `LoraConfig`. Here are some tips to help with this.
+
+As a first step, it is a good idea is to check the existing models for inspiration. You can find them inside of [constants.py](https://github.com/huggingface/peft/blob/main/src/peft/utils/constants.py) in the PEFT repository. Often, you'll find a similar architecture that uses the same names. For example, if the new model architecture is a variation of the "mistral" model and you want to apply LoRA, you can see that the entry for "mistral" in `TRANSFORMERS_MODELS_TO_LORA_TARGET_MODULES_MAPPING` contains `["q_proj", "v_proj"]`. This tells you that for "mistral" models, the `target_modules` for LoRA should be `["q_proj", "v_proj"]`:
+
+```python
+from peft import LoraConfig, get_peft_model
+
+my_mistral_model = ...
+config = LoraConfig(
+    target_modules=["q_proj", "v_proj"],
+    ...,  # other LoRA arguments
+)
+peft_model = get_peft_model(my_mistral_model, config)
+```
+
+If that doesn't help, check the existing modules in your model architecture with the `named_modules` method and try to identify the attention layers, especially the key, query, and value layers. Those will often have names such as `c_attn`, `query`, `q_proj`, etc. The key layer is not always adapted, and ideally, you should check whether including it results in better performance.
+
+Additionally, linear layers are common targets to be adapted (e.g. in [QLoRA paper](https://arxiv.org/abs/2305.14314), authors suggest to adapt them as well). Their names will often contain the strings `fc` or `dense`.
+
+If you want to add a new model to PEFT, please create an entry in [constants.py](https://github.com/huggingface/peft/blob/main/src/peft/utils/constants.py) and open a pull request on the [repository](https://github.com/huggingface/peft/pulls). Don't forget to update the [README](https://github.com/huggingface/peft#models-support-matrix) as well.
+
+## Verify parameters and layers
+
+You can verify whether you've correctly applied a PEFT method to your model in a few ways.
+
+* Check the fraction of parameters that are trainable with the [`~PeftModel.print_trainable_parameters`] method. If this number is lower or higher than expected, check the model `repr` by printing the model. This shows the names of all the layer types in the model. Ensure that only the intended target layers are replaced by the adapter layers. For example, if LoRA is applied to `nn.Linear` layers, then you should only see `lora.Linear` layers being used.
+
+```py
+peft_model.print_trainable_parameters()
+```
+
+* Another way you can view the adapted layers is to use the `targeted_module_names` attribute to list the name of each module that was adapted.
+
+```python
+print(peft_model.targeted_module_names)
+```
+
+## Unsupported module types
+
+Methods like LoRA only work if the target modules are supported by PEFT. For example, it's possible to apply LoRA to `nn.Linear` and `nn.Conv2d` layers, but not, for instance, to `nn.LSTM`. If you find a layer class you want to apply PEFT to is not supported, you can:
+
+ - define a custom mapping to dynamically dispatch custom modules in LoRA
+ -  open an [issue](https://github.com/huggingface/peft/issues) and request the feature where maintainers will implement it or guide you on how to implement it yourself if demand for this module type is sufficiently high
+
+### Experimental support for dynamic dispatch of custom modules in LoRA
+
+> [!WARNING]
+> This feature is experimental and subject to change, depending on its reception by the community. We will introduce a public and stable API if there is significant demand for it.
+
+PEFT supports an experimental API for custom module types for LoRA. Let's assume you have a LoRA implementation for LSTMs. Normally, you would not be able to tell PEFT to use it, even if it would theoretically work with PEFT. However, this is possible with dynamic dispatch of custom layers.
+
+The experimental API currently looks like this:
+
+```python
+class MyLoraLSTMLayer:
+    ...
+
+base_model = ...  # load the base model that uses LSTMs
+
+# add the LSTM layer names to target_modules
+config = LoraConfig(..., target_modules=["lstm"])
+# define a mapping from base layer type to LoRA layer type
+custom_module_mapping = {nn.LSTM: MyLoraLSTMLayer}
+# register the new mapping
+config._register_custom_module(custom_module_mapping)
+# after registration, create the PEFT model
+peft_model = get_peft_model(base_model, config)
+# do training
+```
+
+<Tip>
+
+When you call [`get_peft_model`], you will see a warning because PEFT does not recognize the targeted module type. In this case, you can ignore this warning.
+
+</Tip>
+
+By supplying a custom mapping, PEFT first checks the base model's layers against the custom mapping and dispatches to the custom LoRA layer type if there is a match. If there is no match, PEFT checks the built-in LoRA layer types for a match.
+
+Therefore, this feature can also be used to override existing dispatch logic, e.g. if you want to use your own LoRA layer for `nn.Linear` instead of using the one provided by PEFT.
+
+When creating your custom LoRA module, please follow the same rules as the [existing LoRA modules](https://github.com/huggingface/peft/blob/main/src/peft/tuners/lora/layer.py). Some important constraints to consider:
+
+- The custom module should inherit from `nn.Module` and `peft.tuners.lora.layer.LoraLayer`.
+- The `__init__` method of the custom module should have the positional arguments `base_layer` and `adapter_name`. After this, there are additional `**kwargs` that you are free to use or ignore.
+- The learnable parameters should be stored in an `nn.ModuleDict` or `nn.ParameterDict`, where the key corresponds to the name of the specific adapter (remember that a model can have more than one adapter at a time).
+- The name of these learnable parameter attributes should start with `"lora_"`, e.g. `self.lora_new_param = ...`.
+- Some methods are optional, e.g. you only need to implement `merge` and `unmerge` if you want to support weight merging.
+
+Currently, the information about the custom module does not persist when you save the model. When loading the model, you have to register the custom modules again.
+
+```python
+# saving works as always and includes the parameters of the custom modules
+peft_model.save_pretrained(<model-path>)
+
+# loading the model later:
+base_model = ...
+# load the LoRA config that you saved earlier
+config = LoraConfig.from_pretrained(<model-path>)
+# register the custom module again, the same way as the first time
+custom_module_mapping = {nn.LSTM: MyLoraLSTMLayer}
+config._register_custom_module(custom_module_mapping)
+# pass the config instance to from_pretrained:
+peft_model = PeftModel.from_pretrained(model, tmp_path / "lora-custom-module", config=config)
+```
+
+If you use this feature and find it useful, or if you encounter problems, let us know by creating an issue or a discussion on GitHub. This allows us to estimate the demand for this feature and add a public API if it is sufficiently high.

docs/source/developer_guides/lora.md

@@ -0,0 +1,387 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contains specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# LoRA
+
+LoRA is low-rank decomposition method to reduce the number of trainable parameters which speeds up finetuning large models and uses less memory. In PEFT, using LoRA is as easy as setting up a [`LoraConfig`] and wrapping it with [`get_peft_model`] to create a trainable [`PeftModel`].
+
+This guide explores in more detail other options and features for using LoRA.
+
+## Initialization
+
+The initialization of LoRA weights is controlled by the parameter `init_lora_weights` in [`LoraConfig`]. By default, PEFT initializes LoRA weights with Kaiming-uniform for weight A and zeros for weight B resulting in an identity transform (same as the reference [implementation](https://github.com/microsoft/LoRA)).
+
+It is also possible to pass `init_lora_weights="gaussian"`. As the name suggests, this initializes weight A with a Gaussian distribution and zeros for weight B (this is how [Diffusers](https://huggingface.co/docs/diffusers/index) initializes LoRA weights).
+
+```py
+from peft import LoraConfig
+
+config = LoraConfig(init_lora_weights="gaussian", ...)
+```
+
+There is also an option to set `init_lora_weights=False` which is useful for debugging and testing. This should be the only time you use this option. When choosing this option, the LoRA weights are initialized such that they do *not* result in an identity transform.
+
+```py
+from peft import LoraConfig
+
+config = LoraConfig(init_lora_weights=False, ...)
+```
+
+### PiSSA
+[PiSSA](https://arxiv.org/abs/2404.02948) initializes the LoRA adapter using the principal singular values and singular vectors. This straightforward modification allows PiSSA to converge more rapidly than LoRA and ultimately attain superior performance. Moreover, PiSSA reduces the quantization error compared to QLoRA, leading to further enhancements. 
+
+Configure the initialization method to "pissa", which may take several minutes to execute SVD on the pre-trained model:
+```python
+from peft import LoraConfig
+config = LoraConfig(init_lora_weights="pissa", ...)
+```
+Alternatively, execute fast SVD, which takes only a few seconds. The number of iterations determines the trade-off between the error and computation time:
+```python
+lora_config = LoraConfig(init_lora_weights="pissa_niter_[number of iters]", ...) 
+```
+For detailed instruction on using PiSSA, please follow [these instructions](https://github.com/fxmeng/peft/tree/main/examples/pissa_finetuning).
+
+### OLoRA
+[OLoRA](https://arxiv.org/abs/2406.01775) utilizes QR decomposition to initialize the LoRA adapters. OLoRA translates the base weights of the model by a factor of their QR decompositions, i.e., it mutates the weights before performing any training on them. This approach significantly improves stability, accelerates convergence speed, and ultimately achieves superior performance.
+
+You just need to pass a single additional option to use OLoRA:
+```python
+from peft import LoraConfig
+config = LoraConfig(init_lora_weights="olora", ...)
+```
+For more advanced usage, please refer to our [documentation](https://github.com/huggingface/peft/tree/main/examples/olora_finetuning).
+### LoftQ
+
+#### Standard approach
+
+When quantizing the base model for QLoRA training, consider using the [LoftQ initialization](https://arxiv.org/abs/2310.08659), which has been shown to improve performance when training quantized models. The idea is that the LoRA weights are initialized such that the quantization error is minimized. To use LoftQ, follow [these instructions](https://github.com/huggingface/peft/tree/main/examples/loftq_finetuning).
+
+In general, for LoftQ to work best, it is recommended to target as many layers with LoRA as possible, since those not targeted cannot have LoftQ applied. This means that passing `LoraConfig(..., target_modules="all-linear")` will most likely give the best results. Also, you should use `nf4` as quant type in your quantization config when using 4bit quantization, i.e. `BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4")`.
+
+#### A more convenient way
+
+An easier but more limited way to apply LoftQ initialization is to use the convenience function `replace_lora_weights_loftq`. This takes the quantized PEFT model as input and replaces the LoRA weights in-place with their LoftQ-initialized counterparts.
+
+```python
+from peft import replace_lora_weights_loftq
+from transformers import BitsAndBytesConfig
+
+bnb_config = BitsAndBytesConfig(load_in_4bit=True, ...)
+base_model = AutoModelForCausalLM.from_pretrained(..., quantization_config=bnb_config)
+# note: don't pass init_lora_weights="loftq" or loftq_config!
+lora_config = LoraConfig(task_type="CAUSAL_LM")
+peft_model = get_peft_model(base_model, lora_config)
+replace_lora_weights_loftq(peft_model)
+```
+
+`replace_lora_weights_loftq` also allows you to pass a `callback` argument to give you more control over which layers should be modified or not, which empirically can improve the results quite a lot. To see a more elaborate example of this, check out [this notebook](https://github.com/huggingface/peft/blob/main/examples/loftq_finetuning/LoftQ_weight_replacement.ipynb).
+
+`replace_lora_weights_loftq` implements only one iteration step of LoftQ. This means that only the LoRA weights are updated, instead of iteratevily updating LoRA weights and quantized base model weights. This may lead to lower performance but has the advantage that we can use the original quantized weights derived from the base model, instead of having to keep an extra copy of modified quantized weights. Whether this tradeoff is worthwhile depends on the use case.
+
+At the moment, `replace_lora_weights_loftq` has these additional limitations:
+
+- Model files must be stored as a `safetensors` file.
+- Only bitsandbytes 4bit quantization is supported.
+
+<Tip>
+
+Learn more about how PEFT works with quantization in the [Quantization](quantization) guide.
+
+</Tip>
+
+### Rank-stabilized LoRA
+
+Another way to initialize [`LoraConfig`] is with the [rank-stabilized LoRA (rsLoRA)](https://huggingface.co/papers/2312.03732) method. The LoRA architecture scales each adapter during every forward pass by a fixed scalar which is set at initialization and depends on the rank `r`. The scalar is given by `lora_alpha/r` in the original implementation, but rsLoRA uses `lora_alpha/math.sqrt(r)` which stabilizes the adapters and increases the performance potential from using a higher `r`.
+
+```py
+from peft import LoraConfig
+
+config = LoraConfig(use_rslora=True, ...)
+```
+
+### Weight-Decomposed Low-Rank Adaptation (DoRA)
+
+This technique decomposes the updates of the weights into two parts, magnitude and direction. Direction is handled by normal LoRA, whereas the magnitude is handled by a separate learnable parameter. This can improve the performance of LoRA, especially at low ranks. For more information on DoRA, see  https://arxiv.org/abs/2402.09353.
+
+```py
+from peft import LoraConfig
+
+config = LoraConfig(use_dora=True, ...)
+```
+
+If parts of the model or the DoRA adapter are offloaded to CPU you can get a significant speedup at the cost of some temporary (ephemeral) VRAM overhead by using `ephemeral_gpu_offload=True` in `config.runtime_config`.
+
+```py
+from peft import LoraConfig, LoraRuntimeConfig
+
+config = LoraConfig(use_dora=True, runtime_config=LoraRuntimeConfig(ephemeral_gpu_offload=True), ...)
+```
+
+A `PeftModel` with a DoRA adapter can also be loaded with `ephemeral_gpu_offload=True` flag using the `from_pretrained` method as well as the `load_adapter` method.
+
+```py
+from peft import PeftModel
+
+model = PeftModel.from_pretrained(base_model, peft_model_id, ephemeral_gpu_offload=True)
+```
+
+#### Caveats
+
+- DoRA only supports linear and Conv2d layers at the moment.
+- DoRA introduces a bigger overhead than pure LoRA, so it is recommended to merge weights for inference, see [`LoraModel.merge_and_unload`]. 
+- DoRA should work with weights quantized with bitsandbytes ("QDoRA"). However, issues have been reported when using QDoRA with DeepSpeed Zero2.
+
+### QLoRA-style training
+
+The default LoRA settings in PEFT add trainable weights to the query and value layers of each attention block. But [QLoRA](https://hf.co/papers/2305.14314), which adds trainable weights to all the linear layers of a transformer model, can provide performance equal to a fully finetuned model. To apply LoRA to all the linear layers, like in QLoRA, set `target_modules="all-linear"` (easier than specifying individual modules by name which can vary depending on the architecture).
+
+```py
+config = LoraConfig(target_modules="all-linear", ...)
+```
+
+### Memory efficient Layer Replication with LoRA
+
+An approach used to improve the performance of models is to expand a model by duplicating layers in the model to build a larger model from a pretrained model of a given size. For example increasing a 7B model to a 10B model as described in the [SOLAR](https://arxiv.org/abs/2312.15166) paper. PEFT LoRA supports this kind of expansion in a memory efficient manner that supports further fine-tuning using LoRA adapters attached to the layers post replication of the layers. The replicated layers do not take additional memory as they share the underlying weights so the only additional memory required is the memory for the adapter weights. To use this feature you would create a config with the `layer_replication` argument.
+
+```py
+config = LoraConfig(layer_replication=[[0,4], [2,5]], ...)
+```
+
+Assuming the original model had 5 layers `[0, 1, 2 ,3, 4]`, this would create a model with 7 layers arranged as `[0, 1, 2, 3, 2, 3, 4]`. This follows the [mergekit](https://github.com/arcee-ai/mergekit) pass through merge convention where sequences of layers specified as start inclusive and end exclusive tuples are stacked to build the final model. Each layer in the final model gets its own distinct set of LoRA adapters.
+
+[Fewshot-Metamath-OrcaVicuna-Mistral-10B](https://huggingface.co/abacusai/Fewshot-Metamath-OrcaVicuna-Mistral-10B) is an example of a model trained using this method on Mistral-7B expanded to 10B. The
+[adapter_config.json](https://huggingface.co/abacusai/Fewshot-Metamath-OrcaVicuna-Mistral-10B/blob/main/adapter_config.json) shows a sample LoRA adapter config applying this method for fine-tuning.
+
+## Optimizers
+
+LoRA training can optionally include special purpose optimizers. Currently the only such optimizer is LoRA+.
+
+### LoRA+ optimized LoRA
+
+LoRA training can be optimized using [LoRA+](https://arxiv.org/abs/2402.12354), which uses different learning rates for the adapter matrices A and B, shown to increase finetuning speed by up to 2x and performance by 1-2%.
+
+```py
+from peft import LoraConfig, get_peft_model
+from peft.optimizers import create_loraplus_optimizer
+from transformers import Trainer
+import bitsandbytes as bnb
+
+base_model = ...
+config = LoraConfig(...)
+model = get_peft_model(base_model, config)
+
+optimizer = create_loraplus_optimizer(
+    model=model,
+    optimizer_cls=bnb.optim.Adam8bit,
+    lr=5e-5,
+    loraplus_lr_ratio=16,
+)
+scheduler = None
+
+...
+trainer = Trainer(
+    ...,
+    optimizers=(optimizer, scheduler),
+)
+```
+
+## Merge LoRA weights into the base model
+
+While LoRA is significantly smaller and faster to train, you may encounter latency issues during inference due to separately loading the base model and the LoRA adapter. To eliminate latency, use the [`~LoraModel.merge_and_unload`] function to merge the adapter weights with the base model. This allows you to use the newly merged model as a standalone model. The [`~LoraModel.merge_and_unload`] function doesn't keep the adapter weights in memory.
+
+Below is a diagram that explains the intuition of LoRA adapter merging:
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/lora_diagram.png"/>
+</div>
+
+We show in the snippets below how to run that using PEFT.
+
+```py
+from transformers import AutoModelForCausalLM
+from peft import PeftModel
+
+base_model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
+peft_model_id = "alignment-handbook/zephyr-7b-sft-lora"
+model = PeftModel.from_pretrained(base_model, peft_model_id)
+model.merge_and_unload()
+```
+
+If you need to keep a copy of the weights so you can unmerge the adapter later or delete and load different ones, you should use the [`~LoraModel.merge_adapter`] function instead. Now you have the option to use [`~LoraModel.unmerge_adapter`] to return the base model.
+
+```py
+from transformers import AutoModelForCausalLM
+from peft import PeftModel
+
+base_model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
+peft_model_id = "alignment-handbook/zephyr-7b-sft-lora"
+model = PeftModel.from_pretrained(base_model, peft_model_id)
+model.merge_adapter()
+
+# unmerge the LoRA layers from the base model
+model.unmerge_adapter()
+```
+
+The [`~LoraModel.add_weighted_adapter`] function is useful for merging multiple LoRAs into a new adapter based on a user provided weighting scheme in the `weights` parameter. Below is an end-to-end example.
+
+First load the base model:
+
+```python
+from transformers import AutoModelForCausalLM
+from peft import PeftModel
+import torch
+
+base_model = AutoModelForCausalLM.from_pretrained(
+    "mistralai/Mistral-7B-v0.1", torch_dtype=torch.float16, device_map="auto"
+)
+```
+
+Then we load the first adapter: 
+
+```python
+peft_model_id = "alignment-handbook/zephyr-7b-sft-lora"
+model = PeftModel.from_pretrained(base_model, peft_model_id, adapter_name="sft")
+```
+
+Then load a different adapter and merge it with the first one:
+
+```python
+weighted_adapter_name = "sft-dpo"
+model.load_adapter("alignment-handbook/zephyr-7b-dpo-lora", adapter_name="dpo")
+model.add_weighted_adapter(
+    adapters=["sft", "dpo"],
+    weights=[0.7, 0.3],
+    adapter_name=weighted_adapter_name,
+    combination_type="linear"
+)
+model.set_adapter(weighted_adapter_name)
+```
+
+<Tip>
+
+There are several supported methods for `combination_type`. Refer to the [documentation](../package_reference/lora#peft.LoraModel.add_weighted_adapter) for more details. Note that "svd" as the `combination_type` is not supported when using `torch.float16` or `torch.bfloat16` as the datatype.
+
+</Tip>
+
+Now, perform inference:
+
+```python
+tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
+
+prompt = "Hey, are you conscious? Can you talk to me?"
+inputs = tokenizer(prompt, return_tensors="pt")
+inputs = {k: v.to("cuda") for k, v in inputs.items()}
+
+with torch.no_grad():
+    generate_ids = model.generate(**inputs, max_length=30)
+outputs = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+print(outputs)
+```
+
+## Load adapters
+
+Adapters can be loaded onto a pretrained model with [`~PeftModel.load_adapter`], which is useful for trying out different adapters whose weights aren't merged. Set the active adapter weights with the [`~LoraModel.set_adapter`] function.
+
+```py
+from transformers import AutoModelForCausalLM
+from peft import PeftModel
+
+base_model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
+peft_model_id = "alignment-handbook/zephyr-7b-sft-lora"
+model = PeftModel.from_pretrained(base_model, peft_model_id)
+
+# load different adapter
+model.load_adapter("alignment-handbook/zephyr-7b-dpo-lora", adapter_name="dpo")
+
+# set adapter as active
+model.set_adapter("dpo")
+```
+
+To return the base model, you could use [`~LoraModel.unload`] to unload all of the LoRA modules or [`~LoraModel.delete_adapter`] to delete the adapter entirely.
+
+```py
+# unload adapter
+model.unload()
+
+# delete adapter
+model.delete_adapter("dpo")
+```
+
+## Inference with different LoRA adapters in the same batch
+
+Normally, each inference batch has to use the same adapter(s) in PEFT. This can sometimes be annoying, because we may have batches that contain samples intended to be used with different LoRA adapters. For example, we could have a base model that works well in English and two more LoRA adapters, one for French and one for German. Usually, we would have to split our batches such that each batch only contains samples of one of the languages, we cannot combine different languages in the same batch.
+
+Thankfully, it is possible to mix different LoRA adapters in the same batch using the `adapter_name` argument. Below, we show an example of how this works in practice. First, let's load the base model, English, and the two adapters, French and German, like this:
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from peft import PeftModel
+
+model_id = ...
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+model = AutoModelForCausalLM.from_pretrained(model_id)
+# load the LoRA adapter for French
+peft_model = PeftModel.from_pretrained(model, <path>, adapter_name="adapter_fr")
+# next, load the LoRA adapter for German
+peft_model.load_adapter(<path>, adapter_name="adapter_de")
+```
+
+Now, we want to generate text on a sample that contains all three languages: The first three samples are in English, the next three are in French, and the last three are in German. We can use the `adapter_names` argument to specify which adapter to use for each sample. Since our base model is used for English, we use the special string `"__base__"` for these samples. For the next three samples, we indicate the adapter name of the French LoRA fine-tune, in this case `"adapter_fr"`. For the last three samples, we indicate the adapter name of the German LoRA fine-tune, in this case `"adapter_de"`. This way, we can use the base model and the two adapters in a single batch.
+
+```python
+inputs = tokenizer(
+    [
+        "Hello, my dog is cute",
+        "Hello, my cat is awesome",
+        "Hello, my fish is great",
+        "Salut, mon chien est mignon",
+        "Salut, mon chat est génial",
+        "Salut, mon poisson est super",
+        "Hallo, mein Hund ist süß",
+        "Hallo, meine Katze ist toll",
+        "Hallo, mein Fisch ist großartig",
+    ],
+    return_tensors="pt",
+    padding=True,
+)
+
+adapter_names = [
+    "__base__", "__base__", "__base__",
+    "adapter_fr", "adapter_fr", "adapter_fr",
+    "adapter_de", "adapter_de", "adapter_de",
+]
+output = peft_model.generate(**inputs, adapter_names=adapter_names, max_new_tokens=20)
+```
+
+Note that the order does not matter here, i.e. the samples in the batch don't need to be grouped by adapter as in the example above. We just need to ensure that the `adapter_names` argument is aligned correctly with the samples.
+
+Additionally, the same approach also works with the `modules_to_save` feature, which allows for saving and reusing specific neural network layers, such as custom heads for classification tasks, across different LoRA adapters. 
+
+### Caveats
+
+Using this features has some drawbacks, namely:
+
+- It only works for inference, not for training.
+- Disabling adapters using the `with model.disable_adapter()` context takes precedence over `adapter_names`.
+- You cannot pass `adapter_names` when some adapter weights where merged with base weight using the `merge_adapter` method. Please unmerge all adapters first by calling `model.unmerge_adapter()`.
+- For obvious reasons, this cannot be used after calling `merge_and_unload()`, since all the LoRA adapters will be merged into the base weights in this case.
+- This feature does not currently work with DoRA, so set `use_dora=False` in your `LoraConfig` if you want to use it.
+- The `modules_to_save` feature is currently only supported for the layers of types `Linear`, `Embedding`, `Conv2d` and `Conv1d`.
+- There is an expected overhead for inference with `adapter_names`, especially if the amount of different adapters in the batch is high. This is because the batch size is effectively reduced to the number of samples per adapter. If runtime performance is your top priority, try the following:
+  - Increase the batch size.
+  - Try to avoid having a large number of different adapters in the same batch, prefer homogeneous batches. This can be achieved by buffering samples with the same adapter and only perform inference with a small handfull of different adapters.
+  - Take a look at alternative implementations such as [LoRAX](https://github.com/predibase/lorax), [punica](https://github.com/punica-ai/punica), or [S-LoRA](https://github.com/S-LoRA/S-LoRA), which are specialized to work with a large number of different adapters.

docs/source/developer_guides/low_level_api.md

@@ -0,0 +1,126 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Adapter injection
+
+With PEFT, you can inject trainable adapters into any `torch` module which allows you to use adapter methods without relying on the modeling classes in PEFT. Currently, PEFT supports injecting [LoRA](../conceptual_guides/adapter#low-rank-adaptation-lora), [AdaLoRA](../conceptual_guides/adapter#adaptive-low-rank-adaptation-adalora), and [IA3](../conceptual_guides/ia3) into models because for these adapters, inplace modification of the model is sufficient for finetuning it.
+
+Check the table below to see when you should inject adapters.
+
+| Pros | Cons |
+|---|---|
+| the model is modified inplace, keeping all the original attributes and methods | manually write the `from_pretrained` and `save_pretrained` utility functions from Hugging Face to save and load adapters |
+| works for any `torch` module and modality | doesn't work with any of the utility methods provided by `PeftModel` such as disabling and merging adapters |
+
+## Creating a new PEFT model
+
+To perform the adapter injection, use the [`inject_adapter_in_model`] method. This method takes 3 arguments, the PEFT config, the model, and an optional adapter name. You can also attach multiple adapters to the model if you call [`inject_adapter_in_model`] multiple times with different adapter names.
+
+For example, to inject LoRA adapters into the `linear` submodule of the `DummyModel` module:
+
+```python
+import torch
+from peft import inject_adapter_in_model, LoraConfig
+
+class DummyModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.embedding = torch.nn.Embedding(10, 10)
+        self.linear = torch.nn.Linear(10, 10)
+        self.lm_head = torch.nn.Linear(10, 10)
+
+    def forward(self, input_ids):
+        x = self.embedding(input_ids)
+        x = self.linear(x)
+        x = self.lm_head(x)
+        return x
+
+
+lora_config = LoraConfig(
+    lora_alpha=16,
+    lora_dropout=0.1,
+    r=64,
+    bias="none",
+    target_modules=["linear"],
+)
+
+model = DummyModel()
+model = inject_adapter_in_model(lora_config, model)
+
+dummy_inputs = torch.LongTensor([[0, 1, 2, 3, 4, 5, 6, 7]])
+dummy_outputs = model(dummy_inputs)
+```
+
+Print the model to see that the adapters have been correctly injected.
+
+```bash
+DummyModel(
+  (embedding): Embedding(10, 10)
+  (linear): Linear(
+    in_features=10, out_features=10, bias=True
+    (lora_dropout): ModuleDict(
+      (default): Dropout(p=0.1, inplace=False)
+    )
+    (lora_A): ModuleDict(
+      (default): Linear(in_features=10, out_features=64, bias=False)
+    )
+    (lora_B): ModuleDict(
+      (default): Linear(in_features=64, out_features=10, bias=False)
+    )
+    (lora_embedding_A): ParameterDict()
+    (lora_embedding_B): ParameterDict()
+  )
+  (lm_head): Linear(in_features=10, out_features=10, bias=True)
+)
+```
+
+## Saving the model
+
+To only save the adapter, use the [`get_peft_model_state_dict`] function:
+
+```python
+from peft import get_peft_model_state_dict
+
+peft_state_dict = get_peft_model_state_dict(model)
+print(peft_state_dict)
+```
+
+Otherwise, `model.state_dict()` returns the full state dict of the model.
+
+## Loading the model
+
+After loading the saved `state_dict`, it can be applied using the [`set_peft_model_state_dict`] function:
+
+```python
+from peft import set_peft_model_state_dict
+
+model = DummyModel()
+model = inject_adapter_in_model(lora_config, model)
+outcome = set_peft_model_state_dict(model, peft_state_dict)
+# check that there were no wrong keys
+print(outcome.unexpected_keys)
+```
+
+If injecting the adapter is slow or you need to load a large number of adapters, you may use an optimization that allows to create an "empty" adapter on meta device and only fills the weights with real weights when the [`set_peft_model_state_dict`] is called. To do this, pass `low_cpu_mem_usage=True` to both [`inject_adapter_in_model`] and [`set_peft_model_state_dict`].
+
+```python
+model = DummyModel()
+model = inject_adapter_in_model(lora_config, model, low_cpu_mem_usage=True)
+
+print(model.linear.lora_A["default"].weight.device.type == "meta")  # should be True
+set_peft_model_state_dict(model, peft_state_dict, low_cpu_mem_usage=True)
+print(model.linear.lora_A["default"].weight.device.type == "cpu")  # should be True
+```

docs/source/developer_guides/mixed_models.md

@@ -0,0 +1,37 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Mixed adapter types
+
+Normally, it isn't possible to mix different adapter types in 🤗 PEFT. You can create a PEFT model with two different LoRA adapters (which can have different config options), but it is not possible to combine a LoRA and LoHa adapter. With [`PeftMixedModel`] however, this works as long as the adapter types are compatible. The main purpose of allowing mixed adapter types is to combine trained adapters for inference. While it is possible to train a mixed adapter model, this has not been tested and is not recommended.
+
+To load different adapter types into a PEFT model, use [`PeftMixedModel`] instead of [`PeftModel`]:
+
+```py
+from peft import PeftMixedModel
+
+base_model = ...  # load the base model, e.g. from transformers
+# load first adapter, which will be called "default"
+peft_model = PeftMixedModel.from_pretrained(base_model, <path_to_adapter1>)
+peft_model.load_adapter(<path_to_adapter2>, adapter_name="other")
+peft_model.set_adapter(["default", "other"])
+```
+
+The [`~PeftMixedModel.set_adapter`] method is necessary to activate both adapters, otherwise only the first adapter would be active. You can keep adding more adapters by calling [`~PeftModel.add_adapter`] repeatedly.
+
+[`PeftMixedModel`] does not support saving and loading mixed adapters. The adapters should already be trained, and loading the model requires a script to be run each time.
+
+## Tips
+
+- Not all adapter types can be combined. See [`peft.tuners.mixed.COMPATIBLE_TUNER_TYPES`](https://github.com/huggingface/peft/blob/1c1c7fdaa6e6abaa53939b865dee1eded82ad032/src/peft/tuners/mixed/model.py#L35) for a list of compatible types. An error will be raised if you try to combine incompatible adapter types.
+- It is possible to mix multiple adapters of the same type which can be useful for combining adapters with very different configs.
+- If you want to combine a lot of different adapters, the most performant way to do it is to consecutively add the same adapter types. For example, add LoRA1, LoRA2, LoHa1, LoHa2 in this order, instead of LoRA1, LoHa1, LoRA2, and LoHa2. While the order can affect the output, there is no inherently *best* order, so it is best to choose the fastest one.

docs/source/developer_guides/model_merging.md

@@ -0,0 +1,157 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Model merging
+
+Training a model for each task can be costly, take up storage space, and the models aren't able to learn new information to improve their performance. Multitask learning can overcome some of these limitations by training a model to learn several tasks, but it is expensive to train and designing a dataset for it is challenging. *Model merging* offers a solution to these challenges by combining multiple pretrained models into one model, giving it the combined abilities of each individual model without any additional training.
+
+PEFT provides several methods for merging models like a linear or SVD combination. This guide focuses on two methods that are more efficient for merging LoRA adapters by eliminating redundant parameters:
+
+* [TIES](https://hf.co/papers/2306.01708) - TrIm, Elect, and Merge (TIES) is a three-step method for merging models. First, redundant parameters are trimmed, then conflicting signs are resolved into an aggregated vector, and finally the parameters whose signs are the same as the aggregate sign are averaged. This method takes into account that some values (redundant and sign disagreement) can degrade performance in the merged model.
+* [DARE](https://hf.co/papers/2311.03099) - Drop And REscale is a method that can be used to prepare for other model merging methods like TIES. It works by randomly dropping parameters according to a drop rate and rescaling the remaining parameters. This helps to reduce the number of redundant and potentially interfering parameters among multiple models.
+
+Models are merged with the [`~LoraModel.add_weighted_adapter`] method, and the specific model merging method is specified in the `combination_type` parameter.
+
+## Merge method
+
+With TIES and DARE, merging is enabled by setting `combination_type` and `density` to a value of the weights to keep from the individual models. For example, let's merge three finetuned [TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T](https://huggingface.co/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T) models: [tinyllama_lora_nobots](https://huggingface.co/smangrul/tinyllama_lora_norobots), [tinyllama_lora_sql](https://huggingface.co/smangrul/tinyllama_lora_sql), and [tinyllama_lora_adcopy](https://huggingface.co/smangrul/tinyllama_lora_adcopy).
+
+<Tip warninig={true}>
+
+When you're attempting to merge fully trained models with TIES, you should be aware of any special tokens each model may have added to the embedding layer which are not a part of the original checkpoint's vocabulary. This may cause an issue because each model may have added a special token to the same embedding position. If this is the case, you should use the [`~transformers.PreTrainedModel.resize_token_embeddings`] method to avoid merging the special tokens at the same embedding index.
+
+<br>
+
+This shouldn't be an issue if you're only merging LoRA adapters trained from the same base model.
+
+</Tip>
+
+Load a base model and can use the [`~PeftModel.load_adapter`] method to load and assign each adapter a name:
+
+```py
+from peft import PeftConfig, PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+
+config = PeftConfig.from_pretrained("smangrul/tinyllama_lora_norobots")
+model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, load_in_4bit=True, device_map="auto").eval()
+tokenizer = AutoTokenizer.from_pretrained("smangrul/tinyllama_lora_norobots")
+
+model = PeftModel.from_pretrained(model, "smangrul/tinyllama_lora_norobots", adapter_name="norobots")
+_ = model.load_adapter("smangrul/tinyllama_lora_sql", adapter_name="sql")
+_ = model.load_adapter("smangrul/tinyllama_lora_adcopy", adapter_name="adcopy")
+```
+
+Set the adapters, weights, `adapter_name`, `combination_type`, and `density` with the [`~LoraModel.add_weighted_adapter`] method.
+
+<hfoptions id="merge-method">
+<hfoption id="TIES">
+
+Weight values greater than `1.0` typically produce better results because they preserve the correct scale. A good default starting value for the weights is to set all values to `1.0`.
+
+```py
+adapters = ["norobots", "adcopy", "sql"]
+weights = [2.0, 1.0, 1.0]
+adapter_name = "merge"
+density = 0.2
+model.add_weighted_adapter(adapters, weights, adapter_name, combination_type="ties", density=density)
+```
+
+</hfoption>
+<hfoption id="DARE">
+
+```py
+adapters = ["norobots", "adcopy", "sql"]
+weights = [2.0, 0.3, 0.7]
+adapter_name = "merge"
+density = 0.2
+model.add_weighted_adapter(adapters, weights, adapter_name, combination_type="dare_ties", density=density)
+```
+
+</hfoption>
+</hfoptions>
+
+Set the newly merged model as the active model with the [`~LoraModel.set_adapter`] method.
+
+```py
+model.set_adapter("merge")
+```
+
+Now you can use the merged model as an instruction-tuned model to write ad copy or SQL queries!
+
+<hfoptions id="ties">
+<hfoption id="instruct">
+
+```py
+messages = [
+    {"role": "user", "content": "Write an essay about Generative AI."},
+]
+text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
+inputs = tokenizer(text, return_tensors="pt")
+inputs = {k: v.to("cuda") for k, v in inputs.items()}
+outputs = model.generate(**inputs, max_new_tokens=256, do_sample=True, top_p=0.95, temperature=0.2, repetition_penalty=1.2, eos_token_id=tokenizer.eos_token_id)
+print(tokenizer.decode(outputs[0]))
+```
+
+</hfoption>
+<hfoption id="ad copy">
+
+```py
+messages = [
+    {"role": "system", "content": "Create a text ad given the following product and description."},
+    {"role": "user", "content": "Product: Sony PS5 PlayStation Console\nDescription: The PS5 console unleashes new gaming possibilities that you never anticipated."},
+]
+text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
+inputs = tokenizer(text, return_tensors="pt")
+inputs = {k: v.to("cuda") for k, v in inputs.items()}
+outputs = model.generate(**inputs, max_new_tokens=128, do_sample=True, top_p=0.95, temperature=0.2, repetition_penalty=1.2, eos_token_id=tokenizer.eos_token_id)
+print(tokenizer.decode(outputs[0]))
+```
+
+</hfoption>
+<hfoption id="SQL">
+
+```py
+text = """Table: 2-11365528-2
+Columns: ['Team', 'Head Coach', 'President', 'Home Ground', 'Location']
+Natural Query: Who is the Head Coach of the team whose President is Mario Volarevic?
+SQL Query:"""
+
+inputs = tokenizer(text, return_tensors="pt")
+inputs = {k: v.to("cuda") for k, v in inputs.items()}
+outputs = model.generate(**inputs, max_new_tokens=64, repetition_penalty=1.1, eos_token_id=tokenizer("</s>").input_ids[-1])
+print(tokenizer.decode(outputs[0]))
+```
+
+</hfoption>
+</hfoptions>
+
+
+## Merging (IA)³ Models
+The (IA)³ models facilitate linear merging of adapters. To merge adapters in an (IA)³ model, utilize the `add_weighted_adapter` method from the `IA3Model` class. This method is analogous to the `add_weighted_adapter` method used in `LoraModel`, with the key difference being the absence of the `combination_type` parameter. For example, to merge three (IA)³ adapters into a PEFT model, you would proceed as follows:
+
+```py
+adapters = ["adapter1", "adapter2", "adapter3"]
+weights = [0.4, 0.3, 0.3]
+adapter_name = "merge"
+model.add_weighted_adapter(adapters, weights, adapter_name)
+```
+
+It is recommended that the weights sum to 1.0 to preserve the scale of the model. The merged model can then be set as the active model using the `set_adapter` method:
+
+```py
+model.set_adapter("merge")
+```

docs/source/developer_guides/quantization.md

@@ -0,0 +1,195 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Quantization
+
+Quantization represents data with fewer bits, making it a useful technique for reducing memory-usage and accelerating inference especially when it comes to large language models (LLMs). There are several ways to quantize a model including:
+
+* optimizing which model weights are quantized with the [AWQ](https://hf.co/papers/2306.00978) algorithm
+* independently quantizing each row of a weight matrix with the [GPTQ](https://hf.co/papers/2210.17323) algorithm
+* quantizing to 8-bit and 4-bit precision with the [bitsandbytes](https://github.com/TimDettmers/bitsandbytes) library
+* quantizing to as low as 2-bit precision with the [AQLM](https://arxiv.org/abs/2401.06118) algorithm
+
+However, after a model is quantized it isn't typically further trained for downstream tasks because training can be unstable due to the lower precision of the weights and activations. But since PEFT methods only add *extra* trainable parameters, this allows you to train a quantized model with a PEFT adapter on top! Combining quantization with PEFT can be a good strategy for training even the largest models on a single GPU. For example, [QLoRA](https://hf.co/papers/2305.14314) is a method that quantizes a model to 4-bits and then trains it with LoRA. This method allows you to finetune a 65B parameter model on a single 48GB GPU!
+
+In this guide, you'll see how to quantize a model to 4-bits and train it with LoRA.
+
+## Quantize a model
+
+[bitsandbytes](https://github.com/TimDettmers/bitsandbytes) is a quantization library with a Transformers integration. With this integration, you can quantize a model to 8 or 4-bits and enable many other options by configuring the [`~transformers.BitsAndBytesConfig`] class. For example, you can:
+
+* set `load_in_4bit=True` to quantize the model to 4-bits when you load it
+* set `bnb_4bit_quant_type="nf4"` to use a special 4-bit data type for weights initialized from a normal distribution
+* set `bnb_4bit_use_double_quant=True` to use a nested quantization scheme to quantize the already quantized weights
+* set `bnb_4bit_compute_dtype=torch.bfloat16` to use bfloat16 for faster computation
+
+```py
+import torch
+from transformers import BitsAndBytesConfig
+
+config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_use_double_quant=True,
+    bnb_4bit_compute_dtype=torch.bfloat16,
+)
+```
+
+Pass the `config` to the [`~transformers.AutoModelForCausalLM.from_pretrained`] method.
+
+```py
+from transformers import AutoModelForCausalLM
+
+model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1", quantization_config=config)
+```
+
+Next, you should call the [`~peft.utils.prepare_model_for_kbit_training`] function to preprocess the quantized model for training.
+
+```py
+from peft import prepare_model_for_kbit_training
+
+model = prepare_model_for_kbit_training(model)
+```
+
+Now that the quantized model is ready, let's set up a configuration.
+
+## LoraConfig
+
+Create a [`LoraConfig`] with the following parameters (or choose your own):
+
+```py
+from peft import LoraConfig
+
+config = LoraConfig(
+    r=16,
+    lora_alpha=8,
+    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM"
+)
+```
+
+Then use the [`get_peft_model`] function to create a [`PeftModel`] from the quantized model and configuration.
+
+```py
+from peft import get_peft_model
+
+model = get_peft_model(model, config)
+```
+
+You're all set for training with whichever training method you prefer!
+
+### LoftQ initialization
+
+[LoftQ](https://hf.co/papers/2310.08659) initializes LoRA weights such that the quantization error is minimized, and it can improve performance when training quantized models. To get started, follow [these instructions](https://github.com/huggingface/peft/tree/main/examples/loftq_finetuning).
+
+In general, for LoftQ to work best, it is recommended to target as many layers with LoRA as possible, since those not targeted cannot have LoftQ applied. This means that passing `LoraConfig(..., target_modules="all-linear")` will most likely give the best results. Also, you should use `nf4` as quant type in your quantization config when using 4bit quantization, i.e. `BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4")`.
+
+### QLoRA-style training
+
+QLoRA adds trainable weights to all the linear layers in the transformer architecture. Since the attribute names for these linear layers can vary across architectures, set `target_modules` to `"all-linear"` to add LoRA to all the linear layers:
+
+```py
+config = LoraConfig(target_modules="all-linear", ...)
+```
+
+## AQLM quantization
+
+Additive Quantization of Language Models ([AQLM](https://arxiv.org/abs/2401.06118)) is a Large Language Models compression method. It quantizes multiple weights together and takes advantage of interdependencies between them. AQLM represents groups of 8-16 weights as a sum of multiple vector codes. This allows it to compress models down to as low as 2-bit with considerably low accuracy losses.
+
+Since the AQLM quantization process is computationally expensive, a use of prequantized models is recommended. A partial list of available models can be found in the official aqlm [repository](https://github.com/Vahe1994/AQLM).
+
+The models support LoRA adapter tuning. To tune the quantized model you'll need to install the `aqlm` inference library: `pip install aqlm>=1.0.2`. Finetuned LoRA adapters shall be saved separately, as merging them with AQLM quantized weights is not possible.
+
+```py
+quantized_model = AutoModelForCausalLM.from_pretrained(
+    "BlackSamorez/Mixtral-8x7b-AQLM-2Bit-1x16-hf-test-dispatch",
+    torch_dtype="auto", device_map="auto", low_cpu_mem_usage=True,
+)
+
+peft_config = LoraConfig(...)
+
+quantized_model = get_peft_model(quantized_model, peft_config)
+```
+
+You can refer to the [Google Colab](https://colab.research.google.com/drive/12GTp1FCj5_0SnnNQH18h_2XFh9vS_guX?usp=sharing) example for an overview of AQLM+LoRA finetuning.
+
+## EETQ quantization
+
+You can also perform LoRA fine-tuning on EETQ quantized models. [EETQ](https://github.com/NetEase-FuXi/EETQ) package offers simple and efficient way to perform 8-bit quantization, which is claimed to be faster than the `LLM.int8()` algorithm. First, make sure that you have a transformers version that is compatible with EETQ (e.g. by installing it from latest pypi or from source).
+
+```py
+import torch
+from transformers import EetqConfig
+
+config = EetqConfig("int8")
+```
+
+Pass the `config` to the [`~transformers.AutoModelForCausalLM.from_pretrained`] method.
+
+```py
+from transformers import AutoModelForCausalLM
+
+model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1", quantization_config=config)
+```
+
+and create a `LoraConfig` and pass it to `get_peft_model`:
+
+```py
+from peft import LoraConfig, get_peft_model
+
+config = LoraConfig(
+    r=16,
+    lora_alpha=8,
+    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM"
+)
+
+model = get_peft_model(model, config)
+```
+
+## HQQ quantization
+
+The models that is quantized using Half-Quadratic Quantization of Large Machine Learning Models ([HQQ](https://mobiusml.github.io/hqq_blog/)) support LoRA adapter tuning. To tune the quantized model, you'll need to install the `hqq` library with: `pip install hqq`.
+
+```python
+from hqq.engine.hf import HQQModelForCausalLM
+
+quantized_model = HQQModelForCausalLM.from_quantized(save_dir_or_hfhub, device='cuda')
+peft_config = LoraConfig(...)
+quantized_model = get_peft_model(quantized_model, peft_config)
+```
+
+Or using transformers version that is compatible with HQQ (e.g. by installing it from latest pypi or from source).
+
+```python
+from transformers import HqqConfig, AutoModelForCausalLM
+
+quant_config = HqqConfig(nbits=4, group_size=64)
+quantized_model = AutoModelForCausalLM.from_pretrained(save_dir_or_hfhub, device_map=device_map, quantization_config=quant_config)
+peft_config = LoraConfig(...)
+quantized_model = get_peft_model(quantized_model, peft_config)
+```
+
+## Next steps
+
+If you're interested in learning more about quantization, the following may be helpful:
+
+* Learn more about details about QLoRA and check out some benchmarks on its impact in the [Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA](https://huggingface.co/blog/4bit-transformers-bitsandbytes) blog post.
+* Read more about different quantization schemes in the Transformers [Quantization](https://hf.co/docs/transformers/main/quantization) guide.

docs/source/developer_guides/torch_compile.md

@@ -0,0 +1,76 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# torch.compile
+
+In PEFT, [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) works for some but not all features. The reason why it won't always work is because PEFT is highly dynamic in certain places (loading and switching between multiple adapters, for instance), which can cause trouble for `torch.compile`. In other places, `torch.compile` may work, but won't be as fast as expected because of graph breaks.
+
+If you don't see an error, it doesn't necessarily mean that `torch.compile` worked correctly. It might give you an output, but the output is incorrect. This guide describes what works with `torch.compile` and what doesn't.
+
+> [!TIP]
+> Unless indicated otherwise, the default `torch.compile` settings were used.
+
+## Training and inference with `torch.compile`
+
+These features **work** with `torch.compile`. Everything listed below was tested with a causal LM:
+
+- Training with `Trainer` from 🤗 transformers
+- Training with a custom PyTorch loop
+- Inference
+- Generation
+
+The following adapters were tested successfully:
+
+- AdaLoRA
+- BOFT
+- IA³
+- Layer Norm Tuning
+- LoHa
+- LoRA
+- LoRA + DoRA
+- OFT
+- VeRA
+- HRA
+
+The following adapters **don't work** correctly for training or inference when using `torch.compile`:
+
+- LoKr
+- LoRA targeting embedding layers
+
+## Advanced PEFT features with `torch.compile`
+
+Below are some of the more advanced PEFT features that **work**. They were all tested with LoRA.
+
+- `modules_to_save` (i.e. `config = LoraConfig(..., modules_to_save=...)`)
+- Merging adapters (one or multiple)
+- Merging multiple adapters into one adapter (i.e. calling `model.add_weighted_adapter(...)`)
+
+Generally, we can expect that if a feature works correctly with LoRA and is also supported by other adapter types, it should also work for that adapter type.
+
+The more advanced PEFT features below **don't work** in conjunction with `torch.compile`. Tests were run with LoRA:
+
+- Using PEFT adapters with quantization (bitsandbytes)
+- Inference with multiple adapters
+- Unloading (i.e. calling `model.merge_and_unload()`)
+- Disabling adapters (i.e. using `with model.disable_adapter()`)
+- Mixed adapter batches (i.e. calling `model(batch, adapter_names=["__base__", "default", "other", ...])`)
+
+## Test cases
+
+All the use cases listed above are tested inside of [`peft/tests/test_torch_compile.py`](https://github.com/huggingface/peft/blob/main/tests/test_torch_compile.py). If you want to check in more detail how we tested a certain feature, please go to that file and check the test that corresponds to your use case.
+
+> [!TIP]
+> If you have another use case where you know that `torch.compile` does or does not work with PEFT, please contribute by letting us know or by opening a PR to add this use case to the covered test cases.

docs/source/developer_guides/troubleshooting.md

@@ -0,0 +1,286 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Troubleshooting
+
+If you encounter any issue when using PEFT, please check the following list of common issues and their solutions.
+
+## Examples don't work
+
+Examples often rely on the most recent package versions, so please ensure they're up-to-date. In particular, check the following package versions:
+
+- `peft`
+- `transformers`
+- `accelerate`
+- `torch`
+
+In general, you can update the package version by running this command inside your Python environment:
+
+```bash
+python -m pip install -U <package_name>
+```
+
+Installing PEFT from source is useful for keeping up with the latest developments:
+
+```bash
+python -m pip install git+https://github.com/huggingface/peft
+```
+
+## ValueError: Attempting to unscale FP16 gradients
+
+This error probably occurred because the model was loaded with `torch_dtype=torch.float16` and then used in an automatic mixed precision (AMP) context, e.g. by setting `fp16=True` in the [`~transformers.Trainer`] class from 🤗 Transformers. The reason is that when using AMP, trainable weights should never use fp16. To make this work without loading the whole model in fp32, add the following to your code:
+
+```python
+peft_model = get_peft_model(...)
+
+# add this:
+for param in model.parameters():
+    if param.requires_grad:
+        param.data = param.data.float()
+
+# proceed as usual
+trainer = Trainer(model=peft_model, fp16=True, ...)
+trainer.train()
+```
+
+Alternatively, you can use the [`~utils.cast_mixed_precision_params`] function to correctly cast the weights:
+
+```python
+from peft import cast_mixed_precision_params
+
+peft_model = get_peft_model(...)
+cast_mixed_precision_params(peft_model, dtype=torch.float16)
+
+# proceed as usual
+trainer = Trainer(model=peft_model, fp16=True, ...)
+trainer.train()
+```
+
+<Tip>
+
+Starting from PEFT verion v0.12.0, PEFT automatically promotes the dtype of adapter weights from `torch.float16` and `torch.bfloat16` to `torch.float32` where appropriate. To _prevent_ this behavior, you can pass `autocast_adapter_dtype=False` to [`~get_peft_model`], to [`~PeftModel.from_pretrained`], and to [`~PeftModel.load_adapter`].
+
+</Tip>
+
+## Bad results from a loaded PEFT model
+
+There can be several reasons for getting a poor result from a loaded PEFT model which are listed below. If you're still unable to troubleshoot the problem, see if anyone else had a similar [issue](https://github.com/huggingface/peft/issues) on GitHub, and if you can't find any, open a new issue.
+
+When opening an issue, it helps a lot if you provide a minimal code example that reproduces the issue. Also, please report if the loaded model performs at the same level as the model did before fine-tuning, if it performs at a random level, or if it is only slightly worse than expected. This information helps us identify the problem more quickly.
+
+### Random deviations
+
+If your model outputs are not exactly the same as previous runs, there could be an issue with random elements. For example:
+
+1. please ensure it is in `.eval()` mode, which is important, for instance, if the model uses dropout
+2. if you use [`~transformers.GenerationMixin.generate`] on a language model, there could be random sampling, so obtaining the same result requires setting a random seed
+3. if you used quantization and merged the weights, small deviations are expected due to rounding errors
+
+### Incorrectly loaded model
+
+Please ensure that you load the model correctly. A common error is trying to load a _trained_ model with [`get_peft_model`] which is incorrect. Instead, the loading code should look like this:
+
+```python
+from peft import PeftModel, PeftConfig
+
+base_model = ...  # to load the base model, use the same code as when you trained it
+config = PeftConfig.from_pretrained(peft_model_id)
+peft_model = PeftModel.from_pretrained(base_model, peft_model_id)
+```
+
+### Randomly initialized layers
+
+For some tasks, it is important to correctly configure `modules_to_save` in the config to account for randomly initialized layers. 
+
+As an example, this is necessary if you use LoRA to fine-tune a language model for sequence classification because 🤗 Transformers adds a randomly initialized classification head on top of the model. If you do not add this layer to `modules_to_save`, the classification head won't be saved. The next time you load the model, you'll get a _different_ randomly initialized classification head, resulting in completely different results.
+
+PEFT tries to correctly guess the `modules_to_save` if you provide the `task_type` argument in the config. This should work for transformers models that follow the standard naming scheme. It is always a good idea to double check though because we can't guarantee all models follow the naming scheme.
+
+When you load a transformers model that has randomly initialized layers, you should see a warning along the lines of:
+
+```
+Some weights of <MODEL> were not initialized from the model checkpoint at <ID> and are newly initialized: [<LAYER_NAMES>].
+You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+```
+
+The mentioned layers should be added to `modules_to_save` in the config to avoid the described problem.
+
+### Extending the vocabulary
+
+For many language fine-tuning tasks, extending the model's vocabulary is necessary since new tokens are being introduced. This requires extending the embedding layer to account for the new tokens and also storing the embedding layer in addition to the adapter weights when saving the adapter.
+
+Save the embedding layer by adding it to the `target_modules` of the config. The embedding layer name must follow the standard naming scheme from Transformers. For example, the Mistral config could look like this:
+
+```python
+config = LoraConfig(..., target_modules=["embed_tokens", "lm_head", "q_proj", "v_proj"])
+```
+
+Once added to `target_modules`, PEFT automatically stores the embedding layer when saving the adapter if the model has the [`~transformers.PreTrainedModel.get_input_embeddings`] and [`~transformers.PreTrainedModel.get_output_embeddings`]. This is generally the case for Transformers models.
+
+If the model's embedding layer doesn't follow the Transformer's naming scheme, you can still save it by manually passing `save_embedding_layers=True` when saving the adapter:
+
+```python
+model = get_peft_model(...)
+# train the model
+model.save_pretrained("my_adapter", save_embedding_layers=True)
+```
+
+For inference, load the base model first and resize it the same way you did before you trained the model. After you've resized the base model, you can load the PEFT checkpoint.
+
+For a complete example, please check out [this notebook](https://github.com/huggingface/peft/blob/main/examples/causal_language_modeling/peft_lora_clm_with_additional_tokens.ipynb).
+
+### Check layer and model status
+
+Sometimes a PEFT model can end up in a bad state, especially when handling multiple adapters. There can be some confusion around what adapters exist, which one is active, which one is merged, etc. To help investigate this issue, call the [`~peft.PeftModel.get_layer_status`] and the [`~peft.PeftModel.get_model_status`] methods. 
+
+The [`~peft.PeftModel.get_layer_status`] method gives you a detailed overview of each targeted layer's active, merged, and available adapters.
+
+```python
+>>> from transformers import AutoModel
+>>> from peft import get_peft_model, LoraConfig
+
+>>> model_id = "google/flan-t5-small"
+>>> model = AutoModel.from_pretrained(model_id)
+>>> model = get_peft_model(model, LoraConfig())
+
+>>> model.get_layer_status()
+[TunerLayerStatus(name='model.encoder.block.0.layer.0.SelfAttention.q',
+                  module_type='lora.Linear',
+                  enabled=True,
+                  active_adapters=['default'],
+                  merged_adapters=[],
+                  requires_grad={'default': True},
+                  available_adapters=['default']),
+ TunerLayerStatus(name='model.encoder.block.0.layer.0.SelfAttention.v',
+                  module_type='lora.Linear',
+                  enabled=True,
+                  active_adapters=['default'],
+                  merged_adapters=[],
+                  requires_grad={'default': True},
+                  available_adapters=['default']),
+...]
+
+>>> model.get_model_status()
+TunerModelStatus(
+    base_model_type='T5Model',
+    adapter_model_type='LoraModel',
+    peft_types={'default': 'LORA'},
+    trainable_params=344064,
+    total_params=60855680,
+    num_adapter_layers=48,
+    enabled=True,
+    active_adapters=['default'],
+    merged_adapters=[],
+    requires_grad={'default': True},
+    available_adapters=['default'],
+)
+```
+
+In the model state output, you should look out for entries that say `"irregular"`. This means PEFT detected an inconsistent state in the model. For instance, if `merged_adapters="irregular"`, it means that for at least one adapter, it was merged on some target modules but not on others. The inference results will most likely be incorrect as a result.
+
+The best way to resolve this issue is to reload the whole model and adapter checkpoint(s). Ensure that you don't perform any incorrect operations on the model, e.g. manually merging adapters on some modules but not others.
+
+Convert the layer status into a pandas `DataFrame` for an easier visual inspection.
+
+```python
+from dataclasses import asdict
+import pandas as pd
+
+df = pd.DataFrame(asdict(layer) for layer in model.get_layer_status())
+```
+
+It is possible to get this information for non-PEFT models if they are using PEFT layers under the hood, but some information like the `base_model_type` or the `peft_types` cannot be determined in that case. As an example, you can call this on a [diffusers](https://huggingface.co/docs/diffusers/index) model like so:
+
+```python
+>>> import torch
+>>> from diffusers import StableDiffusionPipeline
+>>> from peft import get_model_status, get_layer_status
+
+>>> path = "runwayml/stable-diffusion-v1-5"
+>>> lora_id = "takuma104/lora-test-text-encoder-lora-target"
+>>> pipe = StableDiffusionPipeline.from_pretrained(path, torch_dtype=torch.float16)
+>>> pipe.load_lora_weights(lora_id, adapter_name="adapter-1")
+>>> pipe.load_lora_weights(lora_id, adapter_name="adapter-2")
+>>> pipe.set_lora_device(["adapter-2"], "cuda")
+>>> get_layer_status(pipe.text_encoder)
+[TunerLayerStatus(name='text_model.encoder.layers.0.self_attn.k_proj',
+                  module_type='lora.Linear',
+                  enabled=True,
+                  active_adapters=['adapter-2'],
+                  merged_adapters=[],
+                  requires_grad={'adapter-1': False, 'adapter-2': True},
+                  available_adapters=['adapter-1', 'adapter-2'],
+                  devices={'adapter-1': ['cpu'], 'adapter-2': ['cuda']}),
+ TunerLayerStatus(name='text_model.encoder.layers.0.self_attn.v_proj',
+                  module_type='lora.Linear',
+                  enabled=True,
+                  active_adapters=['adapter-2'],
+                  merged_adapters=[],
+                  requires_grad={'adapter-1': False, 'adapter-2': True},
+                  devices={'adapter-1': ['cpu'], 'adapter-2': ['cuda']}),
+...]
+
+>>> get_model_status(pipe.unet)
+TunerModelStatus(
+    base_model_type='other',
+    adapter_model_type='None',
+    peft_types={},
+    trainable_params=797184,
+    total_params=861115332,
+    num_adapter_layers=128,
+    enabled=True,
+    active_adapters=['adapter-2'],
+    merged_adapters=[],
+    requires_grad={'adapter-1': False, 'adapter-2': True},
+    available_adapters=['adapter-1', 'adapter-2'],
+    devices={'adapter-1': ['cpu'], 'adapter-2': ['cuda']},
+)
+```
+
+## Speed
+
+### Loading adapter weights is slow
+
+Loading adapters like LoRA weights should generally be fast compared to loading the base model. However, there can be use cases where the adapter weights are quite large or where users need to load a large number of adapters -- the loading time can add up in this case. The reason for this is that the adapter weights are first initialized and then overridden by the loaded weights, which is wasteful. To speed up the loading time, you can pass the `low_cpu_mem_usage=True` argument to [`~PeftModel.from_pretrained`] and [`~PeftModel.load_adapter`].
+
+<Tip>
+
+If this option works well across different use casese, it may become the default for adapter loading in the future.
+
+</Tip>
+
+
+## Reproducibility
+
+### Models using batch norm
+
+When loading a trained PEFT model where the base model uses batch norm (e.g. `torch.nn.BatchNorm1d` or `torch.nn.BatchNorm2d`), you may find that you cannot reproduce the exact same outputs. This is because the batch norm layers keep track of running stats during training, but these stats are not part of the PEFT checkpoint. Therefore, when you load the PEFT model, the running stats of the base model will be used (i.e. from before training with PEFT).
+
+Depending on your use case, this may not be a big deal. If, however, you need your outputs to be 100% reproducible, you can achieve this by adding the batch norm layers to `modules_to_save`. Below is an example of this using resnet and LoRA. Notice that we set `modules_to_save=["classifier", "normalization"]`. We need the `"classifier"` argument because our task is image classification, and we add the `"normalization"` argument to ensure that the batch norm layers are saved in the PEFT checkpoint.
+
+```python
+from transformers import AutoModelForImageClassification
+from peft import LoraConfig, get_peft_model
+
+model_id = "microsoft/resnet-18"
+base_model = AutoModelForImageClassification.from_pretrained(self.model_id)
+config = LoraConfig(
+    target_modules=["convolution"],
+    modules_to_save=["classifier", "normalization"],
+),
+```
+
+Depending on the type of model you use, the batch norm layers could have different names than `"normalization"`, so please ensure that the name matches your model architecture.

docs/source/index.md

@@ -0,0 +1,49 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# PEFT
+
+🤗 PEFT (Parameter-Efficient Fine-Tuning) is a library for efficiently adapting large pretrained models to various downstream applications without fine-tuning all of a model's parameters because it is prohibitively costly. PEFT methods only fine-tune a small number of (extra) model parameters - significantly decreasing computational and storage costs - while yielding performance comparable to a fully fine-tuned model. This makes it more accessible to train and store large language models (LLMs) on consumer hardware.
+
+PEFT is integrated with the Transformers, Diffusers, and Accelerate libraries to provide a faster and easier way to load, train, and use large models for inference.
+
+<div class="mt-10">
+  <div class="w-full flex flex-col space-y-4 md:space-y-0 md:grid md:grid-cols-2 md:gap-y-4 md:gap-x-5">
+    <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="quicktour"
+      ><div class="w-full text-center bg-gradient-to-br from-blue-400 to-blue-500 rounded-lg py-1.5 font-semibold mb-5 text-white text-lg leading-relaxed">Get started</div>
+      <p class="text-gray-700">Start here if you're new to 🤗 PEFT to get an overview of the library's main features, and how to train a model with a PEFT method.</p>
+    </a>
+    <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="./task_guides/image_classification_lora"
+      ><div class="w-full text-center bg-gradient-to-br from-indigo-400 to-indigo-500 rounded-lg py-1.5 font-semibold mb-5 text-white text-lg leading-relaxed">How-to guides</div>
+      <p class="text-gray-700">Practical guides demonstrating how to apply various PEFT methods across different types of tasks like image classification, causal language modeling, automatic speech recognition, and more. Learn how to use 🤗 PEFT with the DeepSpeed and Fully Sharded Data Parallel scripts.</p>
+    </a>
+    <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="./conceptual_guides/lora"
+      ><div class="w-full text-center bg-gradient-to-br from-pink-400 to-pink-500 rounded-lg py-1.5 font-semibold mb-5 text-white text-lg leading-relaxed">Conceptual guides</div>
+      <p class="text-gray-700">Get a better theoretical understanding of how LoRA and various soft prompting methods help reduce the number of trainable parameters to make training more efficient.</p>
+   </a>
+    <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="./package_reference/config"
+      ><div class="w-full text-center bg-gradient-to-br from-purple-400 to-purple-500 rounded-lg py-1.5 font-semibold mb-5 text-white text-lg leading-relaxed">Reference</div>
+      <p class="text-gray-700">Technical descriptions of how 🤗 PEFT classes and methods work.</p>
+    </a>
+  </div>
+</div>
+
+<iframe
+	src="https://stevhliu-peft-methods.hf.space"
+	frameborder="0"
+	width="850"
+	height="620"
+></iframe>

docs/source/install.md

@@ -0,0 +1,47 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Installation
+
+Before you start, you will need to setup your environment, install the appropriate packages, and configure 🤗 PEFT. 🤗 PEFT is tested on **Python 3.8+**.
+
+🤗 PEFT is available on PyPI, as well as GitHub:
+
+## PyPI
+
+To install 🤗 PEFT from PyPI:
+
+```bash
+pip install peft
+```
+
+## Source
+
+New features that haven't been released yet are added every day, which also means there may be some bugs. To try them out, install from the GitHub repository:
+
+```bash
+pip install git+https://github.com/huggingface/peft
+```
+
+If you're working on contributing to the library or wish to play with the source code and see live 
+results as you run the code, an editable version can be installed from a locally-cloned version of the 
+repository:
+
+```bash
+git clone https://github.com/huggingface/peft
+cd peft
+pip install -e .
+```

docs/source/package_reference/adalora.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# AdaLoRA
+
+[AdaLoRA](https://hf.co/papers/2303.10512) is a method for optimizing the number of trainable parameters to assign to weight matrices and layers, unlike LoRA, which distributes parameters evenly across all modules. More parameters are budgeted for important weight matrices and layers while less important ones receive fewer parameters.
+
+The abstract from the paper is:
+
+*Fine-tuning large pre-trained language models on downstream tasks has become an important paradigm in NLP. However, common practice fine-tunes all of the parameters in a pre-trained model, which becomes prohibitive when a large number of downstream tasks are present. Therefore, many fine-tuning methods are proposed to learn incremental updates of pre-trained weights in a parameter efficient way, e.g., low-rank increments. These methods often evenly distribute the budget of incremental updates across all pre-trained weight matrices, and overlook the varying importance of different weight parameters. As a consequence, the fine-tuning performance is suboptimal. To bridge this gap, we propose AdaLoRA, which adaptively allocates the parameter budget among weight matrices according to their importance score. In particular, AdaLoRA parameterizes the incremental updates in the form of singular value decomposition. Such a novel approach allows us to effectively prune the singular values of unimportant updates, which is essentially to reduce their parameter budget but circumvent intensive exact SVD computations. We conduct extensive experiments with several pre-trained models on natural language processing, question answering, and natural language generation to validate the effectiveness of AdaLoRA. Results demonstrate that AdaLoRA manifests notable improvement over baselines, especially in the low budget settings. Our code is publicly available at https://github.com/QingruZhang/AdaLoRA*.
+
+## AdaLoraConfig
+
+[[autodoc]] tuners.adalora.config.AdaLoraConfig
+
+## AdaLoraModel
+
+[[autodoc]] tuners.adalora.model.AdaLoraModel

docs/source/package_reference/adapter_utils.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# LyCORIS
+
+[LyCORIS](https://hf.co/papers/2309.14859) (Lora beYond Conventional methods, Other Rank adaptation Implementations for Stable diffusion) are LoRA-like matrix decomposition adapters that modify the cross-attention layer of the UNet. The [LoHa](loha) and [LoKr](lokr) methods inherit from the `Lycoris` classes here.
+
+## LycorisConfig
+
+[[autodoc]] tuners.lycoris_utils.LycorisConfig
+
+## LycorisLayer
+
+[[autodoc]] tuners.lycoris_utils.LycorisLayer
+
+## LycorisTuner
+
+[[autodoc]] tuners.lycoris_utils.LycorisTuner

docs/source/package_reference/auto_class.md

@@ -0,0 +1,48 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# AutoPeftModels
+
+The `AutoPeftModel` classes loads the appropriate PEFT model for the task type by automatically inferring it from the configuration file. They are designed to quickly and easily load a PEFT model in a single line of code without having to worry about which exact model class you need or manually loading a [`PeftConfig`].
+
+## AutoPeftModel
+
+[[autodoc]] auto.AutoPeftModel
+    - from_pretrained
+
+## AutoPeftModelForCausalLM
+
+[[autodoc]] auto.AutoPeftModelForCausalLM
+
+## AutoPeftModelForSeq2SeqLM
+
+[[autodoc]] auto.AutoPeftModelForSeq2SeqLM
+
+## AutoPeftModelForSequenceClassification
+
+[[autodoc]] auto.AutoPeftModelForSequenceClassification
+
+## AutoPeftModelForTokenClassification
+
+[[autodoc]] auto.AutoPeftModelForTokenClassification
+
+## AutoPeftModelForQuestionAnswering
+
+[[autodoc]] auto.AutoPeftModelForQuestionAnswering
+
+## AutoPeftModelForFeatureExtraction
+
+[[autodoc]] auto.AutoPeftModelForFeatureExtraction

docs/source/package_reference/boft.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# BOFT
+
+[Orthogonal Butterfly (BOFT)](https://hf.co/papers/2311.06243) is a generic method designed for finetuning foundation models. It improves the paramter efficiency of the finetuning paradigm -- Orthogonal Finetuning (OFT), by taking inspiration from Cooley-Tukey fast Fourier transform, showing favorable results across finetuning different foundation models, including large vision transformers, large language models and text-to-image diffusion models.
+
+The abstract from the paper is:
+
+*Large foundation models are becoming ubiquitous, but training them from scratch is prohibitively expensive. Thus, efficiently adapting these powerful models to downstream tasks is increasingly important. In this paper, we study a principled finetuning paradigm -- Orthogonal Finetuning (OFT) -- for downstream task adaptation. Despite demonstrating good generalizability, OFT still uses a fairly large number of trainable parameters due to the high dimensionality of orthogonal matrices. To address this, we start by examining OFT from an information transmission perspective, and then identify a few key desiderata that enable better parameter-efficiency. Inspired by how the Cooley-Tukey fast Fourier transform algorithm enables efficient information transmission, we propose an efficient orthogonal parameterization using butterfly structures. We apply this parameterization to OFT, creating a novel parameter-efficient finetuning method, called Orthogonal Butterfly (BOFT). By subsuming OFT as a special case, BOFT introduces a generalized orthogonal finetuning framework. Finally, we conduct an extensive empirical study of adapting large vision transformers, large language models, and text-to-image diffusion models to various downstream tasks in vision and language*.
+
+## BOFTConfig
+
+[[autodoc]] tuners.boft.config.BOFTConfig
+
+## BOFTModel
+
+[[autodoc]] tuners.boft.model.BOFTModel

docs/source/package_reference/config.md

@@ -0,0 +1,22 @@
+<!--⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# Configuration
+
+[`PeftConfigMixin`] is the base configuration class for storing the adapter configuration of a [`PeftModel`], and [`PromptLearningConfig`] is the base configuration class for soft prompt methods (p-tuning, prefix tuning, and prompt tuning). These base classes contain methods for saving and loading model configurations from the Hub, specifying the PEFT method to use, type of task to perform, and model configurations like number of layers and number of attention heads.
+
+## PeftConfigMixin
+
+[[autodoc]] config.PeftConfigMixin
+    - all
+
+## PeftConfig
+
+[[autodoc]] PeftConfig
+    - all
+
+## PromptLearningConfig
+
+[[autodoc]] PromptLearningConfig
+    - all

docs/source/package_reference/fourierft.md

@@ -0,0 +1,38 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# FourierFT: Discrete Fourier Transformation Fine-Tuning
+
+[FourierFT](https://huggingface.co/papers/2405.03003) is a parameter-efficient fine-tuning technique that leverages Discrete Fourier Transform to compress the model's tunable weights. This method outperforms LoRA in the GLUE benchmark and common ViT classification tasks using much less parameters.
+
+FourierFT currently has the following constraints:
+
+- Only `nn.Linear` layers are supported.
+- Quantized layers are not supported.
+
+If these constraints don't work for your use case, consider other methods instead.
+
+The abstract from the paper is:
+
+> Low-rank adaptation (LoRA) has recently gained much interest in fine-tuning foundation models. It effectively reduces the number of trainable parameters by incorporating low-rank matrices A and B to represent the weight change, i.e., Delta W=BA. Despite LoRA's progress, it faces storage challenges when handling extensive customization adaptations or larger base models. In this work, we aim to further compress trainable parameters by enjoying the powerful expressiveness of the Fourier transform. Specifically, we introduce FourierFT, which treats Delta W as a matrix in the spatial domain and learns only a small fraction of its spectral coefficients. With the trained spectral coefficients, we implement the inverse discrete Fourier transform to recover Delta W. Empirically, our FourierFT method shows comparable or better performance with fewer parameters than LoRA on various tasks, including natural language understanding, natural language generation, instruction tuning, and image classification. For example, when performing instruction tuning on the LLaMA2-7B model, FourierFT surpasses LoRA with only 0.064M trainable parameters, compared to LoRA's 33.5M.
+
+## FourierFTConfig
+
+[[autodoc]] tuners.fourierft.config.FourierFTConfig
+
+## FourierFTModel
+
+[[autodoc]] tuners.fourierft.model.FourierFTModel

docs/source/package_reference/helpers.md

@@ -0,0 +1,17 @@
+<!--⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# Helper methods
+
+A collection of helper functions for PEFT.
+
+## Checking if a model is a PEFT model
+
+[[autodoc]] helpers.check_if_peft_model
+    - all
+
+## Temporarily Rescaling Adapter Scale in LoraLayer Modules
+
+[[autodoc]] helpers.rescale_adapter_scale
+    - all

docs/source/package_reference/ia3.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# IA3
+
+Infused Adapter by Inhibiting and Amplifying Inner Activations, or [IA3](https://hf.co/papers/2205.05638), is a method that adds three learned vectors to rescale the keys and values of the self-attention and encoder-decoder attention layers, and the intermediate activation of the position-wise feed-forward network.
+
+The abstract from the paper is:
+
+*Few-shot in-context learning (ICL) enables pre-trained language models to perform a previously-unseen task without any gradient-based training by feeding a small number of training examples as part of the input. ICL incurs substantial computational, memory, and storage costs because it involves processing all of the training examples every time a prediction is made. Parameter-efficient fine-tuning (PEFT) (e.g. adapter modules, prompt tuning, sparse update methods, etc.) offers an alternative paradigm where a small set of parameters are trained to enable a model to perform the new task. In this paper, we rigorously compare few-shot ICL and PEFT and demonstrate that the latter offers better accuracy as well as dramatically lower computational costs. Along the way, we introduce a new PEFT method called (IA)^3 that scales activations by learned vectors, attaining stronger performance while only introducing a relatively tiny amount of new parameters. We also propose a simple recipe based on the T0 model called T-Few that can be applied to new tasks without task-specific tuning or modifications. We validate the effectiveness of T-Few on completely unseen tasks by applying it to the RAFT benchmark, attaining super-human performance for the first time and outperforming the state-of-the-art by 6% absolute. All of the code used in our experiments is publicly available*.
+
+## IA3Config
+
+[[autodoc]] tuners.ia3.config.IA3Config
+
+## IA3Model
+
+[[autodoc]] tuners.ia3.model.IA3Model

docs/source/package_reference/layernorm_tuning.md

@@ -0,0 +1,34 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# LayerNorm Tuning
+
+LayerNorm Tuning ([LN Tuning](https://huggingface.co/papers/2312.11420)) is a PEFT method that only fine-tunes the parameters of the LayerNorm layers in a model.
+The paper has tested the performance of this method on large language models and has shown that it can achieve strong performance with a significant reduction in the number of trainable parameters and GPU memory usage.
+However, the method is not limited to language models and can be applied to any model that uses LayerNorm layers.
+In this implementation, the default is that all layernorm layers inside a model is finetuned, but it could be used to target other layer types such as `MLP` or `Attention` layers, this can be done by specifying the `target_modules` in the `LNTuningConfig`.
+
+The abstract from the paper is:
+
+*This paper introduces an efficient strategy to transform Large Language Models (LLMs) into Multi-Modal Large Language Models (MLLMs). By conceptualizing this transformation as a domain adaptation process, i.e., transitioning from text understanding to embracing multiple modalities, we intriguingly note that, within each attention block, tuning LayerNorm suffices to yield strong performance. Moreover, when benchmarked against other tuning approaches like full parameter finetuning or LoRA, its benefits on efficiency are substantial. For example, when compared to LoRA on a 13B model scale, performance can be enhanced by an average of over 20% across five multi-modal tasks, and meanwhile, results in a significant reduction of trainable parameters by 41.9% and a decrease in GPU memory usage by 17.6%. On top of this LayerNorm strategy, we showcase that selectively tuning only with conversational data can improve efficiency further. Beyond these empirical outcomes, we provide a comprehensive analysis to explore the role of LayerNorm in adapting LLMs to the multi-modal domain and improving the expressive power of the model.*
+
+## LNTuningConfig
+
+[[autodoc]] tuners.ln_tuning.config.LNTuningConfig
+
+## LNTuningModel
+
+[[autodoc]] tuners.ln_tuning.model.LNTuningModel

docs/source/package_reference/llama_adapter.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Llama-Adapter
+
+[Llama-Adapter](https://hf.co/papers/2303.16199) is a PEFT method specifically designed for turning Llama into an instruction-following model. The Llama model is frozen and only a set of adaptation prompts prefixed to the input instruction tokens are learned. Since randomly initialized modules inserted into the model can cause the model to lose some of its existing knowledge, Llama-Adapter uses zero-initialized attention with zero gating to progressively add the instructional prompts to the model.
+
+The abstract from the paper is:
+
+*We present LLaMA-Adapter, a lightweight adaption method to efficiently fine-tune LLaMA into an instruction-following model. Using 52K self-instruct demonstrations, LLaMA-Adapter only introduces 1.2M learnable parameters upon the frozen LLaMA 7B model, and costs less than one hour for fine-tuning on 8 A100 GPUs. Specifically, we adopt a set of learnable adaption prompts, and prepend them to the input text tokens at higher transformer layers. Then, a zero-init attention mechanism with zero gating is proposed, which adaptively injects the new instructional cues into LLaMA, while effectively preserves its pre-trained knowledge. With efficient training, LLaMA-Adapter generates high-quality responses, comparable to Alpaca with fully fine-tuned 7B parameters. Furthermore, our approach can be simply extended to multi-modal input, e.g., images, for image-conditioned LLaMA, which achieves superior reasoning capacity on ScienceQA. We release our code at https://github.com/ZrrSkywalker/LLaMA-Adapter*.
+
+## AdaptionPromptConfig
+
+[[autodoc]] tuners.adaption_prompt.config.AdaptionPromptConfig
+
+## AdaptionPromptModel
+
+[[autodoc]] tuners.adaption_prompt.model.AdaptionPromptModel

docs/source/package_reference/loha.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# LoHa
+
+Low-Rank Hadamard Product ([LoHa](https://huggingface.co/papers/2108.06098)), is similar to LoRA except it approximates the large weight matrix with more low-rank matrices and combines them with the Hadamard product. This method is even more parameter-efficient than LoRA and achieves comparable performance.
+
+The abstract from the paper is:
+
+*In this work, we propose a communication-efficient parameterization, FedPara, for federated learning (FL) to overcome the burdens on frequent model uploads and downloads. Our method re-parameterizes weight parameters of layers using low-rank weights followed by the Hadamard product. Compared to the conventional low-rank parameterization, our FedPara method is not restricted to low-rank constraints, and thereby it has a far larger capacity. This property enables to achieve comparable performance while requiring 3 to 10 times lower communication costs than the model with the original layers, which is not achievable by the traditional low-rank methods. The efficiency of our method can be further improved by combining with other efficient FL optimizers. In addition, we extend our method to a personalized FL application, pFedPara, which separates parameters into global and local ones. We show that pFedPara outperforms competing personalized FL methods with more than three times fewer parameters*.
+
+## LoHaConfig
+
+[[autodoc]] tuners.loha.config.LoHaConfig
+
+## LoHaModel
+
+[[autodoc]] tuners.loha.model.LoHaModel

docs/source/package_reference/lokr.md

@@ -0,0 +1,27 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# LoKr
+
+Low-Rank Kronecker Product ([LoKr](https://hf.co/papers/2309.14859)), is a LoRA-variant method that approximates the large weight matrix with two low-rank matrices and combines them with the Kronecker product. LoKr also provides an optional third low-rank matrix to provide better control during fine-tuning.
+
+## LoKrConfig
+
+[[autodoc]] tuners.lokr.config.LoKrConfig
+
+## LoKrModel
+
+[[autodoc]] tuners.lokr.model.LoKrModel

docs/source/package_reference/lora.md

@@ -0,0 +1,35 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# LoRA
+
+Low-Rank Adaptation ([LoRA](https://huggingface.co/papers/2309.15223)) is a PEFT method that decomposes a large matrix into two smaller low-rank matrices in the attention layers. This drastically reduces the number of parameters that need to be fine-tuned.
+
+The abstract from the paper is:
+
+*We propose a neural language modeling system based on low-rank adaptation (LoRA) for speech recognition output rescoring. Although pretrained language models (LMs) like BERT have shown superior performance in second-pass rescoring, the high computational cost of scaling up the pretraining stage and adapting the pretrained models to specific domains limit their practical use in rescoring. Here we present a method based on low-rank decomposition to train a rescoring BERT model and adapt it to new domains using only a fraction (0.08%) of the pretrained parameters. These inserted matrices are optimized through a discriminative training objective along with a correlation-based regularization loss. The proposed low-rank adaptation Rescore-BERT (LoRB) architecture is evaluated on LibriSpeech and internal datasets with decreased training times by factors between 5.4 and 3.6.*.
+
+## LoraConfig
+
+[[autodoc]] tuners.lora.config.LoraConfig
+
+## LoraModel
+
+[[autodoc]] tuners.lora.model.LoraModel
+
+## Utility
+
+[[autodoc]] utils.loftq_utils.replace_lora_weights_loftq

docs/source/package_reference/merge_utils.md

@@ -0,0 +1,33 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Model merge
+
+PEFT provides several internal utilities for [merging LoRA adapters](../developer_guides/model_merging) with the TIES and DARE methods.
+
+[[autodoc]] utils.merge_utils.prune
+
+[[autodoc]] utils.merge_utils.calculate_majority_sign_mask
+
+[[autodoc]] utils.merge_utils.disjoint_merge
+
+[[autodoc]] utils.merge_utils.task_arithmetic
+
+[[autodoc]] utils.merge_utils.ties
+
+[[autodoc]] utils.merge_utils.dare_linear
+
+[[autodoc]] utils.merge_utils.dare_ties

docs/source/package_reference/multitask_prompt_tuning.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Multitask prompt tuning
+
+[Multitask prompt tuning](https://huggingface.co/papers/2303.02861)  decomposes the soft prompts of each task into a single learned transferable prompt instead of a separate prompt for each task. The single learned prompt can be adapted for each task by multiplicative low rank updates.
+
+The abstract from the paper is:
+
+*Prompt tuning, in which a base pretrained model is adapted to each task via conditioning on learned prompt vectors, has emerged as a promising approach for efficiently adapting large language models to multiple downstream tasks. However, existing methods typically learn soft prompt vectors from scratch, and it has not been clear how to exploit the rich cross-task knowledge with prompt vectors in a multitask learning setting. We propose multitask prompt tuning (MPT), which first learns a single transferable prompt by distilling knowledge from multiple task-specific source prompts. We then learn multiplicative low rank updates to this shared prompt to efficiently adapt it to each downstream target task. Extensive experiments on 23 NLP datasets demonstrate that our proposed approach outperforms the state-of-the-art methods, including the full finetuning baseline in some cases, despite only tuning 0.035% as many task-specific parameters*.
+
+## MultitaskPromptTuningConfig
+
+[[autodoc]] tuners.multitask_prompt_tuning.config.MultitaskPromptTuningConfig
+
+## MultitaskPromptEmbedding
+
+[[autodoc]] tuners.multitask_prompt_tuning.model.MultitaskPromptEmbedding

docs/source/package_reference/oft.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# OFT
+
+[Orthogonal Finetuning (OFT)](https://hf.co/papers/2306.07280) is a method developed for adapting text-to-image diffusion models. It works by reparameterizing the pretrained weight matrices with it's orthogonal matrix to preserve information in the pretrained model. To reduce the number of parameters, OFT introduces a block-diagonal structure in the orthogonal matrix.
+
+The abstract from the paper is:
+
+*Large text-to-image diffusion models have impressive capabilities in generating photorealistic images from text prompts. How to effectively guide or control these powerful models to perform different downstream tasks becomes an important open problem. To tackle this challenge, we introduce a principled finetuning method -- Orthogonal Finetuning (OFT), for adapting text-to-image diffusion models to downstream tasks. Unlike existing methods, OFT can provably preserve hyperspherical energy which characterizes the pairwise neuron relationship on the unit hypersphere. We find that this property is crucial for preserving the semantic generation ability of text-to-image diffusion models. To improve finetuning stability, we further propose Constrained Orthogonal Finetuning (COFT) which imposes an additional radius constraint to the hypersphere. Specifically, we consider two important finetuning text-to-image tasks: subject-driven generation where the goal is to generate subject-specific images given a few images of a subject and a text prompt, and controllable generation where the goal is to enable the model to take in additional control signals. We empirically show that our OFT framework outperforms existing methods in generation quality and convergence speed*.
+
+## OFTConfig
+
+[[autodoc]] tuners.oft.config.OFTConfig
+
+## OFTModel
+
+[[autodoc]] tuners.oft.model.OFTModel

docs/source/package_reference/p_tuning.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# P-tuning
+
+[P-tuning](https://hf.co/papers/2103.10385) adds trainable prompt embeddings to the input that is optimized by a prompt encoder to find a better prompt, eliminating the need to manually design prompts. The prompt tokens can be added anywhere in the input sequence, and p-tuning also introduces anchor tokens for improving performance.
+
+The abstract from the paper is:
+
+*While GPTs with traditional fine-tuning fail to achieve strong results on natural language understanding (NLU), we show that GPTs can be better than or comparable to similar-sized BERTs on NLU tasks with a novel method P-tuning -- which employs trainable continuous prompt embeddings. On the knowledge probing (LAMA) benchmark, the best GPT recovers 64\% (P@1) of world knowledge without any additional text provided during test time, which substantially improves the previous best by 20+ percentage points. On the SuperGlue benchmark, GPTs achieve comparable and sometimes better performance to similar-sized BERTs in supervised learning. Importantly, we find that P-tuning also improves BERTs' performance in both few-shot and supervised settings while largely reducing the need for prompt engineering. Consequently, P-tuning outperforms the state-of-the-art approaches on the few-shot SuperGlue benchmark.*.
+
+## PromptEncoderConfig
+
+[[autodoc]] tuners.p_tuning.config.PromptEncoderConfig
+
+## PromptEncoder
+
+[[autodoc]] tuners.p_tuning.model.PromptEncoder

docs/source/package_reference/peft_model.md

@@ -0,0 +1,77 @@
+<!--⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# Models
+
+[`PeftModel`] is the base model class for specifying the base Transformer model and configuration to apply a PEFT method to. The base `PeftModel` contains methods for loading and saving models from the Hub.
+
+## PeftModel
+
+[[autodoc]] PeftModel
+    - all
+
+## PeftModelForSequenceClassification
+
+A `PeftModel` for sequence classification tasks.
+
+[[autodoc]] PeftModelForSequenceClassification
+    - all
+
+## PeftModelForTokenClassification
+
+A `PeftModel` for token classification tasks.
+
+[[autodoc]] PeftModelForTokenClassification
+    - all
+
+## PeftModelForCausalLM
+
+A `PeftModel` for causal language modeling.
+
+[[autodoc]] PeftModelForCausalLM
+    - all
+
+## PeftModelForSeq2SeqLM
+
+A `PeftModel` for sequence-to-sequence language modeling.
+
+[[autodoc]] PeftModelForSeq2SeqLM
+    - all
+
+## PeftModelForQuestionAnswering
+
+A `PeftModel` for question answering.
+
+[[autodoc]] PeftModelForQuestionAnswering
+    - all
+
+## PeftModelForFeatureExtraction
+
+A `PeftModel` for getting extracting features/embeddings from transformer models.
+
+[[autodoc]] PeftModelForFeatureExtraction
+    - all
+
+## PeftMixedModel
+
+A `PeftModel` for mixing different adapter types (e.g. LoRA and LoHa).
+
+[[autodoc]] PeftMixedModel
+    - all
+
+## Utilities
+
+[[autodoc]] utils.cast_mixed_precision_params
+
+[[autodoc]] get_peft_model
+
+[[autodoc]] inject_adapter_in_model
+
+[[autodoc]] utils.get_peft_model_state_dict
+
+[[autodoc]] utils.prepare_model_for_kbit_training
+
+[[autodoc]] get_layer_status
+
+[[autodoc]] get_model_status

docs/source/package_reference/peft_types.md

@@ -0,0 +1,27 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# PEFT types
+
+[`PeftType`] includes the supported adapters in PEFT, and [`TaskType`] includes PEFT-supported tasks.
+
+## PeftType
+
+[[autodoc]] utils.peft_types.PeftType
+
+## TaskType
+
+[[autodoc]] utils.peft_types.TaskType

docs/source/package_reference/poly.md

@@ -0,0 +1,44 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Polytropon
+
+[Polytropon](https://hf.co/papers/2202.13914) is a multitask model with a number of different LoRA adapters in it's "inventory". The model learns the correct combination of adapters from the inventory with a routing function to choose the best subset of modules for a specific task. PEFT also supports [Multi-Head Adapter Routing (MHR)](https://hf.co/papers/2211.03831) for Polytropon which builds on and improves the routing function by combining the adapter heads more granularly. The adapter heads are separated into disjoint blocks and a different routing function is learned for each one, allowing for more expressivity.
+
+<hfoptions id="paper">
+<hfoption id="Combining Modular Skills in Multitask Learning">
+
+The abstract from the paper is:
+
+*A modular design encourages neural models to disentangle and recombine different facets of knowledge to generalise more systematically to new tasks. In this work, we assume that each task is associated with a subset of latent discrete skills from a (potentially small) inventory. In turn, skills correspond to parameter-efficient (sparse / low-rank) model parameterisations. By jointly learning these and a task-skill allocation matrix, the network for each task is instantiated as the average of the parameters of active skills. To favour non-trivial soft partitions of skills across tasks, we experiment with a series of inductive biases, such as an Indian Buffet Process prior and a two-speed learning rate. We evaluate our latent-skill model on two main settings: 1) multitask reinforcement learning for grounded instruction following on 8 levels of the BabyAI platform; and 2) few-shot adaptation of pre-trained text-to-text generative models on CrossFit, a benchmark comprising 160 NLP tasks. We find that the modular design of a network significantly increases sample efficiency in reinforcement learning and few-shot generalisation in supervised learning, compared to baselines with fully shared, task-specific, or conditionally generated parameters where knowledge is entangled across tasks. In addition, we show how discrete skills help interpretability, as they yield an explicit hierarchy of tasks.*
+
+</hfoption>
+<hfoption id="Multi-Head Adapter Routing for Cross-Task Generalization">
+
+The abstract from the paper is:
+
+*Parameter-efficient fine-tuning (PEFT) for cross-task generalization consists in pre-training adapters on a multi-task training set before few-shot adaptation to test tasks. Polytropon [Ponti et al., 2023] (Poly) jointly learns an inventory of adapters and a routing function that selects a (variable-size) subset of adapters for each task during both pre-training and few-shot adaptation. In this paper, we investigate the role that adapter routing plays in its success and design new variants based on our findings. First, we build on the intuition that finer-grained routing provides more expressivity. Hence, we propose MHR (Multi-Head Routing), which combines subsets of adapter parameters and outperforms Poly under a comparable parameter budget; by only fine-tuning the routing function and not the adapters (MHR-z), we achieve competitive performance with extreme parameter efficiency. Second, we find that Poly/MHR performance is a result of better multi-task optimization, rather than modular inductive biases that facilitate adapter recombination and local adaptation, as previously hypothesized. In fact, we find that MHR exhibits higher gradient alignment between tasks than any other method. Since this implies that routing is only crucial during multi-task pre-training, we propose MHR-mu, which discards routing and fine-tunes the average of the pre-trained adapters during few-shot adaptation. This establishes MHR-mu as an effective method for single-adapter fine-tuning.*.
+
+</hfoption>
+</hfoptions>
+
+## PolyConfig
+
+[[autodoc]] tuners.poly.config.PolyConfig
+
+## PolyModel
+
+[[autodoc]] tuners.poly.model.PolyModel

docs/source/package_reference/prefix_tuning.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Prefix tuning
+
+[Prefix tuning](https://hf.co/papers/2101.00190) prefixes a series of task-specific vectors to the input sequence that can be learned while keeping the pretrained model frozen. The prefix parameters are inserted in all of the model layers.
+
+The abstract from the paper is:
+
+*Fine-tuning is the de facto way to leverage large pretrained language models to perform downstream tasks. However, it modifies all the language model parameters and therefore necessitates storing a full copy for each task. In this paper, we propose prefix-tuning, a lightweight alternative to fine-tuning for natural language generation tasks, which keeps language model parameters frozen, but optimizes a small continuous task-specific vector (called the prefix). Prefix-tuning draws inspiration from prompting, allowing subsequent tokens to attend to this prefix as if it were "virtual tokens". We apply prefix-tuning to GPT-2 for table-to-text generation and to BART for summarization. We find that by learning only 0.1\% of the parameters, prefix-tuning obtains comparable performance in the full data setting, outperforms fine-tuning in low-data settings, and extrapolates better to examples with topics unseen during training*.
+
+## PrefixTuningConfig
+
+[[autodoc]] tuners.prefix_tuning.config.PrefixTuningConfig
+
+## PrefixEncoder
+
+[[autodoc]] tuners.prefix_tuning.model.PrefixEncoder

docs/source/package_reference/prompt_tuning.md

@@ -0,0 +1,31 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Prompt tuning
+
+[Prompt tuning](https://hf.co/papers/2104.08691) adds task-specific prompts to the input, and these prompt parameters are updated independently of the pretrained model parameters which are frozen.
+
+The abstract from the paper is:
+
+*In this work, we explore "prompt tuning", a simple yet effective mechanism for learning "soft prompts" to condition frozen language models to perform specific downstream tasks. Unlike the discrete text prompts used by GPT-3, soft prompts are learned through backpropagation and can be tuned to incorporate signal from any number of labeled examples. Our end-to-end learned approach outperforms GPT-3's "few-shot" learning by a large margin. More remarkably, through ablations on model size using T5, we show that prompt tuning becomes more competitive with scale: as models exceed billions of parameters, our method "closes the gap" and matches the strong performance of model tuning (where all model weights are tuned). This finding is especially relevant in that large models are costly to share and serve, and the ability to reuse one frozen model for multiple downstream tasks can ease this burden. Our method can be seen as a simplification of the recently proposed "prefix tuning" of Li and Liang (2021), and we provide a comparison to this and other similar approaches. Finally, we show that conditioning a frozen model with soft prompts confers benefits in robustness to domain transfer, as compared to full model tuning*.
+
+## PromptTuningConfig
+
+[[autodoc]] tuners.prompt_tuning.config.PromptTuningConfig
+
+## PromptEmbedding
+
+[[autodoc]] tuners.prompt_tuning.model.PromptEmbedding

docs/source/package_reference/tuners.md

@@ -0,0 +1,27 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Tuners
+
+A tuner (or adapter) is a module that can be plugged into a `torch.nn.Module`. [`BaseTuner`] base class for other tuners and provides shared methods and attributes for preparing an adapter configuration and replacing a target module with the adapter module. [`BaseTunerLayer`] is a base class for adapter layers. It offers methods and attributes for managing adapters such as activating and disabling adapters.
+
+## BaseTuner
+
+[[autodoc]] tuners.tuners_utils.BaseTuner
+
+## BaseTunerLayer
+
+[[autodoc]] tuners.tuners_utils.BaseTunerLayer

docs/source/package_reference/vblora.md

@@ -0,0 +1,40 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks
+
+## Overview
+
+[VB-LoRA](https://arxiv.org/abs/2405.15179) is a parameter-efficient fine-tuning technique that extends LoRA by learning a fine-grained parameter-sharing scheme at the sub-vector level, achieving significantly higher parameter efficiency. This makes VB-LoRA especially useful in scenarios where storage and transmission costs are critical. It works by decomposing low-rank matrices—from different layers and modules such as K, Q, V, and FFN—into sub-vectors, which are then globally shared through a vector bank.
+
+The abstract from the paper is:
+
+*As the adoption of large language models increases and the need for per-user or per-task model customization grows, the parameter-efficient fine-tuning (PEFT) methods, such as low-rank adaptation (LoRA) and its variants, incur substantial storage and transmission costs. To further reduce stored parameters, we introduce a "divide-and-share" paradigm that breaks the barriers of low-rank decomposition across matrix dimensions, modules and layers by sharing parameters globally via a vector bank. As an instantiation of the paradigm to LoRA, our proposed VB-LoRA composites all the low-rank matrices of LoRA from a shared vector bank with a differentiable top-k admixture module. VB-LoRA achieves extreme parameter efficiency while maintaining comparable or better performance compared to state-of-the-art PEFT methods. Extensive experiments demonstrate the effectiveness of VB-LoRA on natural language understanding, natural language generation, and instruction tuning tasks. When fine-tuning the Llama2-13B model, VB-LoRA only uses 0.4% of LoRA's stored parameters, yet achieves superior results.*
+
+## Usage Tips
+
+- VB-LoRA utilizes a sparse top-k module to learn the sharing machanism. When saving adapter parameters, you can either save only the top-k weights and their indices by setting `save_only_topk_weights = True` in `VBLoRAConfig`, or save all the trainable logits by setting it to `False`. Enabling `save_only_topk_weights = True` significantly reduces storage space; for instance, in Llama2-7B, the storage file size decreases from 308MB to 2.5MB. Note that models saved with `save_only_topk_weights = True` are intended for merging or inference only and cannot be used to resume training.
+
+- VB-LoRA has two sets of training parameters: vector bank parameters and logit parameters. In practice, we found that logit parameters require a higher learning rate, while vector bank parameters require a lower learning rate. When using the AdamW optimizer, typical learning rates are 0.01 for logits and 0.001 for vector bank parameters.
+
+## VBLoRAConfig
+
+[[autodoc]] tuners.vblora.config.VBLoRAConfig
+
+## VBLoRAModel
+
+[[autodoc]] tuners.vblora.model.VBLoRAModel
+

docs/source/package_reference/vera.md

@@ -0,0 +1,42 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# VeRA: Vector-based Random Matrix Adaptation
+
+[VeRA](https://huggingface.co/papers/2310.11454) is a parameter-efficient fine-tuning technique that is similar to LoRA but requires even fewer extra parameters while promising similar or even better performance. As such, it is particularly useful when the parameter budget is very limited, e.g. when scaling to very large models. The reduction of the count of trainable parameters is achieved by sharing the same low-rank matrices across all layers, and only training two additional vectors per layer.
+
+When saving the adapter parameters, it's possible to eschew storing the low rank matrices by setting `save_projection=False` on the `VeraConfig`. In that case, these matrices will be restored based on the fixed random seed from the `projection_prng_key` argument. This cuts down on the size of the checkpoint, but we cannot guarantee reproducibility on all devices and for all future versions of PyTorch. If you want to ensure reproducibility, set `save_projection=True` (which is the default).
+
+To handle different shapes of adapted layers, VeRA initializes shared A and B matrices with the largest required size for each dimension. During the forward pass, submatrices A and B for a given layer are sliced out from these shared matrices and used as described in the paper. For example, adapting two linear layers of shapes (100, 20) and (80, 50) will create A and B matrices of shapes (rank, 50) and (100, rank) respectively. Then, to adapt a layer of shape (100, 20), submatrices A and B of shapes (rank, 20) and (100, rank) will be extracted.
+
+VeRA currently has the following constraints:
+
+- Only `nn.Linear` layers are supported.
+- Quantized layers are not supported.
+
+If these constraints don't work for your use case, use LoRA instead.
+
+The abstract from the paper is:
+
+> Low-rank adapation (LoRA) is a popular method that reduces the number of trainable parameters when finetuning large language models, but still faces acute storage challenges when scaling to even larger models or deploying numerous per-user or per-task adapted models. In this work, we present Vector-based Random Matrix Adaptation (VeRA), which significantly reduces the number of trainable parameters compared to LoRA, yet maintains the same performance. It achieves this by using a single pair of low-rank matrices shared across all layers and learning small scaling vectors instead. We demonstrate its effectiveness on the GLUE and E2E benchmarks, image classification tasks, and show its application in instruction-tuning of 7B and 13B language models.
+
+## VeRAConfig
+
+[[autodoc]] tuners.vera.config.VeraConfig
+
+## VeRAModel
+
+[[autodoc]] tuners.vera.model.VeraModel

docs/source/package_reference/xlora.md

@@ -0,0 +1,56 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# X-LoRA
+
+Mixture of LoRA Experts ([X-LoRA](https://arxiv.org/abs/2402.07148)) is a PEFT method enabling sparse or dense mixture of LoRA experts based on a high granularity (token, layer, sequence) scalings matrix. This leverages frozen LoRA adapters and a frozen base model to drastically reduces the number of parameters that need to be fine-tuned.
+
+A unique aspect of X-LoRA is its versatility: it can be applied to any `transformers` base model with LoRA adapters. This means that, despite the mixture of experts strategy, no changes to the model code must be made.
+
+The below graphic demonstrates how the scalings change for different prompts for each token. This highlights the activation of different adapters as the generation progresses and the sequence creates new context.
+
+![Token-by-token scalings](https://github.com/EricLBuehler/xlora/raw/master/res/token_by_token_scalings.gif)
+
+The abstract from the paper is:
+
+*We report a mixture of expert strategy to create fine-tuned large language models using a deep layer-wise token-level approach based on low-rank adaptation (LoRA). Starting with a set of pre-trained LoRA adapters, our gating strategy uses the hidden states to dynamically mix adapted layers, allowing the resulting X-LoRA model to draw upon different capabilities and create never-before-used deep layer-wise combinations to solve tasks. The design is inspired by the biological principles of universality and diversity, where neural network building blocks are reused in different hierarchical manifestations. Hence, the X-LoRA model can be easily implemented for any existing large language model (LLM) without a need for modifications of the underlying structure. We develop a tailored X-LoRA model that offers scientific capabilities including forward/inverse analysis tasks and enhanced reasoning capability, focused on biomaterial analysis, protein mechanics and design. The impact of this work include access to readily expandable and adaptable models with strong domain knowledge and the capability to integrate across areas of knowledge. Featuring experts in biology, mathematics, reasoning, bio-inspired materials, mechanics and materials, chemistry, protein biophysics, mechanics and quantum-mechanics based molecular properties, we conduct a series of physics-focused case studies. We examine knowledge recall, protein mechanics forward/inverse tasks, protein design, adversarial agentic modeling including ontological knowledge graph construction, as well as molecular design. The model is capable not only of making quantitative predictions of nanomechanical properties of proteins or quantum mechanical molecular properties, but also reasons over the results and correctly predicts likely mechanisms that explain distinct molecular behaviors.*.
+
+Please cite X-LoRA as:
+```bibtex
+@article{10.1063/5.0203126,
+    author = {Buehler, Eric L. and Buehler, Markus J.},
+    title = "{X-LoRA: Mixture of low-rank adapter experts, a flexible framework for large language models with applications in protein mechanics and molecular design}",
+    journal = {APL Machine Learning},
+    volume = {2},
+    number = {2},
+    pages = {026119},
+    year = {2024},
+    month = {05},
+    abstract = "{We report a mixture of expert strategy to create fine-tuned large language models using a deep layer-wise token-level approach based on low-rank adaptation (LoRA). Starting with a set of pre-trained LoRA adapters, our gating strategy uses the hidden states to dynamically mix adapted layers, allowing the resulting X-LoRA model to draw upon different capabilities and create never-before-used deep layer-wise combinations to solve tasks. The design is inspired by the biological principles of universality and diversity, where neural network building blocks are reused in different hierarchical manifestations. Hence, the X-LoRA model can be easily implemented for any existing large language model without a need for modifications of the underlying structure. We develop a tailored X-LoRA model that offers scientific capabilities, including forward/inverse analysis tasks and enhanced reasoning capability, focused on biomaterial analysis, protein mechanics, and design. The impact of this work includes access to readily expandable and adaptable models with strong domain knowledge and the capability to integrate across areas of knowledge. Featuring experts in biology, mathematics, reasoning, bio-inspired materials, mechanics and materials, chemistry, protein biophysics, mechanics, and quantum-mechanics based molecular properties, we conduct a series of physics-focused case studies. We examine knowledge recall, protein mechanics forward/inverse tasks, protein design, adversarial agentic modeling including ontological knowledge graph construction, and molecular design. The model is capable not only of making quantitative predictions of nanomechanical properties of proteins or quantum mechanical molecular properties but also reasoning over the results and correctly predicting likely mechanisms that explain distinct molecular behaviors.}",
+    issn = {2770-9019},
+    doi = {10.1063/5.0203126},
+    url = {https://doi.org/10.1063/5.0203126},
+    eprint = {https://pubs.aip.org/aip/aml/article-pdf/doi/10.1063/5.0203126/19964043/026119\_1\_5.0203126.pdf},
+}
+```
+
+## XLoraConfig
+
+[[autodoc]] tuners.xlora.config.XLoraConfig
+
+## XLoraModel
+
+[[autodoc]] tuners.xlora.model.XLoraModel

docs/source/quicktour.md

@@ -0,0 +1,170 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Quicktour
+
+PEFT offers parameter-efficient methods for finetuning large pretrained models. The traditional paradigm is to finetune all of a model's parameters for each downstream task, but this is becoming exceedingly costly and impractical because of the enormous number of parameters in models today. Instead, it is more efficient to train a smaller number of prompt parameters or use a reparametrization method like low-rank adaptation (LoRA) to reduce the number of trainable parameters.
+
+This quicktour will show you PEFT's main features and how you can train or run inference on large models that would typically be inaccessible on consumer devices.
+
+## Train
+
+Each PEFT method is defined by a [`PeftConfig`] class that stores all the important parameters for building a [`PeftModel`]. For example, to train with LoRA, load and create a [`LoraConfig`] class and specify the following parameters:
+
+- `task_type`: the task to train for (sequence-to-sequence language modeling in this case)
+- `inference_mode`: whether you're using the model for inference or not
+- `r`: the dimension of the low-rank matrices
+- `lora_alpha`: the scaling factor for the low-rank matrices
+- `lora_dropout`: the dropout probability of the LoRA layers
+
+```python
+from peft import LoraConfig, TaskType
+
+peft_config = LoraConfig(task_type=TaskType.SEQ_2_SEQ_LM, inference_mode=False, r=8, lora_alpha=32, lora_dropout=0.1)
+```
+
+<Tip>
+
+See the [`LoraConfig`] reference for more details about other parameters you can adjust, such as the modules to target or the bias type.
+
+</Tip>
+
+Once the [`LoraConfig`] is setup, create a [`PeftModel`] with the [`get_peft_model`] function. It takes a base model - which you can load from the Transformers library - and the [`LoraConfig`] containing the parameters for how to configure a model for training with LoRA.
+
+Load the base model you want to finetune.
+
+```python
+from transformers import AutoModelForSeq2SeqLM
+
+model = AutoModelForSeq2SeqLM.from_pretrained("bigscience/mt0-large")
+```
+
+Wrap the base model and `peft_config` with the [`get_peft_model`] function to create a [`PeftModel`]. To get a sense of the number of trainable parameters in your model, use the [`print_trainable_parameters`] method.
+
+```python
+from peft import get_peft_model
+
+model = get_peft_model(model, peft_config)
+model.print_trainable_parameters()
+"output: trainable params: 2359296 || all params: 1231940608 || trainable%: 0.19151053100118282"
+```
+
+Out of [bigscience/mt0-large's](https://huggingface.co/bigscience/mt0-large) 1.2B parameters, you're only training 0.19% of them!
+
+That is it 🎉! Now you can train the model with the Transformers [`~transformers.Trainer`], Accelerate, or any custom PyTorch training loop.
+
+For example, to train with the [`~transformers.Trainer`] class, setup a [`~transformers.TrainingArguments`] class with some training hyperparameters.
+
+```py
+training_args = TrainingArguments(
+    output_dir="your-name/bigscience/mt0-large-lora",
+    learning_rate=1e-3,
+    per_device_train_batch_size=32,
+    per_device_eval_batch_size=32,
+    num_train_epochs=2,
+    weight_decay=0.01,
+    eval_strategy="epoch",
+    save_strategy="epoch",
+    load_best_model_at_end=True,
+)
+```
+
+Pass the model, training arguments, dataset, tokenizer, and any other necessary component to the [`~transformers.Trainer`], and call [`~transformers.Trainer.train`] to start training.
+
+```py
+trainer = Trainer(
+    model=model,
+    args=training_args,
+    train_dataset=tokenized_datasets["train"],
+    eval_dataset=tokenized_datasets["test"],
+    tokenizer=tokenizer,
+    data_collator=data_collator,
+    compute_metrics=compute_metrics,
+)
+
+trainer.train()
+```
+
+### Save model
+
+After your model is finished training, you can save your model to a directory using the [`~transformers.PreTrainedModel.save_pretrained`] function.
+
+```py
+model.save_pretrained("output_dir")
+```
+
+You can also save your model to the Hub (make sure you're logged in to your Hugging Face account first) with the [`~transformers.PreTrainedModel.push_to_hub`] function.
+
+```python
+from huggingface_hub import notebook_login
+
+notebook_login()
+model.push_to_hub("your-name/bigscience/mt0-large-lora")
+```
+
+Both methods only save the extra PEFT weights that were trained, meaning it is super efficient to store, transfer, and load. For example, this [facebook/opt-350m](https://huggingface.co/ybelkada/opt-350m-lora) model trained with LoRA only contains two files: `adapter_config.json` and `adapter_model.safetensors`. The `adapter_model.safetensors` file is just 6.3MB!
+
+<div class="flex flex-col justify-center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/PEFT-hub-screenshot.png"/>
+  <figcaption class="text-center">The adapter weights for a opt-350m model stored on the Hub are only ~6MB compared to the full size of the model weights, which can be ~700MB.</figcaption>
+</div>
+
+## Inference
+
+<Tip>
+
+Take a look at the [AutoPeftModel](package_reference/auto_class) API reference for a complete list of available `AutoPeftModel` classes.
+
+</Tip>
+
+Easily load any PEFT-trained model for inference with the [`AutoPeftModel`] class and the [`~transformers.PreTrainedModel.from_pretrained`] method:
+
+```py
+from peft import AutoPeftModelForCausalLM
+from transformers import AutoTokenizer
+import torch
+
+model = AutoPeftModelForCausalLM.from_pretrained("ybelkada/opt-350m-lora")
+tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
+
+model = model.to("cuda")
+model.eval()
+inputs = tokenizer("Preheat the oven to 350 degrees and place the cookie dough", return_tensors="pt")
+
+outputs = model.generate(input_ids=inputs["input_ids"].to("cuda"), max_new_tokens=50)
+print(tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True)[0])
+
+"Preheat the oven to 350 degrees and place the cookie dough in the center of the oven. In a large bowl, combine the flour, baking powder, baking soda, salt, and cinnamon. In a separate bowl, combine the egg yolks, sugar, and vanilla."
+```
+
+For other tasks that aren't explicitly supported with an `AutoPeftModelFor` class - such as automatic speech recognition - you can still use the base [`AutoPeftModel`] class to load a model for the task.
+
+```py
+from peft import AutoPeftModel
+
+model = AutoPeftModel.from_pretrained("smangrul/openai-whisper-large-v2-LORA-colab")
+```
+
+## Next steps
+
+Now that you've seen how to train a model with one of the PEFT methods, we encourage you to try out some of the other methods like prompt tuning. The steps are very similar to the ones shown in the quicktour:
+
+1. prepare a [`PeftConfig`] for a PEFT method
+2. use the [`get_peft_model`] method to create a [`PeftModel`] from the configuration and base model
+
+Then you can train it however you like! To load a PEFT model for inference, you can use the [`AutoPeftModel`] class.
+
+Feel free to also take a look at the task guides if you're interested in training a model with another PEFT method for a specific task such as semantic segmentation, multilingual automatic speech recognition, DreamBooth, token classification, and more.

docs/source/task_guides/ia3.md

@@ -0,0 +1,239 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# IA3
+
+[IA3](../conceptual_guides/ia3) multiplies the model's activations (the keys and values in the self-attention and encoder-decoder attention blocks, and the intermediate activation of the position-wise feedforward network) by three learned vectors. This PEFT method introduces an even smaller number of trainable parameters than LoRA which introduces weight matrices instead of vectors. The original model's parameters are kept frozen and only these vectors are updated. As a result, it is faster, cheaper and more efficient to finetune for a new downstream task.
+
+This guide will show you how to train a sequence-to-sequence model with IA3 to *generate a sentiment* given some financial news.
+
+<Tip>
+
+Some familiarity with the general process of training a sequence-to-sequence would be really helpful and allow you to focus on how to apply IA3. If you’re new, we recommend taking a look at the [Translation](https://huggingface.co/docs/transformers/tasks/translation) and [Summarization](https://huggingface.co/docs/transformers/tasks/summarization) guides first from the Transformers documentation. When you’re ready, come back and see how easy it is to drop PEFT in to your training!
+
+</Tip>
+
+## Dataset
+
+You'll use the sentences_allagree subset of the [financial_phrasebank](https://huggingface.co/datasets/financial_phrasebank) dataset. This subset contains financial news with 100% annotator agreement on the sentiment label. Take a look at the [dataset viewer](https://huggingface.co/datasets/financial_phrasebank/viewer/sentences_allagree) for a better idea of the data and sentences you'll be working with.
+
+Load the dataset with the [`~datasets.load_dataset`] function. This subset of the dataset only contains a train split, so use the [`~datasets.train_test_split`] function to create a train and validation split. Create a new `text_label` column so it is easier to understand what the `label` values `0`, `1`, and `2` mean.
+
+```py
+from datasets import load_dataset
+
+ds = load_dataset("financial_phrasebank", "sentences_allagree")
+ds = ds["train"].train_test_split(test_size=0.1)
+ds["validation"] = ds["test"]
+del ds["test"]
+
+classes = ds["train"].features["label"].names
+ds = ds.map(
+    lambda x: {"text_label": [classes[label] for label in x["label"]]},
+    batched=True,
+    num_proc=1,
+)
+
+ds["train"][0]
+{'sentence': 'It will be operated by Nokia , and supported by its Nokia NetAct network and service management system .',
+ 'label': 1,
+ 'text_label': 'neutral'}
+```
+
+Load a tokenizer and create a preprocessing function that:
+
+1. tokenizes the inputs, pads and truncates the sequence to the `max_length`
+2. apply the same tokenizer to the labels but with a shorter `max_length` that corresponds to the label
+3. mask the padding tokens
+
+```py
+from transformers import AutoTokenizer
+
+text_column = "sentence"
+label_column = "text_label"
+max_length = 128
+
+tokenizer = AutoTokenizer.from_pretrained("bigscience/mt0-large")
+
+def preprocess_function(examples):
+    inputs = examples[text_column]
+    targets = examples[label_column]
+    model_inputs = tokenizer(inputs, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt")
+    labels = tokenizer(targets, max_length=3, padding="max_length", truncation=True, return_tensors="pt")
+    labels = labels["input_ids"]
+    labels[labels == tokenizer.pad_token_id] = -100
+    model_inputs["labels"] = labels
+    return model_inputs
+```
+
+Use the [`~datasets.Dataset.map`] function to apply the preprocessing function to the entire dataset.
+
+```py
+processed_ds = ds.map(
+    preprocess_function,
+    batched=True,
+    num_proc=1,
+    remove_columns=ds["train"].column_names,
+    load_from_cache_file=False,
+    desc="Running tokenizer on dataset",
+)
+```
+
+Create a training and evaluation [`DataLoader`](https://pytorch.org/docs/stable/data.html#torch.utils.data.DataLoader), and set `pin_memory=True` to speed up data transfer to the GPU during training if your dataset samples are on a CPU.
+
+```py
+from torch.utils.data import DataLoader
+from transformers import default_data_collator
+
+train_ds = processed_ds["train"]
+eval_ds = processed_ds["validation"]
+
+batch_size = 8
+
+train_dataloader = DataLoader(
+    train_ds, shuffle=True, collate_fn=default_data_collator, batch_size=batch_size, pin_memory=True
+)
+eval_dataloader = DataLoader(eval_ds, collate_fn=default_data_collator, batch_size=batch_size, pin_memory=True)
+```
+
+## Model
+
+Now you can load a pretrained model to use as the base model for IA3. This guide uses the [bigscience/mt0-large](https://huggingface.co/bigscience/mt0-large) model, but you can use any sequence-to-sequence model you like.
+
+```py
+from transformers import AutoModelForSeq2SeqLM
+
+model = AutoModelForSeq2SeqLM.from_pretrained("bigscience/mt0-large")
+```
+
+### PEFT configuration and model
+
+All PEFT methods need a configuration that contains and specifies all the parameters for how the PEFT method should be applied. Create an [`IA3Config`] with the task type and set the inference mode to `False`. You can find additional parameters for this configuration in the [API reference](../package_reference/ia3#ia3config).
+
+<Tip>
+
+Call the [`~PeftModel.print_trainable_parameters`] method to compare the number of trainable parameters of [`PeftModel`] versus the number of parameters in the base model!
+
+</Tip>
+
+Once the configuration is setup, pass it to the [`get_peft_model`] function along with the base model to create a trainable [`PeftModel`].
+
+```py
+from peft import IA3Config, get_peft_model
+
+peft_config = IA3Config(task_type="SEQ_2_SEQ_LM")
+model = get_peft_model(model, peft_config)
+model.print_trainable_parameters()
+"trainable params: 282,624 || all params: 1,229,863,936 || trainable%: 0.022980103060766553"
+```
+
+### Training
+
+Set up an optimizer and learning rate scheduler.
+
+```py
+import torch
+from transformers import get_linear_schedule_with_warmup
+
+lr = 8e-3
+num_epochs = 3
+
+optimizer = torch.optim.AdamW(model.parameters(), lr=lr)
+lr_scheduler = get_linear_schedule_with_warmup(
+    optimizer=optimizer,
+    num_warmup_steps=0,
+    num_training_steps=(len(train_dataloader) * num_epochs),
+)
+```
+
+Move the model to the GPU and create a training loop that reports the loss and perplexity for each epoch.
+
+```py
+from tqdm import tqdm
+
+device = "cuda"
+model = model.to(device)
+
+for epoch in range(num_epochs):
+    model.train()
+    total_loss = 0
+    for step, batch in enumerate(tqdm(train_dataloader)):
+        batch = {k: v.to(device) for k, v in batch.items()}
+        outputs = model(**batch)
+        loss = outputs.loss
+        total_loss += loss.detach().float()
+        loss.backward()
+        optimizer.step()
+        lr_scheduler.step()
+        optimizer.zero_grad()
+
+    model.eval()
+    eval_loss = 0
+    eval_preds = []
+    for step, batch in enumerate(tqdm(eval_dataloader)):
+        batch = {k: v.to(device) for k, v in batch.items()}
+        with torch.no_grad():
+            outputs = model(**batch)
+        loss = outputs.loss
+        eval_loss += loss.detach().float()
+        eval_preds.extend(
+            tokenizer.batch_decode(torch.argmax(outputs.logits, -1).detach().cpu().numpy(), skip_special_tokens=True)
+        )
+
+    eval_epoch_loss = eval_loss / len(eval_dataloader)
+    eval_ppl = torch.exp(eval_epoch_loss)
+    train_epoch_loss = total_loss / len(train_dataloader)
+    train_ppl = torch.exp(train_epoch_loss)
+    print(f"{epoch=}: {train_ppl=} {train_epoch_loss=} {eval_ppl=} {eval_epoch_loss=}")
+```
+
+## Share your model
+
+After training is complete, you can upload your model to the Hub with the [`~transformers.PreTrainedModel.push_to_hub`] method. You'll need to login to your Hugging Face account first and enter your token when prompted.
+
+```py
+from huggingface_hub import notebook_login
+
+account = <your-hf-account-name>
+peft_model_id = f"{account}/mt0-large-ia3"
+model.push_to_hub(peft_model_id)
+```
+
+## Inference
+
+To load the model for inference, use the [`~AutoPeftModelForSeq2SeqLM.from_pretrained`] method. Let's also load a sentence of financial news from the dataset to generate a sentiment for.
+
+```py
+from peft import AutoPeftModelForSeq2SeqLM
+
+model = AutoPeftModelForSeq2SeqLM.from_pretrained("<your-hf-account-name>/mt0-large-ia3").to("cuda")
+tokenizer = AutoTokenizer.from_pretrained("bigscience/mt0-large")
+
+i = 15
+inputs = tokenizer(ds["validation"][text_column][i], return_tensors="pt")
+print(ds["validation"][text_column][i])
+"The robust growth was the result of the inclusion of clothing chain Lindex in the Group in December 2007 ."
+```
+
+Call the [`~transformers.GenerationMixin.generate`] method to generate the predicted sentiment label.
+
+```py
+with torch.no_grad():
+    inputs = {k: v.to(device) for k, v in inputs.items()}
+    outputs = model.generate(input_ids=inputs["input_ids"], max_new_tokens=10)
+    print(tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True))
+['positive']
+```

docs/source/task_guides/lora_based_methods.md

@@ -0,0 +1,350 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# LoRA methods
+
+A popular way to efficiently train large models is to insert (typically in the attention blocks) smaller trainable matrices that are a low-rank decomposition of the delta weight matrix to be learnt during finetuning. The pretrained model's original weight matrix is frozen and only the smaller matrices are updated during training. This reduces the number of trainable parameters, reducing memory usage and training time which can be very expensive for large models.
+
+There are several different ways to express the weight matrix as a low-rank decomposition, but [Low-Rank Adaptation (LoRA)](../conceptual_guides/adapter#low-rank-adaptation-lora) is the most common method. The PEFT library supports several other LoRA variants, such as [Low-Rank Hadamard Product (LoHa)](../conceptual_guides/adapter#low-rank-hadamard-product-loha), [Low-Rank Kronecker Product (LoKr)](../conceptual_guides/adapter#low-rank-kronecker-product-lokr), and [Adaptive Low-Rank Adaptation (AdaLoRA)](../conceptual_guides/adapter#adaptive-low-rank-adaptation-adalora). You can learn more about how these methods work conceptually in the [Adapters](../conceptual_guides/adapter) guide. If you're interested in applying these methods to other tasks and use cases like semantic segmentation, token classification, take a look at our [notebook collection](https://huggingface.co/collections/PEFT/notebooks-6573b28b33e5a4bf5b157fc1)!
+
+Additionally, PEFT supports the [X-LoRA](../conceptual_guides/adapter#mixture-of-lora-experts-x-lora) Mixture of LoRA Experts method.
+
+This guide will show you how to quickly train an image classification model - with a low-rank decomposition method - to identify the class of food shown in an image.
+
+<Tip>
+
+Some familiarity with the general process of training an image classification model would be really helpful and allow you to focus on the low-rank decomposition methods. If you're new, we recommend taking a look at the [Image classification](https://huggingface.co/docs/transformers/tasks/image_classification) guide first from the Transformers documentation. When you're ready, come back and see how easy it is to drop PEFT in to your training!
+
+</Tip>
+
+Before you begin, make sure you have all the necessary libraries installed.
+
+```bash
+pip install -q peft transformers datasets
+```
+
+## Dataset
+
+In this guide, you'll use the [Food-101](https://huggingface.co/datasets/food101) dataset which contains images of 101 food classes (take a look at the [dataset viewer](https://huggingface.co/datasets/food101/viewer/default/train) to get a better idea of what the dataset looks like).
+
+Load the dataset with the [`~datasets.load_dataset`] function.
+
+```py
+from datasets import load_dataset
+
+ds = load_dataset("food101")
+```
+
+Each food class is labeled with an integer, so to make it easier to understand what these integers represent, you'll create a `label2id` and `id2label` dictionary to map the integer to its class label.
+
+```py
+labels = ds["train"].features["label"].names
+label2id, id2label = dict(), dict()
+for i, label in enumerate(labels):
+    label2id[label] = i
+    id2label[i] = label
+
+id2label[2]
+"baklava"
+```
+
+Load an image processor to properly resize and normalize the pixel values of the training and evaluation images.
+
+```py
+from transformers import AutoImageProcessor
+
+image_processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224-in21k")
+```
+
+You can also use the image processor to prepare some transformation functions for data augmentation and pixel scaling.
+
+```py
+from torchvision.transforms import (
+    CenterCrop,
+    Compose,
+    Normalize,
+    RandomHorizontalFlip,
+    RandomResizedCrop,
+    Resize,
+    ToTensor,
+)
+
+normalize = Normalize(mean=image_processor.image_mean, std=image_processor.image_std)
+train_transforms = Compose(
+    [
+        RandomResizedCrop(image_processor.size["height"]),
+        RandomHorizontalFlip(),
+        ToTensor(),
+        normalize,
+    ]
+)
+
+val_transforms = Compose(
+    [
+        Resize(image_processor.size["height"]),
+        CenterCrop(image_processor.size["height"]),
+        ToTensor(),
+        normalize,
+    ]
+)
+
+def preprocess_train(example_batch):
+    example_batch["pixel_values"] = [train_transforms(image.convert("RGB")) for image in example_batch["image"]]
+    return example_batch
+
+def preprocess_val(example_batch):
+    example_batch["pixel_values"] = [val_transforms(image.convert("RGB")) for image in example_batch["image"]]
+    return example_batch
+```
+
+Define the training and validation datasets, and use the [`~datasets.Dataset.set_transform`] function to apply the transformations on-the-fly.
+
+```py
+train_ds = ds["train"]
+val_ds = ds["validation"]
+
+train_ds.set_transform(preprocess_train)
+val_ds.set_transform(preprocess_val)
+```
+
+Finally, you'll need a data collator to create a batch of training and evaluation data and convert the labels to `torch.tensor` objects.
+
+```py
+import torch
+
+def collate_fn(examples):
+    pixel_values = torch.stack([example["pixel_values"] for example in examples])
+    labels = torch.tensor([example["label"] for example in examples])
+    return {"pixel_values": pixel_values, "labels": labels}
+```
+
+## Model
+
+Now let's load a pretrained model to use as the base model. This guide uses the [google/vit-base-patch16-224-in21k](https://huggingface.co/google/vit-base-patch16-224-in21k) model, but you can use any image classification model you want. Pass the `label2id` and `id2label` dictionaries to the model so it knows how to map the integer labels to their class labels, and you can optionally pass the `ignore_mismatched_sizes=True` parameter if you're finetuning a checkpoint that has already been finetuned.
+
+```py
+from transformers import AutoModelForImageClassification, TrainingArguments, Trainer
+
+model = AutoModelForImageClassification.from_pretrained(
+    "google/vit-base-patch16-224-in21k",
+    label2id=label2id,
+    id2label=id2label,
+    ignore_mismatched_sizes=True,
+)
+```
+
+### PEFT configuration and model
+
+Every PEFT method requires a configuration that holds all the parameters specifying how the PEFT method should be applied. Once the configuration is setup, pass it to the [`~peft.get_peft_model`] function along with the base model to create a trainable [`PeftModel`].
+
+<Tip>
+
+Call the [`~PeftModel.print_trainable_parameters`] method to compare the number of parameters of [`PeftModel`] versus the number of parameters in the base model!
+
+</Tip>
+
+<hfoptions id="loras">
+<hfoption id="LoRA">
+
+[LoRA](../conceptual_guides/adapter#low-rank-adaptation-lora) decomposes the weight update matrix into *two* smaller matrices. The size of these low-rank matrices is determined by its *rank* or `r`. A higher rank means the model has more parameters to train, but it also means the model has more learning capacity. You'll also want to specify the `target_modules` which determine where the smaller matrices are inserted. For this guide, you'll target the *query* and *value* matrices of the attention blocks. Other important parameters to set are `lora_alpha` (scaling factor), `bias` (whether `none`, `all` or only the LoRA bias parameters should be trained), and `modules_to_save` (the modules apart from the LoRA layers to be trained and saved). All of these parameters - and more - are found in the [`LoraConfig`].
+
+```py
+from peft import LoraConfig, get_peft_model
+
+config = LoraConfig(
+    r=16,
+    lora_alpha=16,
+    target_modules=["query", "value"],
+    lora_dropout=0.1,
+    bias="none",
+    modules_to_save=["classifier"],
+)
+model = get_peft_model(model, config)
+model.print_trainable_parameters()
+"trainable params: 667,493 || all params: 86,543,818 || trainable%: 0.7712775047664294"
+```
+
+</hfoption>
+<hfoption id="LoHa">
+
+[LoHa](../conceptual_guides/adapter#low-rank-hadamard-product-loha) decomposes the weight update matrix into *four* smaller matrices and each pair of smaller matrices is combined with the Hadamard product. This allows the weight update matrix to keep the same number of trainable parameters when compared to LoRA, but with a higher rank (`r^2` for LoHA when compared to `2*r` for LoRA). The size of the smaller matrices is determined by its *rank* or `r`. You'll also want to specify the `target_modules` which determines where the smaller matrices are inserted. For this guide, you'll target the *query* and *value* matrices of the attention blocks. Other important parameters to set are `alpha` (scaling factor), and `modules_to_save` (the modules apart from the LoHa layers to be trained and saved). All of these parameters - and more - are found in the [`LoHaConfig`].
+
+```py
+from peft import LoHaConfig, get_peft_model
+
+config = LoHaConfig(
+    r=16,
+    alpha=16,
+    target_modules=["query", "value"],
+    module_dropout=0.1,
+    modules_to_save=["classifier"],
+)
+model = get_peft_model(model, config)
+model.print_trainable_parameters()
+"trainable params: 1,257,317 || all params: 87,133,642 || trainable%: 1.4429753779831676"
+```
+
+</hfoption>
+<hfoption id="LoKr">
+
+[LoKr](../conceptual_guides/adapter#low-rank-kronecker-product-lokr) expresses the weight update matrix as a decomposition of a Kronecker product, creating a block matrix that is able to preserve the rank of the original weight matrix. The size of the smaller matrices are determined by its *rank* or `r`. You'll also want to specify the `target_modules` which determines where the smaller matrices are inserted. For this guide, you'll target the *query* and *value* matrices of the attention blocks. Other important parameters to set are `alpha` (scaling factor), and `modules_to_save` (the modules apart from the LoKr layers to be trained and saved). All of these parameters - and more - are found in the [`LoKrConfig`].
+
+```py
+from peft import LoKrConfig, get_peft_model
+
+config = LoKrConfig(
+    r=16,
+    alpha=16,
+    target_modules=["query", "value"],
+    module_dropout=0.1,
+    modules_to_save=["classifier"],
+)
+model = get_peft_model(model, config)
+model.print_trainable_parameters()
+"trainable params: 116,069 || all params: 87,172,042 || trainable%: 0.13314934162033282"
+```
+
+</hfoption>
+<hfoption id="AdaLoRA">
+
+[AdaLoRA](../conceptual_guides/adapter#adaptive-low-rank-adaptation-adalora) efficiently manages the LoRA parameter budget by assigning important weight matrices more parameters and pruning less important ones. In contrast, LoRA evenly distributes parameters across all modules. You can control the average desired *rank* or `r` of the matrices, and which modules to apply AdaLoRA to with `target_modules`. Other important parameters to set are `lora_alpha` (scaling factor), and `modules_to_save` (the modules apart from the AdaLoRA layers to be trained and saved). All of these parameters - and more - are found in the [`AdaLoraConfig`].
+
+```py
+from peft import AdaLoraConfig, get_peft_model
+
+config = AdaLoraConfig(
+    r=8,
+    init_r=12,
+    tinit=200,
+    tfinal=1000,
+    deltaT=10,
+    target_modules=["query", "value"],
+    modules_to_save=["classifier"],
+)
+model = get_peft_model(model, config)
+model.print_trainable_parameters()
+"trainable params: 520,325 || all params: 87,614,722 || trainable%: 0.5938785036606062"
+```
+
+</hfoption>
+</hfoptions>
+
+### Training
+
+For training, let's use the [`~transformers.Trainer`] class from Transformers. The [`Trainer`] contains a PyTorch training loop, and when you're ready, call [`~transformers.Trainer.train`] to start training. To customize the training run, configure the training hyperparameters in the [`~transformers.TrainingArguments`] class. With LoRA-like methods, you can afford to use a higher batch size and learning rate.
+
+> [!WARNING]
+> AdaLoRA has an [`~AdaLoraModel.update_and_allocate`] method that should be called at each training step to update the parameter budget and mask, otherwise the adaptation step is not performed. This requires writing a custom training loop or subclassing the [`~transformers.Trainer`] to incorporate this method. As an example, take a look at this [custom training loop](https://github.com/huggingface/peft/blob/912ad41e96e03652cabf47522cd876076f7a0c4f/examples/conditional_generation/peft_adalora_seq2seq.py#L120).
+
+```py
+from transformers import TrainingArguments, Trainer
+
+account = "stevhliu"
+peft_model_id = f"{account}/google/vit-base-patch16-224-in21k-lora"
+batch_size = 128
+
+args = TrainingArguments(
+    peft_model_id,
+    remove_unused_columns=False,
+    eval_strategy="epoch",
+    save_strategy="epoch",
+    learning_rate=5e-3,
+    per_device_train_batch_size=batch_size,
+    gradient_accumulation_steps=4,
+    per_device_eval_batch_size=batch_size,
+    fp16=True,
+    num_train_epochs=5,
+    logging_steps=10,
+    load_best_model_at_end=True,
+    label_names=["labels"],
+)
+```
+
+Begin training with [`~transformers.Trainer.train`].
+
+```py
+trainer = Trainer(
+    model,
+    args,
+    train_dataset=train_ds,
+    eval_dataset=val_ds,
+    tokenizer=image_processor,
+    data_collator=collate_fn,
+)
+trainer.train()
+```
+
+## Share your model
+
+Once training is complete, you can upload your model to the Hub with the [`~transformers.PreTrainedModel.push_to_hub`] method. You’ll need to login to your Hugging Face account first and enter your token when prompted.
+
+```py
+from huggingface_hub import notebook_login
+
+notebook_login()
+```
+
+Call [`~transformers.PreTrainedModel.push_to_hub`] to save your model to your repositoy.
+
+```py
+model.push_to_hub(peft_model_id)
+```
+
+## Inference
+
+Let's load the model from the Hub and test it out on a food image.
+
+```py
+from peft import PeftConfig, PeftModel
+from transformers import AutoImageProcessor
+from PIL import Image
+import requests
+
+config = PeftConfig.from_pretrained("stevhliu/vit-base-patch16-224-in21k-lora")
+model = AutoModelForImageClassification.from_pretrained(
+    config.base_model_name_or_path,
+    label2id=label2id,
+    id2label=id2label,
+    ignore_mismatched_sizes=True,
+)
+model = PeftModel.from_pretrained(model, "stevhliu/vit-base-patch16-224-in21k-lora")
+
+url = "https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/beignets.jpeg"
+image = Image.open(requests.get(url, stream=True).raw)
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/beignets.jpeg">
+</div>
+
+Convert the image to RGB and return the underlying PyTorch tensors.
+
+```py
+encoding = image_processor(image.convert("RGB"), return_tensors="pt")
+```
+
+Now run the model and return the predicted class!
+
+```py
+with torch.no_grad():
+    outputs = model(**encoding)
+    logits = outputs.logits
+
+predicted_class_idx = logits.argmax(-1).item()
+print("Predicted class:", model.config.id2label[predicted_class_idx])
+"Predicted class: beignets"
+```

docs/source/task_guides/prompt_based_methods.md

@@ -0,0 +1,305 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Prompt-based methods
+
+A prompt can describe a task or provide an example of a task you want the model to learn. Instead of manually creating these prompts, soft prompting methods add learnable parameters to the input embeddings that can be optimized for a specific task while keeping the pretrained model's parameters frozen. This makes it both faster and easier to finetune large language models (LLMs) for new downstream tasks.
+
+The PEFT library supports several types of prompting methods (p-tuning, prefix tuning, prompt tuning) and you can learn more about how these methods work conceptually in the [Soft prompts](../conceptual_guides/prompting) guide. If you're interested in applying these methods to other tasks and use cases, take a look at our [notebook collection](https://huggingface.co/spaces/PEFT/soft-prompting)!
+
+This guide will show you how to train a causal language model - with a soft prompting method - to *generate a classification* for whether a tweet is a complaint or not.
+
+<Tip>
+
+Some familiarity with the general process of training a causal language model would be really helpful and allow you to focus on the soft prompting methods. If you're new, we recommend taking a look at the [Causal language modeling](https://huggingface.co/docs/transformers/tasks/language_modeling) guide first from the Transformers documentation. When you're ready, come back and see how easy it is to drop PEFT in to your training!
+
+</Tip>
+
+Before you begin, make sure you have all the necessary libraries installed.
+
+```bash
+pip install -q peft transformers datasets
+```
+
+## Dataset
+
+For this guide, you'll use the `twitter_complaints` subset of the [RAFT](https://huggingface.co/datasets/ought/raft) dataset. The `twitter_complaints` subset contains tweets labeled as `complaint` and `no complaint` and you can check out the [dataset viewer](https://huggingface.co/datasets/ought/raft/viewer/twitter_complaints) for a better idea of what the data looks like.
+
+Use the [`~datasets.load_dataset`] function to load the dataset and create a new `text_label` column so it is easier to understand what the `Label` values, `1` and `2` mean.
+
+```py
+from datasets import load_dataset
+
+ds = load_dataset("ought/raft", "twitter_complaints")
+
+classes = [k.replace("_", " ") for k in ds["train"].features["Label"].names]
+ds = ds.map(
+    lambda x: {"text_label": [classes[label] for label in x["Label"]]},
+    batched=True,
+    num_proc=1,
+)
+ds["train"][0]
+{"Tweet text": "@HMRCcustomers No this is my first job", "ID": 0, "Label": 2, "text_label": "no complaint"}
+```
+
+Load a tokenizer, define the padding token to use, and determine the maximum length of the tokenized label.
+
+```py
+from transformers import AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("bigscience/bloomz-560m")
+if tokenizer.pad_token_id is None:
+    tokenizer.pad_token_id = tokenizer.eos_token_id
+target_max_length = max([len(tokenizer(class_label)["input_ids"]) for class_label in classes])
+print(target_max_length)
+```
+
+Create a preprocessing function that tokenizes the tweet text and labels, pad the inputs and labels in each batch, create an attention mask, and truncate sequences to the `max_length`. Then convert the `input_ids`, `attention_mask`, and `labels` to PyTorch tensors.
+
+```py
+import torch
+
+max_length = 64
+
+def preprocess_function(examples, text_column="Tweet text", label_column="text_label"):
+    batch_size = len(examples[text_column])
+    inputs = [f"{text_column} : {x} Label : " for x in examples[text_column]]
+    targets = [str(x) for x in examples[label_column]]
+    model_inputs = tokenizer(inputs)
+    labels = tokenizer(targets)
+    classes = [k.replace("_", " ") for k in ds["train"].features["Label"].names]
+    for i in range(batch_size):
+        sample_input_ids = model_inputs["input_ids"][i]
+        label_input_ids = labels["input_ids"][i]
+        model_inputs["input_ids"][i] = [tokenizer.pad_token_id] * (
+            max_length - len(sample_input_ids)
+        ) + sample_input_ids
+        model_inputs["attention_mask"][i] = [0] * (max_length - len(sample_input_ids)) + model_inputs[
+            "attention_mask"
+        ][i]
+        labels["input_ids"][i] = [-100] * (max_length - len(label_input_ids)) + label_input_ids
+        model_inputs["input_ids"][i] = torch.tensor(model_inputs["input_ids"][i][:max_length])
+        model_inputs["attention_mask"][i] = torch.tensor(model_inputs["attention_mask"][i][:max_length])
+        labels["input_ids"][i] = torch.tensor(labels["input_ids"][i][:max_length])
+    model_inputs["labels"] = labels["input_ids"]
+    return model_inputs
+```
+
+Apply the preprocessing function to the entire dataset with the [`~datasets.Dataset.map`] function, and remove the unprocessed columns because the model won't need them.
+
+```py
+processed_ds = ds.map(
+    preprocess_function,
+    batched=True,
+    num_proc=1,
+    remove_columns=ds["train"].column_names,
+    load_from_cache_file=False,
+    desc="Running tokenizer on dataset",
+)
+```
+
+Finally, create a training and evaluation [`DataLoader`](https://pytorch.org/docs/stable/data.html#torch.utils.data.DataLoader). You can set `pin_memory=True` to speed up the data transfer to the GPU during training if the samples in your dataset are on a CPU.
+
+```py
+from torch.utils.data import DataLoader
+from transformers import default_data_collator
+
+train_ds = processed_ds["train"]
+eval_ds = processed_ds["test"]
+
+batch_size = 16
+
+train_dataloader = DataLoader(train_ds, shuffle=True, collate_fn=default_data_collator, batch_size=batch_size, pin_memory=True)
+eval_dataloader = DataLoader(eval_ds, collate_fn=default_data_collator, batch_size=batch_size, pin_memory=True)
+```
+
+## Model
+
+Now let's load a pretrained model to use as the base model for the soft prompt method. This guide uses the [bigscience/bloomz-560m](https://huggingface.co/bigscience/bloomz-560m) model, but you can use any causal language model you want.
+
+```py
+from transformers import AutoModelForCausalLM
+
+model = AutoModelForCausalLM.from_pretrained("bigscience/bloomz-560m")
+```
+
+### PEFT configuration and model
+
+For any PEFT method, you'll need to create a configuration which contains all the parameters that specify how the PEFT method should be applied. Once the configuration is setup, pass it to the [`~peft.get_peft_model`] function along with the base model to create a trainable [`PeftModel`].
+
+<Tip>
+
+Call the [`~PeftModel.print_trainable_parameters`] method to compare the number of trainable parameters of [`PeftModel`] versus the number of parameters in the base model!
+
+</Tip>
+
+<hfoptions id="configurations">
+<hfoption id="p-tuning">
+
+[P-tuning](../conceptual_guides/prompting#p-tuning) adds a trainable embedding tensor where the prompt tokens can be added anywhere in the input sequence. Create a [`PromptEncoderConfig`] with the task type, the number of virtual tokens to add and learn, and the hidden size of the encoder for learning the prompt parameters.
+
+```py
+from peft import PromptEncoderConfig, get_peft_model
+
+peft_config = PromptEncoderConfig(task_type="CAUSAL_LM", num_virtual_tokens=20, encoder_hidden_size=128)
+model = get_peft_model(model, peft_config)
+model.print_trainable_parameters()
+"trainable params: 300,288 || all params: 559,514,880 || trainable%: 0.05366935013417338"
+```
+
+</hfoption>
+<hfoption id="prefix tuning">
+
+[Prefix tuning](../conceptual_guides/prompting#prefix-tuning) adds task-specific parameters in all of the model layers, which are optimized by a separate feed-forward network. Create a [`PrefixTuningConfig`] with the task type and number of virtual tokens to add and learn.
+
+```py
+from peft import PrefixTuningConfig, get_peft_model
+
+peft_config = PrefixTuningConfig(task_type="CAUSAL_LM", num_virtual_tokens=20)
+model = get_peft_model(model, peft_config)
+model.print_trainable_parameters()
+"trainable params: 983,040 || all params: 560,197,632 || trainable%: 0.1754809274167014"
+```
+
+</hfoption>
+<hfoption id="prompt tuning">
+
+[Prompt tuning](../conceptual_guides/prompting#prompt-tuning) formulates all tasks as a *generation* task and it adds a task-specific prompt to the input which is updated independently. The `prompt_tuning_init_text` parameter specifies how to finetune the model (in this case, it is classifying whether tweets are complaints or not). For the best results, the `prompt_tuning_init_text` should have the same number of tokens that should be predicted. To do this, you can set `num_virtual_tokens` to the number of tokens of the `prompt_tuning_init_text`.
+
+Create a [`PromptTuningConfig`] with the task type, the initial prompt tuning text to train the model with, the number of virtual tokens to add and learn, and a tokenizer.
+
+```py
+from peft import PromptTuningConfig, PromptTuningInit, get_peft_model
+
+prompt_tuning_init_text = "Classify if the tweet is a complaint or no complaint.\n"
+peft_config = PromptTuningConfig(
+    task_type="CAUSAL_LM",
+    prompt_tuning_init=PromptTuningInit.TEXT,
+    num_virtual_tokens=len(tokenizer(prompt_tuning_init_text)["input_ids"]),
+    prompt_tuning_init_text=prompt_tuning_init_text,
+    tokenizer_name_or_path="bigscience/bloomz-560m",
+)
+model = get_peft_model(model, peft_config)
+model.print_trainable_parameters()
+"trainable params: 8,192 || all params: 559,222,784 || trainable%: 0.0014648902430985358"
+```
+
+</hfoption>
+</hfoptions>
+
+### Training
+
+Set up an optimizer and learning rate scheduler.
+
+```py
+from transformers import get_linear_schedule_with_warmup
+
+lr = 3e-2
+num_epochs = 50
+
+optimizer = torch.optim.AdamW(model.parameters(), lr=lr)
+lr_scheduler = get_linear_schedule_with_warmup(
+    optimizer=optimizer,
+    num_warmup_steps=0,
+    num_training_steps=(len(train_dataloader) * num_epochs),
+)
+```
+
+Move the model to the GPU and create a training loop that reports the loss and perplexity for each epoch.
+
+```py
+from tqdm import tqdm
+
+device = "cuda"
+model = model.to(device)
+
+for epoch in range(num_epochs):
+    model.train()
+    total_loss = 0
+    for step, batch in enumerate(tqdm(train_dataloader)):
+        batch = {k: v.to(device) for k, v in batch.items()}
+        outputs = model(**batch)
+        loss = outputs.loss
+        total_loss += loss.detach().float()
+        loss.backward()
+        optimizer.step()
+        lr_scheduler.step()
+        optimizer.zero_grad()
+
+    model.eval()
+    eval_loss = 0
+    eval_preds = []
+    for step, batch in enumerate(tqdm(eval_dataloader)):
+        batch = {k: v.to(device) for k, v in batch.items()}
+        with torch.no_grad():
+            outputs = model(**batch)
+        loss = outputs.loss
+        eval_loss += loss.detach().float()
+        eval_preds.extend(
+            tokenizer.batch_decode(torch.argmax(outputs.logits, -1).detach().cpu().numpy(), skip_special_tokens=True)
+        )
+
+    eval_epoch_loss = eval_loss / len(eval_dataloader)
+    eval_ppl = torch.exp(eval_epoch_loss)
+    train_epoch_loss = total_loss / len(train_dataloader)
+    train_ppl = torch.exp(train_epoch_loss)
+    print(f"{epoch=}: {train_ppl=} {train_epoch_loss=} {eval_ppl=} {eval_epoch_loss=}")
+```
+
+## Share your model
+
+Once training is complete, you can upload your model to the Hub with the [`~transformers.PreTrainedModel.push_to_hub`] method. You'll need to login to your Hugging Face account first and enter your token when prompted.
+
+```py
+from huggingface_hub import notebook_login
+
+account = <your-hf-account-name>
+peft_model_id = f"{account}/bloomz-560-m-peft-method"
+model.push_to_hub(peft_model_id)
+```
+
+If you check the model file size in the repository, you’ll see that it is a lot smaller than a full sized model!
+
+<div class="flex flex-col justify-center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/PEFT-hub-screenshot.png"/>
+  <figcaption class="text-center">For example, the adapter weights for a opt-350m model stored on the Hub are only ~6MB compared to the full model size which can be ~700MB.</figcaption>
+</div>
+
+## Inference
+
+Let's load the model for inference and test it out on a tweet!
+
+```py
+from peft import AutoPeftModelForCausalLM
+
+model = AutoPeftModelForCausalLM.from_pretrained("peft_model_id").to("cuda")
+tokenizer = AutoTokenizer.from_pretrained("bigscience/bloomz-560m")
+
+i = 15
+inputs = tokenizer(f'{text_column} : {ds["test"][i]["Tweet text"]} Label : ', return_tensors="pt")
+print(ds["test"][i]["Tweet text"])
+"@NYTsupport i have complained a dozen times &amp; yet my papers are still thrown FAR from my door. Why is this so hard to resolve?"
+```
+
+Call the [`~transformers.GenerationMixin.generate`] method to generate the predicted classification label.
+
+```py
+with torch.no_grad():
+    inputs = {k: v.to(device) for k, v in inputs.items()}
+    outputs = model.generate(input_ids=inputs["input_ids"], max_new_tokens=10)
+    print(tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True))
+"['Tweet text : @NYTsupport i have complained a dozen times &amp; yet my papers are still thrown FAR from my door. Why is this so hard to resolve? Label : complaint']"
+```

docs/source/tutorial/peft_integrations.md

@@ -0,0 +1,152 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# PEFT integrations
+
+PEFT's practical benefits extends to other Hugging Face libraries like [Diffusers](https://hf.co/docs/diffusers) and [Transformers](https://hf.co/docs/transformers). One of the main benefits of PEFT is that an adapter file generated by a PEFT method is a lot smaller than the original model, which makes it super easy to manage and use multiple adapters. You can use one pretrained base model for multiple tasks by simply loading a new adapter finetuned for the task you're solving. Or you can combine multiple adapters with a text-to-image diffusion model to create new effects.
+
+This tutorial will show you how PEFT can help you manage adapters in Diffusers and Transformers.
+
+## Diffusers
+
+Diffusers is a generative AI library for creating images and videos from text or images with diffusion models. LoRA is an especially popular training method for diffusion models because you can very quickly train and share diffusion models to generate images in new styles. To make it easier to use and try multiple LoRA models, Diffusers uses the PEFT library to help manage different adapters for inference.
+
+For example, load a base model and then load the [artificialguybr/3DRedmond-V1](https://huggingface.co/artificialguybr/3DRedmond-V1) adapter for inference with the [`load_lora_weights`](https://huggingface.co/docs/diffusers/v0.24.0/en/api/loaders/lora#diffusers.loaders.LoraLoaderMixin.load_lora_weights) method. The `adapter_name` argument in the loading method is enabled by PEFT and allows you to set a name for the adapter so it is easier to reference.
+
+```py
+import torch
+from diffusers import DiffusionPipeline
+
+pipeline = DiffusionPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+).to("cuda")
+pipeline.load_lora_weights(
+    "peft-internal-testing/artificialguybr__3DRedmond-V1", 
+    weight_name="3DRedmond-3DRenderStyle-3DRenderAF.safetensors", 
+    adapter_name="3d"
+)
+image = pipeline("sushi rolls shaped like kawaii cat faces").images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/test-lora-diffusers.png"/>
+</div>
+
+Now let's try another cool LoRA model, [ostris/super-cereal-sdxl-lora](https://huggingface.co/ostris/super-cereal-sdxl-lora). All you need to do is load and name this new adapter with `adapter_name`, and use the [`set_adapters`](https://huggingface.co/docs/diffusers/api/loaders/unet#diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters) method to set it as the currently active adapter.
+
+```py
+pipeline.load_lora_weights(
+    "ostris/super-cereal-sdxl-lora", 
+    weight_name="cereal_box_sdxl_v1.safetensors", 
+    adapter_name="cereal"
+)
+pipeline.set_adapters("cereal")
+image = pipeline("sushi rolls shaped like kawaii cat faces").images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/test-lora-diffusers-2.png"/>
+</div>
+
+Finally, you can call the [`disable_lora`](https://huggingface.co/docs/diffusers/api/loaders/unet#diffusers.loaders.UNet2DConditionLoadersMixin.disable_lora) method to restore the base model.
+
+```py
+pipeline.disable_lora()
+```
+
+Learn more about how PEFT supports Diffusers in the [Inference with PEFT](https://huggingface.co/docs/diffusers/tutorials/using_peft_for_inference) tutorial.
+
+## Transformers
+
+🤗 [Transformers](https://hf.co/docs/transformers) is a collection of pretrained models for all types of tasks in all modalities. You can load these models for training or inference. Many of the models are large language models (LLMs), so it makes sense to integrate PEFT with Transformers to manage and train adapters.
+
+Load a base pretrained model to train.
+
+```py
+from transformers import AutoModelForCausalLM
+
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m")
+```
+
+Next, add an adapter configuration to specify how to adapt the model parameters. Call the [`~PeftModel.add_adapter`] method to add the configuration to the base model.
+
+```py
+from peft import LoraConfig
+
+peft_config = LoraConfig(
+    lora_alpha=16,
+    lora_dropout=0.1,
+    r=64,
+    bias="none",
+    task_type="CAUSAL_LM"
+)
+model.add_adapter(peft_config)
+```
+
+Now you can train the model with Transformer's [`~transformers.Trainer`] class or whichever training framework you prefer.
+
+To use the newly trained model for inference, the [`~transformers.AutoModel`] class uses PEFT on the backend to load the adapter weights and configuration file into a base pretrained model.
+
+```py
+from transformers import AutoModelForCausalLM
+
+model = AutoModelForCausalLM.from_pretrained("peft-internal-testing/opt-350m-lora")
+```
+
+Alternatively, you can use transformers [Pipelines](https://huggingface.co/docs/transformers/en/main_classes/pipelines) to load the model for conveniently running inference:
+
+```py
+from transformers import pipeline
+
+model = pipeline("text-generation", "peft-internal-testing/opt-350m-lora")
+print(model("Hello World"))
+```
+
+If you're interested in comparing or using more than one adapter, you can call the [`~PeftModel.add_adapter`] method to add the adapter configuration to the base model. The only requirement is the adapter type must be the same (you can't mix a LoRA and LoHa adapter).
+
+```py
+from transformers import AutoModelForCausalLM
+from peft import LoraConfig
+
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m")
+model.add_adapter(lora_config_1, adapter_name="adapter_1")
+```
+
+Call [`~PeftModel.add_adapter`] again to attach a new adapter to the base model.
+
+```py
+model.add_adapter(lora_config_2, adapter_name="adapter_2")
+```
+
+Then you can use [`~PeftModel.set_adapter`] to set the currently active adapter.
+
+```py
+model.set_adapter("adapter_1")
+output = model.generate(**inputs)
+print(tokenizer.decode(output_disabled[0], skip_special_tokens=True))
+```
+
+To disable the adapter, call the [disable_adapters](https://github.com/huggingface/transformers/blob/4e3490f79b40248c53ee54365a9662611e880892/src/transformers/integrations/peft.py#L313) method.
+
+```py
+model.disable_adapters()
+```
+
+The [enable_adapters](https://github.com/huggingface/transformers/blob/4e3490f79b40248c53ee54365a9662611e880892/src/transformers/integrations/peft.py#L336) can be used to enable the adapters again.
+
+If you're curious, check out the [Load and train adapters with PEFT](https://huggingface.co/docs/transformers/main/peft) tutorial to learn more.

docs/source/tutorial/peft_model_config.md

@@ -0,0 +1,182 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# PEFT configurations and models
+
+The sheer size of today's large pretrained models - which commonly have billions of parameters - present a significant training challenge because they require more storage space and more computational power to crunch all those calculations. You'll need access to powerful GPUs or TPUs to train these large pretrained models which is expensive, not widely accessible to everyone, not environmentally friendly, and not very practical. PEFT methods address many of these challenges. There are several types of PEFT methods (soft prompting, matrix decomposition, adapters), but they all focus on the same thing, reduce the number of trainable parameters. This makes it more accessible to train and store large models on consumer hardware.
+
+The PEFT library is designed to help you quickly train large models on free or low-cost GPUs, and in this tutorial, you'll learn how to setup a configuration to apply a PEFT method to a pretrained base model for training. Once the PEFT configuration is setup, you can use any training framework you like (Transformer's [`~transformers.Trainer`] class, [Accelerate](https://hf.co/docs/accelerate), a custom PyTorch training loop).
+
+## PEFT configurations
+
+<Tip>
+
+Learn more about the parameters you can configure for each PEFT method in their respective API reference page.
+
+</Tip>
+
+A configuration stores important parameters that specify how a particular PEFT method should be applied.
+
+For example, take a look at the following [`LoraConfig`](https://huggingface.co/ybelkada/opt-350m-lora/blob/main/adapter_config.json) for applying LoRA and [`PromptEncoderConfig`](https://huggingface.co/smangrul/roberta-large-peft-p-tuning/blob/main/adapter_config.json) for applying p-tuning (these configuration files are already JSON-serialized). Whenever you load a PEFT adapter, it is a good idea to check whether it has an associated adapter_config.json file which is required.
+
+<hfoptions id="config">
+<hfoption id="LoraConfig">
+
+```json
+{
+  "base_model_name_or_path": "facebook/opt-350m", #base model to apply LoRA to
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "lora_alpha": 32,
+  "lora_dropout": 0.05,
+  "modules_to_save": null,
+  "peft_type": "LORA", #PEFT method type
+  "r": 16,
+  "revision": null,
+  "target_modules": [
+    "q_proj", #model modules to apply LoRA to (query and value projection layers)
+    "v_proj"
+  ],
+  "task_type": "CAUSAL_LM" #type of task to train model on
+}
+```
+
+You can create your own configuration for training by initializing a [`LoraConfig`].
+
+```py
+from peft import LoraConfig, TaskType
+
+lora_config = LoraConfig(
+    r=16,
+    target_modules=["q_proj", "v_proj"],
+    task_type=TaskType.CAUSAL_LM,
+    lora_alpha=32,
+    lora_dropout=0.05
+)
+```
+
+</hfoption>
+<hfoption id="PromptEncoderConfig">
+
+```json
+{
+  "base_model_name_or_path": "roberta-large", #base model to apply p-tuning to
+  "encoder_dropout": 0.0,
+  "encoder_hidden_size": 128,
+  "encoder_num_layers": 2,
+  "encoder_reparameterization_type": "MLP",
+  "inference_mode": true,
+  "num_attention_heads": 16,
+  "num_layers": 24,
+  "num_transformer_submodules": 1,
+  "num_virtual_tokens": 20,
+  "peft_type": "P_TUNING", #PEFT method type
+  "task_type": "SEQ_CLS", #type of task to train model on
+  "token_dim": 1024
+}
+```
+
+You can create your own configuration for training by initializing a [`PromptEncoderConfig`].
+
+```py
+from peft import PromptEncoderConfig, TaskType
+
+p_tuning_config = PromptEncoderConfig(
+    encoder_reparameterization_type="MLP",
+    encoder_hidden_size=128,
+    num_attention_heads=16,
+    num_layers=24,
+    num_transformer_submodules=1,
+    num_virtual_tokens=20,
+    token_dim=1024,
+    task_type=TaskType.SEQ_CLS
+)
+```
+
+</hfoption>
+</hfoptions>
+
+## PEFT models
+
+With a PEFT configuration in hand, you can now apply it to any pretrained model to create a [`PeftModel`]. Choose from any of the state-of-the-art models from the [Transformers](https://hf.co/docs/transformers) library, a custom model, and even new and unsupported transformer architectures.
+
+For this tutorial, load a base [facebook/opt-350m](https://huggingface.co/facebook/opt-350m) model to finetune.
+
+```py
+from transformers import AutoModelForCausalLM
+
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m")
+```
+
+Use the [`get_peft_model`] function to create a [`PeftModel`] from the base facebook/opt-350m model and the `lora_config` you created earlier.
+
+```py
+from peft import get_peft_model
+
+lora_model = get_peft_model(model, lora_config)
+lora_model.print_trainable_parameters()
+"trainable params: 1,572,864 || all params: 332,769,280 || trainable%: 0.472659014678278"
+```
+
+Now you can train the [`PeftModel`] with your preferred training framework! After training, you can save your model locally with [`~PeftModel.save_pretrained`] or upload it to the Hub with the [`~transformers.PreTrainedModel.push_to_hub`] method.
+
+```py
+# save locally
+lora_model.save_pretrained("your-name/opt-350m-lora")
+
+# push to Hub
+lora_model.push_to_hub("your-name/opt-350m-lora")
+```
+
+To load a [`PeftModel`] for inference, you'll need to provide the [`PeftConfig`] used to create it and the base model it was trained from.
+
+```py
+from peft import PeftModel, PeftConfig
+
+config = PeftConfig.from_pretrained("ybelkada/opt-350m-lora")
+model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path)
+lora_model = PeftModel.from_pretrained(model, "ybelkada/opt-350m-lora")
+```
+
+<Tip>
+
+By default, the [`PeftModel`] is set for inference, but if you'd like to train the adapter some more you can set `is_trainable=True`.
+
+```py
+lora_model = PeftModel.from_pretrained(model, "ybelkada/opt-350m-lora", is_trainable=True)
+```
+
+</Tip>
+
+The [`PeftModel.from_pretrained`] method is the most flexible way to load a [`PeftModel`] because it doesn't matter what model framework was used (Transformers, timm, a generic PyTorch model). Other classes, like [`AutoPeftModel`], are just a convenient wrapper around the base [`PeftModel`], and makes it easier to load PEFT models directly from the Hub or locally where the PEFT weights are stored.
+
+```py
+from peft import AutoPeftModelForCausalLM
+
+lora_model = AutoPeftModelForCausalLM.from_pretrained("ybelkada/opt-350m-lora")
+```
+
+Take a look at the [AutoPeftModel](package_reference/auto_class) API reference to learn more about the [`AutoPeftModel`] classes.
+
+## Next steps
+
+With the appropriate [`PeftConfig`], you can apply it to any pretrained model to create a [`PeftModel`] and train large powerful models faster on freely available GPUs! To learn more about PEFT configurations and models, the following guide may be helpful:
+
+* Learn how to configure a PEFT method for models that aren't from Transformers in the [Working with custom models](../developer_guides/custom_models) guide.

examples/boft_controlnet/boft_controlnet.md

@@ -0,0 +1,177 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+
+# Fine-tuning for controllable generation with BOFT (ControlNet)
+
+This guide demonstrates how to use BOFT, an orthogonal fine-tuning method, to fine-tune Stable Diffusion with either `stabilityai/stable-diffusion-2-1` or `runwayml/stable-diffusion-v1-5` model for controllable generation.
+
+By using BOFT from 🤗 PEFT, we can significantly reduce the number of trainable parameters while still achieving impressive results in various fine-tuning tasks across different foundation models. BOFT enhances model efficiency by integrating full-rank orthogonal matrices with a butterfly structure into specific model blocks, such as attention blocks, mirroring the approach used in LoRA. During fine-tuning, only these inserted matrices are trained, leaving the original model parameters untouched. During inference, the trainable BOFT paramteres can be merged into the original model, eliminating any additional computational costs.
+
+As a member of the **orthogonal finetuning** class, BOFT presents a systematic and principled method for fine-tuning. It possesses several unique properties and has demonstrated superior performance compared to LoRA in a variety of scenarios. For further details on BOFT, please consult the [PEFT's GitHub repo's concept guide OFT](https://https://huggingface.co/docs/peft/index), the [original BOFT paper](https://arxiv.org/abs/2311.06243) and the [original OFT paper](https://arxiv.org/abs/2306.07280).
+
+In this guide we provide a controllable generation (ControlNet) fine-tuning script that is available in [PEFT's GitHub repo examples](https://github.com/huggingface/peft/tree/main/examples/boft_controlnet). This implementation is adapted from [diffusers's ControlNet](https://github.com/huggingface/diffusers/tree/main/examples/controlnet) and [Hecong Wu's ControlLoRA](https://github.com/HighCWu/ControlLoRA). You can try it out and finetune on your custom images.
+
+## Set up your environment
+Start by cloning the PEFT repository:
+
+```bash
+git clone https://github.com/huggingface/peft
+```
+
+Navigate to the directory containing the training scripts for fine-tuning Dreambooth with BOFT:
+```bash
+cd peft/examples/boft_controlnet
+```
+
+Set up your environment: install PEFT, and all the required libraries. At the time of writing this guide we recommend installing PEFT from source.
+
+```bash
+conda create --name peft python=3.10
+conda activate peft
+conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=11.8 -c pytorch -c nvidia
+conda install xformers -c xformers
+pip install -r requirements.txt
+pip install git+https://github.com/huggingface/peft
+```
+
+## Data
+
+We use the [control-celeba-hq](https://huggingface.co/datasets/oftverse/control-celeba-hq) dataset for landmark-to-face controllable generation. We also provide evaluation scripts to evaluate the controllable generation performance. This task can be used to quantitatively compare different fine-tuning techniques.
+
+```bash
+export DATASET_NAME="oftverse/control-celeba-hq"
+```
+
+## Train controllable generation (ControlNet) with BOFT
+
+Start with setting some hyperparamters for BOFT:
+```bash
+PEFT_TYPE="boft"
+BLOCK_NUM=8
+BLOCK_SIZE=0
+N_BUTTERFLY_FACTOR=0
+```
+
+Here:
+
+
+Navigate to the directory containing the training scripts for fine-tuning Stable Diffusion with BOFT for controllable generation:
+
+```bash
+./train_controlnet.sh
+```
+or
+```bash
+export MODEL_NAME="stabilityai/stable-diffusion-2-1"
+# export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+
+export DATASET_NAME="oftverse/control-celeba-hq"
+export PROJECT_NAME="controlnet_${PEFT_TYPE}"
+export RUN_NAME="${PEFT_TYPE}_${BLOCK_NUM}${BLOCK_SIZE}${N_BUTTERFLY_FACTOR}"
+export CONTROLNET_PATH=""
+export OUTPUT_DIR="./output/${DATASET_NAME}/${RUN_NAME}"
+
+accelerate launch train_controlnet.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --resume_from_checkpoint=$RESUME_PATH \
+  --controlnet_model_name_or_path=$CONTROLNET_PATH \
+  --output_dir=$OUTPUT_DIR \
+  --report_to="wandb" \
+  --dataset_name=$DATASET_NAME \
+  --resolution=512 \
+  --learning_rate=1e-5 \
+  --checkpointing_steps=5000 \
+  --max_train_steps=50000 \
+  --validation_steps=2000 \
+  --num_validation_images=12 \
+  --train_batch_size=4 \
+  --dataloader_num_workers=2 \
+  --seed="0" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --wandb_project_name=$PROJECT_NAME \
+  --wandb_run_name=$RUN_NAME \
+  --enable_xformers_memory_efficient_attention \
+  --use_boft \
+  --boft_block_num=$BLOCK_NUM \
+  --boft_block_size=$BLOCK_SIZE \
+  --boft_n_butterfly_factor=$N_BUTTERFLY_FACTOR \
+  --boft_dropout=0.1 \
+  --boft_bias="boft_only" \
+  --report_to="wandb" \
+```
+
+Run inference on the saved model to sample new images from the validation set:
+
+```bash
+./test_controlnet.sh
+```
+or
+```bash
+ITER_NUM=50000
+
+export MODEL_NAME="stabilityai/stable-diffusion-2-1"
+# export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+
+export RUN_NAME="${PEFT_TYPE}_${BLOCK_NUM}${BLOCK_SIZE}${N_BUTTERFLY_FACTOR}"
+export DATASET_NAME="oftverse/control-celeba-hq"
+export CKPT_NAME="checkpoint-${ITER_NUM}"
+export OUTPUT_DIR="./output/${DATASET_NAME}/${RUN_NAME}/${CKPT_NAME}"
+export CONTROLNET_PATH="${OUTPUT_DIR}/controlnet/model.safetensors"
+export UNET_PATH="${OUTPUT_DIR}/unet/${RUN_NAME}"
+export RESULTS_PATH="${OUTPUT_DIR}/results"
+
+accelerate launch test_controlnet.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --dataset_name=$DATASET_NAME \
+  --controlnet_path=$CONTROLNET_PATH \
+  --unet_path=$UNET_PATH \
+  --adapter_name=$RUN_NAME \
+  --output_dir=$RESULTS_PATH \
+  --dataset_name=$DATASET_NAME \
+
+```
+
+Run evaluation on the sampled images to evaluate the landmark reprojection error:
+
+```bash
+./eval.sh
+```
+or
+```bash
+ITER_NUM=50000
+
+export MODEL_NAME="stabilityai/stable-diffusion-2-1"
+# export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+
+export RUN_NAME="${PEFT_TYPE}_${BLOCK_NUM}${BLOCK_SIZE}${N_BUTTERFLY_FACTOR}"
+export DATASET_NAME="oftverse/control-celeba-hq"
+export CKPT_NAME="checkpoint-${ITER_NUM}"
+export OUTPUT_DIR="./output/${DATASET_NAME}/${RUN_NAME}/${CKPT_NAME}"
+export CONTROLNET_PATH="${OUTPUT_DIR}/controlnet/model.safetensors"
+export UNET_PATH="${OUTPUT_DIR}/unet/${RUN_NAME}"
+
+accelerate launch eval.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --dataset_name=$DATASET_NAME \
+  --controlnet_path=$CONTROLNET_PATH \
+  --unet_path=$UNET_PATH \
+  --adapter_name=$RUN_NAME \
+  --output_dir=$OUTPUT_DIR \
+  --dataset_name=$DATASET_NAME \
+  --vis_overlays \
+```

examples/boft_controlnet/eval.py

@@ -0,0 +1,200 @@
+# Copyright 2023-present the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# The implementation is based on "Parameter-Efficient Orthogonal Finetuning
+# via Butterfly Factorization" (https://arxiv.org/abs/2311.06243) in ICLR 2024.
+
+import glob
+import os
+from pathlib import Path
+
+import cv2
+import face_alignment
+import numpy as np
+import torch
+from accelerate import Accelerator
+from skimage.io import imread
+from torchvision.utils import save_image
+from tqdm import tqdm
+from transformers import AutoTokenizer
+from utils.args_loader import parse_args
+from utils.dataset import make_dataset
+
+
+detect_model = face_alignment.FaceAlignment(face_alignment.LandmarksType.TWO_D, device="cuda:0", flip_input=False)
+
+# with open('./data/celebhq-text/prompt_val_blip_full.json', 'rt') as f:    # fill50k, COCO
+#     for line in f:
+#         val_data = json.loads(line)
+
+end_list = np.array([17, 22, 27, 42, 48, 31, 36, 68], dtype=np.int32) - 1
+
+
+def count_txt_files(directory):
+    pattern = os.path.join(directory, "*.txt")
+    txt_files = glob.glob(pattern)
+    return len(txt_files)
+
+
+def plot_kpts(image, kpts, color="g"):
+    """Draw 68 key points
+    Args:
+        image: the input image
+        kpt: (68, 3).
+    """
+    if color == "r":
+        c = (255, 0, 0)
+    elif color == "g":
+        c = (0, 255, 0)
+    elif color == "b":
+        c = (255, 0, 0)
+    image = image.copy()
+    kpts = kpts.copy()
+    radius = max(int(min(image.shape[0], image.shape[1]) / 200), 1)
+    for i in range(kpts.shape[0]):
+        st = kpts[i, :2]
+        if kpts.shape[1] == 4:
+            if kpts[i, 3] > 0.5:
+                c = (0, 255, 0)
+            else:
+                c = (0, 0, 255)
+        image = cv2.circle(image, (int(st[0]), int(st[1])), radius, c, radius * 2)
+        if i in end_list:
+            continue
+        ed = kpts[i + 1, :2]
+        image = cv2.line(image, (int(st[0]), int(st[1])), (int(ed[0]), int(ed[1])), (255, 255, 255), radius)
+    return image
+
+
+def generate_landmark2d(dataset, input_dir, pred_lmk_dir, gt_lmk_dir, vis=False):
+    print("Generate 2d landmarks ...")
+    os.makedirs(pred_lmk_dir, exist_ok=True)
+
+    imagepath_list = sorted(glob.glob(f"{input_dir}/pred*.png"))
+
+    for imagepath in tqdm(imagepath_list):
+        name = Path(imagepath).stem
+        idx = int(name.split("_")[-1])
+        pred_txt_path = os.path.join(pred_lmk_dir, f"{idx}.txt")
+        gt_lmk_path = os.path.join(gt_lmk_dir, f"{idx}_gt_lmk.jpg")
+        gt_txt_path = os.path.join(gt_lmk_dir, f"{idx}.txt")
+        gt_img_path = os.path.join(gt_lmk_dir, f"{idx}_gt_img.jpg")
+
+        if (not os.path.exists(pred_txt_path)) or (not os.path.exists(gt_txt_path)):
+            image = imread(imagepath)  # [:, :, :3]
+            out = detect_model.get_landmarks(image)
+            if out is None:
+                continue
+
+            pred_kpt = out[0].squeeze()
+            np.savetxt(pred_txt_path, pred_kpt)
+
+            # Your existing code for obtaining the image tensor
+            gt_lmk_img = dataset[idx]["conditioning_pixel_values"]
+            save_image(gt_lmk_img, gt_lmk_path)
+
+            gt_img = (dataset[idx]["pixel_values"]) * 0.5 + 0.5
+            save_image(gt_img, gt_img_path)
+
+            gt_img = (gt_img.permute(1, 2, 0) * 255).type(torch.uint8).cpu().numpy()
+            out = detect_model.get_landmarks(gt_img)
+            if out is None:
+                continue
+
+            gt_kpt = out[0].squeeze()
+            np.savetxt(gt_txt_path, gt_kpt)
+            # gt_image = cv2.resize(cv2.imread(gt_lmk_path), (512, 512))
+
+            if vis:
+                gt_lmk_image = cv2.imread(gt_lmk_path)
+
+                # visualize predicted landmarks
+                vis_path = os.path.join(pred_lmk_dir, f"{idx}_overlay.jpg")
+                image = cv2.imread(imagepath)
+                image_point = plot_kpts(image, pred_kpt)
+                cv2.imwrite(vis_path, np.concatenate([image_point, gt_lmk_image], axis=1))
+
+                # visualize gt landmarks
+                vis_path = os.path.join(gt_lmk_dir, f"{idx}_overlay.jpg")
+                image = cv2.imread(gt_img_path)
+                image_point = plot_kpts(image, gt_kpt)
+                cv2.imwrite(vis_path, np.concatenate([image_point, gt_lmk_image], axis=1))
+
+
+def landmark_comparison(val_dataset, lmk_dir, gt_lmk_dir):
+    print("Calculating reprojection error")
+    lmk_err = []
+
+    pbar = tqdm(range(len(val_dataset)))
+    for i in pbar:
+        # line = val_dataset[i]
+        # img_name = line["image"].split(".")[0]
+        lmk1_path = os.path.join(gt_lmk_dir, f"{i}.txt")
+        lmk1 = np.loadtxt(lmk1_path)
+        lmk2_path = os.path.join(lmk_dir, f"{i}.txt")
+
+        if not os.path.exists(lmk2_path):
+            print(f"{lmk2_path} not exist")
+            continue
+
+        lmk2 = np.loadtxt(lmk2_path)
+        lmk_err.append(np.mean(np.linalg.norm(lmk1 - lmk2, axis=1)))
+        pbar.set_description(f"lmk_err: {np.mean(lmk_err):.5f}")
+
+    print("Reprojection error:", np.mean(lmk_err))
+    np.save(os.path.join(lmk_dir, "lmk_err.npy"), lmk_err)
+
+
+def main(args):
+    logging_dir = Path(args.output_dir, args.logging_dir)
+
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_dir=logging_dir,
+    )
+
+    # Load the tokenizer
+    if args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
+    elif args.pretrained_model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.pretrained_model_name_or_path,
+            subfolder="tokenizer",
+            revision=args.revision,
+            use_fast=False,
+        )
+
+    val_dataset = make_dataset(args, tokenizer, accelerator, "test")
+
+    gt_lmk_dir = os.path.join(args.output_dir, "gt_lmk")
+    if not os.path.exists(gt_lmk_dir):
+        os.makedirs(gt_lmk_dir, exist_ok=True)
+
+    pred_lmk_dir = os.path.join(args.output_dir, "pred_lmk")
+    if not os.path.exists(pred_lmk_dir):
+        os.makedirs(pred_lmk_dir, exist_ok=True)
+
+    input_dir = os.path.join(args.output_dir, "results")
+
+    generate_landmark2d(val_dataset, input_dir, pred_lmk_dir, gt_lmk_dir, args.vis_overlays)
+
+    if count_txt_files(pred_lmk_dir) == len(val_dataset) and count_txt_files(gt_lmk_dir) == len(val_dataset):
+        landmark_comparison(val_dataset, pred_lmk_dir, gt_lmk_dir)
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

examples/boft_controlnet/eval.sh

@@ -0,0 +1,29 @@
+PEFT_TYPE="boft"
+BLOCK_NUM=8
+BLOCK_SIZE=0
+N_BUTTERFLY_FACTOR=1
+ITER_NUM=50000
+
+export RUN_NAME="${PEFT_TYPE}_${BLOCK_NUM}${BLOCK_SIZE}${N_BUTTERFLY_FACTOR}"
+
+export MODEL_NAME="stabilityai/stable-diffusion-2-1"
+# export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+
+export DATASET_NAME="oftverse/control-celeba-hq"
+export CKPT_NAME="checkpoint-${ITER_NUM}"
+export OUTPUT_DIR="./output/${DATASET_NAME}/${RUN_NAME}/${CKPT_NAME}"
+export CONTROLNET_PATH="${OUTPUT_DIR}/controlnet/model.safetensors"
+export UNET_PATH="${OUTPUT_DIR}/unet/${RUN_NAME}"
+
+
+accelerate launch eval.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --dataset_name=$DATASET_NAME \
+  --controlnet_path=$CONTROLNET_PATH \
+  --unet_path=$UNET_PATH \
+  --adapter_name=$RUN_NAME \
+  --output_dir=$OUTPUT_DIR \
+  --dataset_name=$DATASET_NAME \
+  --vis_overlays \
+
+

examples/boft_controlnet/requirements.txt

@@ -0,0 +1,8 @@
+datasets==2.16.1
+diffusers==0.17.1
+transformers==4.36.2
+accelerate==0.25.0
+wandb==0.16.1
+scikit-image==0.22.0
+opencv-python==4.9.0.80
+face-alignment==1.4.1

examples/boft_controlnet/test_controlnet.py

@@ -0,0 +1,129 @@
+# Copyright 2023-present the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# The implementation is based on "Parameter-Efficient Orthogonal Finetuning
+# via Butterfly Factorization" (https://arxiv.org/abs/2311.06243) in ICLR 2024.
+
+import os
+import sys
+import time
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+from accelerate import Accelerator
+from diffusers import DDIMScheduler
+from diffusers.utils import check_min_version
+from safetensors.torch import load_file
+from tqdm import tqdm
+from transformers import AutoTokenizer
+from utils.args_loader import parse_args
+from utils.dataset import make_dataset
+from utils.light_controlnet import ControlNetModel
+from utils.pipeline_controlnet import LightControlNetPipeline
+from utils.unet_2d_condition import UNet2DConditionNewModel
+
+
+sys.path.append("../../src")
+from peft import PeftModel  # noqa: E402
+
+
+# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
+check_min_version("0.10.0.dev0")
+device = torch.device("cuda:0")
+
+
+def main(args):
+    logging_dir = Path(args.output_dir, args.logging_dir)
+
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_dir=logging_dir,
+    )
+
+    # Load the tokenizer
+    if args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
+    elif args.pretrained_model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.pretrained_model_name_or_path,
+            subfolder="tokenizer",
+            revision=args.revision,
+            use_fast=False,
+        )
+
+    val_dataset = make_dataset(args, tokenizer, accelerator, "test")
+
+    controlnet_path = args.controlnet_path
+    unet_path = args.unet_path
+
+    controlnet = ControlNetModel()
+    controlnet.load_state_dict(load_file(controlnet_path))
+    unet = UNet2DConditionNewModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")
+    unet = PeftModel.from_pretrained(unet, unet_path, adapter_name=args.adapter_name)
+
+    pipe = LightControlNetPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        controlnet=controlnet,
+        unet=unet.model,
+        torch_dtype=torch.float32,
+        requires_safety_checker=False,
+    ).to(device)
+
+    pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+
+    if not os.path.exists(args.output_dir):
+        os.makedirs(args.output_dir, exist_ok=True)
+
+    exist_lst = [int(img.split("_")[-1][:-4]) for img in os.listdir(args.output_dir)]
+    all_lst = np.arange(len(val_dataset))
+    idx_lst = [item for item in all_lst if item not in exist_lst]
+
+    print("Number of images to be processed: ", len(idx_lst))
+
+    np.random.seed(seed=int(time.time()))
+    np.random.shuffle(idx_lst)
+
+    for idx in tqdm(idx_lst):
+        output_path = os.path.join(args.output_dir, f"pred_img_{idx:04d}.png")
+
+        if not os.path.exists(output_path):
+            data = val_dataset[idx.item()]
+            negative_prompt = "low quality, blurry, unfinished"
+
+            with torch.no_grad():
+                pred_img = pipe(
+                    data["text"],
+                    [data["conditioning_pixel_values"]],
+                    num_inference_steps=50,
+                    guidance_scale=7,
+                    negative_prompt=negative_prompt,
+                ).images[0]
+
+            pred_img.save(output_path)
+
+    # control_img = Image.fromarray(
+    #     (data["conditioning_pixel_value"] * 255).numpy().transpose(1, 2, 0).astype(np.uint8)
+    # )
+    # gt_img = Image.fromarray(
+    #     ((data["pixel_value"] + 1.0) * 0.5 * 255).numpy().transpose(1, 2, 0).astype(np.uint8)
+    # )
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

examples/boft_controlnet/test_controlnet.sh

@@ -0,0 +1,29 @@
+PEFT_TYPE="boft"
+BLOCK_NUM=8
+BLOCK_SIZE=0
+N_BUTTERFLY_FACTOR=1
+ITER_NUM=50000
+
+export RUN_NAME="${PEFT_TYPE}_${BLOCK_NUM}${BLOCK_SIZE}${N_BUTTERFLY_FACTOR}"
+
+export MODEL_NAME="stabilityai/stable-diffusion-2-1"
+# export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+
+export DATASET_NAME="oftverse/control-celeba-hq"
+export CKPT_NAME="checkpoint-${ITER_NUM}"
+export OUTPUT_DIR="./output/${DATASET_NAME}/${RUN_NAME}/${CKPT_NAME}"
+export CONTROLNET_PATH="${OUTPUT_DIR}/controlnet/model.safetensors"
+export UNET_PATH="${OUTPUT_DIR}/unet/${RUN_NAME}"
+export RESULTS_PATH="${OUTPUT_DIR}/results"
+
+
+accelerate launch test_controlnet.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --dataset_name=$DATASET_NAME \
+  --controlnet_path=$CONTROLNET_PATH \
+  --unet_path=$UNET_PATH \
+  --adapter_name=$RUN_NAME \
+  --output_dir=$RESULTS_PATH \
+  --dataset_name=$DATASET_NAME \
+
+

examples/boft_controlnet/train_controlnet.py

@@ -0,0 +1,537 @@
+#!/usr/bin/env python
+# Copyright 2023-present the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# The implementation is based on "Parameter-Efficient Orthogonal Finetuning
+# via Butterfly Factorization" (https://arxiv.org/abs/2311.06243) in ICLR 2024.
+
+import itertools
+import logging
+import math
+import os
+from pathlib import Path
+
+import datasets
+import diffusers
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+import transformers
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from accelerate.utils import set_seed
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+)
+from diffusers.optimization import get_scheduler
+from diffusers.utils import check_min_version
+from diffusers.utils.import_utils import is_xformers_available
+from packaging import version
+from tqdm.auto import tqdm
+from transformers import AutoTokenizer
+from utils.args_loader import (
+    import_model_class_from_model_name_or_path,
+    parse_args,
+)
+from utils.dataset import collate_fn, log_validation, make_dataset
+from utils.light_controlnet import ControlNetModel
+from utils.tracemalloc import TorchTracemalloc, b2mb
+from utils.unet_2d_condition import UNet2DConditionNewModel
+
+from peft import BOFTConfig, get_peft_model
+from peft.peft_model import PeftModel
+
+
+# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
+check_min_version("0.16.0.dev0")
+
+logger = get_logger(__name__)
+
+UNET_TARGET_MODULES = ["to_q", "to_v", "to_k", "query", "value", "key"]
+
+TEXT_ENCODER_TARGET_MODULES = ["q_proj", "v_proj"]
+
+
+@torch.no_grad()
+def save_adaptor(accelerator, output_dir, nets_dict):
+    for net_key in nets_dict.keys():
+        net_model = nets_dict[net_key]
+        unwarpped_net = accelerator.unwrap_model(net_model)
+
+        if isinstance(unwarpped_net, PeftModel):
+            unwarpped_net.save_pretrained(
+                os.path.join(output_dir, net_key),
+                state_dict=accelerator.get_state_dict(net_model),
+                safe_serialization=True,
+            )
+        else:
+            accelerator.save_model(
+                unwarpped_net,
+                os.path.join(output_dir, net_key),
+                safe_serialization=True,
+            )
+
+
+def main(args):
+    logging_dir = Path(args.output_dir, args.logging_dir)
+
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_dir=logging_dir,
+    )
+
+    if args.report_to == "wandb":
+        wandb_init = {
+            "wandb": {
+                "name": args.wandb_run_name,
+                "mode": "online",
+            }
+        }
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+
+    logger.info(accelerator.state, main_process_only=False)
+
+    if accelerator.is_local_main_process:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_warning()
+        diffusers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+        diffusers.utils.logging.set_verbosity_error()
+
+    # If passed along, set the training seed now.
+    if args.seed is not None:
+        set_seed(args.seed)
+
+    # Handle the repository creation
+    if accelerator.is_main_process:
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+    # Load the tokenizer
+    if args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
+    elif args.pretrained_model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.pretrained_model_name_or_path,
+            subfolder="tokenizer",
+            revision=args.revision,
+            use_fast=False,
+        )
+
+    # import correct text encoder class
+    text_encoder_cls = import_model_class_from_model_name_or_path(args.pretrained_model_name_or_path, args.revision)
+
+    # Load scheduler and models
+    noise_scheduler = DDIMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+
+    text_encoder = text_encoder_cls.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
+    )
+    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision)
+    unet = UNet2DConditionNewModel.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="unet",
+        revision=args.revision,
+    )
+
+    controlnet = ControlNetModel()
+
+    if args.controlnet_model_name_or_path != "":
+        logger.info(f"Loading existing controlnet weights from {args.controlnet_model_name_or_path}")
+        controlnet.load_state_dict(torch.load(args.controlnet_model_name_or_path))
+
+    if args.use_boft:
+        config = BOFTConfig(
+            boft_block_size=args.boft_block_size,
+            boft_block_num=args.boft_block_num,
+            boft_n_butterfly_factor=args.boft_n_butterfly_factor,
+            target_modules=UNET_TARGET_MODULES,
+            boft_dropout=args.boft_dropout,
+            bias=args.boft_bias,
+        )
+        unet = get_peft_model(unet, config)
+        unet.print_trainable_parameters()
+
+    vae.requires_grad_(False)
+    controlnet.requires_grad_(True)
+
+    if not args.train_text_encoder:
+        text_encoder.requires_grad_(False)
+
+    unet.train()
+    controlnet.train()
+
+    if args.train_text_encoder and args.use_boft:
+        config = BOFTConfig(
+            boft_block_size=args.boft_block_size,
+            boft_block_num=args.boft_block_num,
+            boft_n_butterfly_factor=args.boft_n_butterfly_factor,
+            target_modules=TEXT_ENCODER_TARGET_MODULES,
+            boft_dropout=args.boft_dropout,
+            bias=args.boft_bias,
+        )
+        text_encoder = get_peft_model(text_encoder, config, adapter_name=args.wandb_run_name)
+        text_encoder.print_trainable_parameters()
+
+    if args.train_text_encoder:
+        text_encoder.train()
+
+    # For mixed precision training we cast the text_encoder and vae weights to half-precision
+    # as these models are only used for inference, keeping weights in full precision is not required.
+    weight_dtype = torch.float32
+    if accelerator.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif accelerator.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
+    # Move unet, vae and text_encoder to device and cast to weight_dtype
+    unet.to(accelerator.device, dtype=weight_dtype)
+    vae.to(accelerator.device, dtype=weight_dtype)
+    controlnet.to(accelerator.device, dtype=weight_dtype)
+
+    if not args.train_text_encoder:
+        text_encoder.to(accelerator.device, dtype=weight_dtype)
+
+    if args.enable_xformers_memory_efficient_attention:
+        if is_xformers_available():
+            import xformers
+
+            xformers_version = version.parse(xformers.__version__)
+            if xformers_version == version.parse("0.0.16"):
+                logger.warning(
+                    "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
+                )
+            unet.enable_xformers_memory_efficient_attention()
+            controlnet.enable_xformers_memory_efficient_attention()
+            if args.train_text_encoder and not (args.use_lora or args.use_boft or args.use_oft):
+                text_encoder.enable_xformers_memory_efficient_attention()
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
+
+    if args.gradient_checkpointing:
+        controlnet.enable_gradient_checkpointing()
+        unet.enable_gradient_checkpointing()
+        if args.train_text_encoder and not (args.use_lora or args.use_boft or args.use_oft):
+            text_encoder.gradient_checkpointing_enable()
+
+    # Check that all trainable models are in full precision
+    low_precision_error_string = (
+        " Please make sure to always have all model weights in full float32 precision when starting training - even if"
+        " doing mixed precision training, copy of the weights should still be float32."
+    )
+
+    if accelerator.unwrap_model(controlnet).dtype != torch.float32:
+        raise ValueError(
+            f"Controlnet loaded as datatype {accelerator.unwrap_model(controlnet).dtype}. {low_precision_error_string}"
+        )
+
+    if accelerator.unwrap_model(unet).dtype != torch.float32:
+        raise ValueError(
+            f"UNet loaded as datatype {accelerator.unwrap_model(unet).dtype}. {low_precision_error_string}"
+        )
+
+    # Enable TF32 for faster training on Ampere GPUs,
+    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
+    if args.allow_tf32:
+        torch.backends.cuda.matmul.allow_tf32 = True
+
+    if args.scale_lr:
+        args.learning_rate = (
+            args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
+        )
+
+    # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
+    if args.use_8bit_adam:
+        try:
+            import bitsandbytes as bnb
+        except ImportError:
+            raise ImportError(
+                "To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
+            )
+
+        optimizer_class = bnb.optim.AdamW8bit
+    else:
+        optimizer_class = torch.optim.AdamW
+
+    params_to_optimize = [param for param in controlnet.parameters() if param.requires_grad]
+    params_to_optimize += [param for param in unet.parameters() if param.requires_grad]
+
+    if args.train_text_encoder:
+        params_to_optimize += [param for param in text_encoder.parameters() if param.requires_grad]
+
+    # Optimizer creation
+    optimizer = optimizer_class(
+        params_to_optimize,
+        lr=args.learning_rate,
+        betas=(args.adam_beta1, args.adam_beta2),
+        weight_decay=args.adam_weight_decay,
+        eps=args.adam_epsilon,
+    )
+
+    # Load the dataset
+    train_dataset = make_dataset(args, tokenizer, accelerator, "train")
+    val_dataset = make_dataset(args, tokenizer, accelerator, "test")
+
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        shuffle=True,
+        collate_fn=collate_fn,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+
+    # Scheduler and math around the number of training steps.
+    overrode_max_train_steps = False
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        overrode_max_train_steps = True
+
+    lr_scheduler = get_scheduler(
+        args.lr_scheduler,
+        optimizer=optimizer,
+        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
+        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_cycles=args.lr_num_cycles,
+        power=args.lr_power,
+    )
+
+    # Prepare everything with our `accelerator`.
+    controlnet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        controlnet, optimizer, train_dataloader, lr_scheduler
+    )
+
+    if args.train_text_encoder:
+        text_encoder = accelerator.prepare(text_encoder)
+
+    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if overrode_max_train_steps:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+    # Afterwards we recalculate our number of training epochs
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    # We need to initialize the trackers we use, and also store our configuration.
+    # The trackers initializes automatically on the main process.
+    if accelerator.is_main_process:
+        accelerator.init_trackers(args.wandb_project_name, config=vars(args), init_kwargs=wandb_init)
+
+    # Train!
+    total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num batches each epoch = {len(train_dataloader)}")
+    logger.info(f"  Num Epochs = {args.num_train_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
+    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
+    logger.info(f"  Total optimization steps = {args.max_train_steps}")
+
+    global_step = 0
+    first_epoch = 0
+
+    # Potentially load in the weights and states from a previous save
+
+    if args.resume_from_checkpoint:
+        if args.resume_from_checkpoint != "latest":
+            path = os.path.basename(args.resume_from_checkpoint)
+        else:
+            # Get the most recent checkpoint
+            dirs = os.listdir(args.output_dir)
+            if "checkpoint-current" in dirs:
+                path = "checkpoint-current"
+                dirs = [d for d in dirs if d.startswith("checkpoint") and d.endswith("0")]
+                dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
+
+            else:
+                dirs = [d for d in dirs if d.startswith("checkpoint")]
+                dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
+                path = dirs[-1] if len(dirs) > 0 else None
+
+        if path is None:
+            accelerator.print(
+                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
+            )
+            args.resume_from_checkpoint = None
+            initial_global_step = 0
+        else:
+            accelerator.print(f"Resuming from checkpoint {path}")
+            accelerator.load_state(os.path.join(args.output_dir, path))
+            if path.split("-")[1] == "current":
+                global_step = int(dirs[-1].split("-")[1])
+            else:
+                global_step = int(path.split("-")[1])
+
+            initial_global_step = global_step
+            resume_global_step = global_step * args.gradient_accumulation_steps
+            first_epoch = global_step // num_update_steps_per_epoch
+            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        disable=not accelerator.is_local_main_process,
+    )
+
+    progress_bar.set_description("Steps")
+
+    for epoch in range(first_epoch, args.num_train_epochs):
+        with TorchTracemalloc() as tracemalloc:
+            for step, batch in enumerate(train_dataloader):
+                # Skip steps until we reach the resumed step
+                if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
+                    if step % args.gradient_accumulation_steps == 0:
+                        progress_bar.update(1)
+                        if args.report_to == "wandb":
+                            accelerator.print(progress_bar)
+                    continue
+
+                with accelerator.accumulate(controlnet), accelerator.accumulate(unet):
+                    # Convert images to latent space
+                    latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
+                    latents = latents * vae.config.scaling_factor
+
+                    # Sample noise that we'll add to the latents
+                    noise = torch.randn_like(latents)
+                    bsz = latents.shape[0]
+
+                    # Sample a random timestep for each image
+                    timesteps = torch.randint(
+                        0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device
+                    )
+                    timesteps = timesteps.long()
+
+                    # Add noise to the latents according to the noise magnitude at each timestep
+                    # (this is the forward diffusion process)
+                    noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+                    # Get the text embedding for conditioning
+                    encoder_hidden_states = text_encoder(batch["input_ids"])[0]
+
+                    controlnet_image = batch["conditioning_pixel_values"].to(dtype=weight_dtype)
+
+                    # Get the guided hint for the UNet (320 dim)
+                    guided_hint = controlnet(
+                        controlnet_cond=controlnet_image,
+                    )
+
+                    # Predict the noise residual
+                    model_pred = unet(
+                        noisy_latents,
+                        timesteps,
+                        guided_hint=guided_hint,
+                        encoder_hidden_states=encoder_hidden_states,
+                    ).sample
+
+                    # Get the target for loss depending on the prediction type
+                    if noise_scheduler.config.prediction_type == "epsilon":
+                        target = noise
+                    elif noise_scheduler.config.prediction_type == "v_prediction":
+                        target = noise_scheduler.get_velocity(latents, noise, timesteps)
+                    else:
+                        raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
+
+                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
+
+                    accelerator.backward(loss)
+
+                    if accelerator.sync_gradients:
+                        params_to_clip = (
+                            itertools.chain(controlnet.parameters(), text_encoder.parameters())
+                            if args.train_text_encoder
+                            else itertools.chain(
+                                controlnet.parameters(),
+                            )
+                        )
+
+                        accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+
+                    optimizer.step()
+                    lr_scheduler.step()
+                    optimizer.zero_grad(set_to_none=args.set_grads_to_none)
+
+                # Checks if the accelerator has performed an optimization step behind the scenes
+                if accelerator.sync_gradients:
+                    progress_bar.update(1)
+                    if args.report_to == "wandb":
+                        accelerator.print(progress_bar)
+                    global_step += 1
+
+                    step_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
+
+                    if accelerator.is_main_process:
+                        if global_step % args.validation_steps == 0 or global_step == 1:
+                            logger.info(f"Running validation... \n Generating {args.num_validation_images} images.")
+                            logger.info("Running validation... ")
+
+                            with torch.no_grad():
+                                log_validation(val_dataset, text_encoder, unet, controlnet, args, accelerator)
+
+                        if global_step % args.checkpointing_steps == 0:
+                            save_adaptor(accelerator, step_save_path, {"controlnet": controlnet, "unet": unet})
+
+                            # save text_encoder if any
+                            if args.train_text_encoder:
+                                save_adaptor(accelerator, step_save_path, {"text_encoder": text_encoder})
+
+                            accelerator.save_state(step_save_path)
+
+                            logger.info(f"Saved {global_step} state to {step_save_path}")
+                            logger.info(f"Saved current state to {step_save_path}")
+
+                logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
+                progress_bar.set_postfix(**logs)
+                accelerator.log(logs, step=global_step)
+
+                if global_step >= args.max_train_steps:
+                    break
+
+        # Printing the GPU memory usage details such as allocated memory, peak memory, and total memory usage
+        accelerator.print(f"GPU Memory before entering the train : {b2mb(tracemalloc.begin)}")
+        accelerator.print(f"GPU Memory consumed at the end of the train (end-begin): {tracemalloc.used}")
+        accelerator.print(f"GPU Peak Memory consumed during the train (max-begin): {tracemalloc.peaked}")
+        accelerator.print(
+            f"GPU Total Peak Memory consumed during the train (max): {tracemalloc.peaked + b2mb(tracemalloc.begin)}"
+        )
+
+        accelerator.print(f"CPU Memory before entering the train : {b2mb(tracemalloc.cpu_begin)}")
+        accelerator.print(f"CPU Memory consumed at the end of the train (end-begin): {tracemalloc.cpu_used}")
+        accelerator.print(f"CPU Peak Memory consumed during the train (max-begin): {tracemalloc.cpu_peaked}")
+        accelerator.print(
+            f"CPU Total Peak Memory consumed during the train (max): {tracemalloc.cpu_peaked + b2mb(tracemalloc.cpu_begin)}"
+        )
+
+    # Create the pipeline using using the trained modules and save it.
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

examples/boft_controlnet/train_controlnet.sh

@@ -0,0 +1,42 @@
+PEFT_TYPE="boft"
+BLOCK_NUM=8
+BLOCK_SIZE=0
+N_BUTTERFLY_FACTOR=1
+
+export DATASET_NAME="oftverse/control-celeba-hq"
+export PROJECT_NAME="controlnet_${PEFT_TYPE}"
+export RUN_NAME="${PEFT_TYPE}_${BLOCK_NUM}${BLOCK_SIZE}${N_BUTTERFLY_FACTOR}"
+export CONTROLNET_PATH=""
+
+export MODEL_NAME="stabilityai/stable-diffusion-2-1"
+# export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+
+export OUTPUT_DIR="./output/${DATASET_NAME}/${RUN_NAME}"
+
+accelerate launch train_controlnet.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --resume_from_checkpoint=$RESUME_PATH \
+  --controlnet_model_name_or_path=$CONTROLNET_PATH \
+  --output_dir=$OUTPUT_DIR \
+  --report_to="wandb" \
+  --dataset_name=$DATASET_NAME \
+  --resolution=512 \
+  --learning_rate=1e-5 \
+  --checkpointing_steps=500 \
+  --max_train_steps=50000 \
+  --validation_steps=5000 \
+  --num_validation_images=12 \
+  --train_batch_size=4 \
+  --dataloader_num_workers=2 \
+  --seed="0" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --wandb_project_name=$PROJECT_NAME \
+  --wandb_run_name=$RUN_NAME \
+  --enable_xformers_memory_efficient_attention \
+  --use_boft \
+  --boft_block_num=$BLOCK_NUM \
+  --boft_block_size=$BLOCK_SIZE \
+  --boft_n_butterfly_factor=$N_BUTTERFLY_FACTOR \
+  --boft_dropout=0.1 \
+  --boft_bias="boft_only" \

examples/boft_controlnet/utils/__init__.py

@@ -0,0 +1 @@
+

examples/boft_controlnet/utils/args_loader.py

@@ -0,0 +1,447 @@
+import argparse
+import os
+from typing import Optional
+
+from huggingface_hub import HfFolder, whoami
+from transformers import PretrainedConfig
+
+
+def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
+    if token is None:
+        token = HfFolder.get_token()
+    if organization is None:
+        username = whoami(token)["name"]
+        return f"{username}/{model_id}"
+    else:
+        return f"{organization}/{model_id}"
+
+
+def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: str, revision: str):
+    text_encoder_config = PretrainedConfig.from_pretrained(
+        pretrained_model_name_or_path,
+        subfolder="text_encoder",
+        revision=revision,
+    )
+    model_class = text_encoder_config.architectures[0]
+
+    if model_class == "CLIPTextModel":
+        from transformers import CLIPTextModel
+
+        return CLIPTextModel
+    elif model_class == "RobertaSeriesModelWithTransformation":
+        from diffusers.pipelines.alt_diffusion.modeling_roberta_series import (
+            RobertaSeriesModelWithTransformation,
+        )
+
+        return RobertaSeriesModelWithTransformation
+    else:
+        raise ValueError(f"{model_class} is not supported.")
+
+
+def parse_args(input_args=None):
+    parser = argparse.ArgumentParser(description="Simple example of a ControlNet training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--controlnet_model_name_or_path",
+        type=str,
+        default=None,
+        help="Path to pretrained controlnet model or model identifier from huggingface.co/models."
+        " If not specified controlnet weights are initialized from unet.",
+    )
+    parser.add_argument(
+        "--revision",
+        type=str,
+        default=None,
+        required=False,
+        help=(
+            "Revision of pretrained model identifier from huggingface.co/models. Trainable model components should be"
+            " float32 precision."
+        ),
+    )
+    parser.add_argument(
+        "--tokenizer_name",
+        type=str,
+        default=None,
+        help="Pretrained tokenizer name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="controlnet-model",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--cache_dir",
+        type=str,
+        default=None,
+        help="The directory where the downloaded models and datasets will be stored.",
+    )
+    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=(
+            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument("--train_text_encoder", action="store_true", help="Whether to train the text encoder")
+
+    parser.add_argument(
+        "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument(
+        "--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images."
+    )
+
+    parser.add_argument("--num_train_epochs", type=int, default=1)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--checkpointing_steps",
+        type=int,
+        default=500,
+        help=(
+            "Save a checkpoint of the training state every X updates. Checkpoints can be used for resuming training via `--resume_from_checkpoint`. "
+            "In the case that the checkpoint is better than the final trained model, the checkpoint can also be used for inference."
+            "Using a checkpoint for inference requires separate loading of the original pipeline and the individual checkpointed model components."
+            "See https://huggingface.co/docs/diffusers/main/en/training/dreambooth#performing-inference-using-a-saved-checkpoint for step by step"
+            "instructions."
+        ),
+    )
+    parser.add_argument(
+        "--checkpoints_total_limit",
+        type=int,
+        default=None,
+        help=("Max number of checkpoints to store."),
+    )
+    parser.add_argument(
+        "--resume_from_checkpoint",
+        type=str,
+        default=None,
+        help=(
+            "Whether training should be resumed from a previous checkpoint. Use a path saved by"
+            ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
+        ),
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--gradient_checkpointing",
+        action="store_true",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=5e-6,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=False,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        help=(
+            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+            ' "constant", "constant_with_warmup"]'
+        ),
+    )
+    parser.add_argument(
+        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
+    )
+    parser.add_argument(
+        "--lr_num_cycles",
+        type=int,
+        default=1,
+        help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
+    )
+    parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
+    parser.add_argument(
+        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
+    )
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--allow_tf32",
+        action="store_true",
+        help=(
+            "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
+            " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
+        ),
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="wandb",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
+            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
+        ),
+    )
+    parser.add_argument(
+        "--wandb_key",
+        type=str,
+        default=None,
+        help=("If report to option is set to wandb, api-key for wandb used for login to wandb "),
+    )
+    parser.add_argument(
+        "--wandb_project_name",
+        type=str,
+        default=None,
+        help=("If report to option is set to wandb, project name in wandb for log tracking  "),
+    )
+    parser.add_argument(
+        "--wandb_run_name",
+        type=str,
+        default=None,
+        help=("If report to option is set to wandb, project name in wandb for log tracking  "),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
+        ),
+    )
+    parser.add_argument(
+        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
+    )
+    parser.add_argument(
+        "--set_grads_to_none",
+        action="store_true",
+        help=(
+            "Save more memory by using setting grads to None instead of zero. Be aware, that this changes certain"
+            " behaviors, so disable this argument if it causes any problems. More info:"
+            " https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html"
+        ),
+    )
+    parser.add_argument(
+        "--dataset_name",
+        type=str,
+        default=None,
+        help=(
+            "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
+            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
+            " or to a folder containing files that 🤗 Datasets can understand."
+        ),
+    )
+    parser.add_argument(
+        "--dataset_config_name",
+        type=str,
+        default=None,
+        help="The config of the Dataset, leave as None if there's only one config.",
+    )
+    parser.add_argument(
+        "--train_data_dir",
+        type=str,
+        default=None,
+        help=(
+            "A folder containing the training data. Folder contents must follow the structure described in"
+            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
+            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
+        ),
+    )
+    parser.add_argument(
+        "--image_column", type=str, default="image", help="The column of the dataset containing the target image."
+    )
+    parser.add_argument(
+        "--conditioning_image_column",
+        type=str,
+        default="conditioning_image",
+        help="The column of the dataset containing the controlnet conditioning image.",
+    )
+    parser.add_argument(
+        "--caption_column",
+        type=str,
+        default="text",
+        help="The column of the dataset containing a caption or a list of captions.",
+    )
+    parser.add_argument(
+        "--max_train_samples",
+        type=int,
+        default=None,
+        help=(
+            "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        ),
+    )
+    parser.add_argument(
+        "--proportion_empty_prompts",
+        type=float,
+        default=0,
+        help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).",
+    )
+    parser.add_argument(
+        "--validation_prompt",
+        type=str,
+        default=None,
+        nargs="+",
+        help=(
+            "A set of prompts evaluated every `--validation_steps` and logged to `--report_to`."
+            " Provide either a matching number of `--validation_image`s, a single `--validation_image`"
+            " to be used with all prompts, or a single prompt that will be used with all `--validation_image`s."
+        ),
+    )
+    parser.add_argument(
+        "--validation_image",
+        type=str,
+        default=None,
+        nargs="+",
+        help=(
+            "A set of paths to the controlnet conditioning image be evaluated every `--validation_steps`"
+            " and logged to `--report_to`. Provide either a matching number of `--validation_prompt`s, a"
+            " a single `--validation_prompt` to be used with all `--validation_image`s, or a single"
+            " `--validation_image` that will be used with all `--validation_prompt`s."
+        ),
+    )
+    parser.add_argument(
+        "--num_validation_images",
+        type=int,
+        default=4,
+        help="Number of images to be generated for each `--validation_image`, `--validation_prompt` pair",
+    )
+    parser.add_argument(
+        "--validation_steps",
+        type=int,
+        default=100,
+        help=(
+            "Run validation every X steps. Validation consists of running the prompt"
+            " `args.validation_prompt` multiple times: `args.num_validation_images`"
+            " and logging the images."
+        ),
+    )
+    parser.add_argument(
+        "--tracker_project_name",
+        type=str,
+        default="train_controlnet",
+        help=(
+            "The `project_name` argument passed to Accelerator.init_trackers for"
+            " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
+        ),
+    )
+
+    # evaluation arguments
+    parser.add_argument("--controlnet_path", type=str, default=None, help="Path to pretrained controlnet.")
+    parser.add_argument("--unet_path", type=str, default=None, help="Path to pretrained unet.")
+    parser.add_argument("--adapter_name", type=str, default=None, help="Name of the adapter to use.")
+    parser.add_argument("--vis_overlays", action="store_true", help="Whether to visualize the landmarks.")
+
+    # self-invented arguments
+
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    parser.add_argument(
+        "--name",
+        type=str,
+        help=("The name of the current experiment run, consists of [data]-[prompt]"),
+    )
+
+    # BOFT args
+    parser.add_argument("--use_boft", action="store_true", help="Whether to use BOFT for parameter efficient tuning")
+    parser.add_argument("--boft_block_num", type=int, default=8, help="The number of BOFT blocks")
+    parser.add_argument("--boft_block_size", type=int, default=0, help="The size of BOFT blocks")
+    parser.add_argument("--boft_n_butterfly_factor", type=int, default=0, help="The number of butterfly factors")
+    parser.add_argument("--boft_dropout", type=float, default=0.1, help="BOFT dropout, only used if use_boft is True")
+    parser.add_argument(
+        "--boft_bias",
+        type=str,
+        default="none",
+        help="Bias type for BOFT. Can be 'none', 'all' or 'boft_only', only used if use_boft is True",
+    )
+
+    if input_args is not None:
+        args = parser.parse_args(input_args)
+    else:
+        args = parser.parse_args()
+
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError("Specify either `--dataset_name` or `--train_data_dir`")
+
+    if args.dataset_name is not None and args.train_data_dir is not None:
+        raise ValueError("Specify only one of `--dataset_name` or `--train_data_dir`")
+
+    if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1:
+        raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].")
+
+    if args.validation_prompt is not None and args.validation_image is None:
+        raise ValueError("`--validation_image` must be set if `--validation_prompt` is set")
+
+    if args.validation_prompt is None and args.validation_image is not None:
+        raise ValueError("`--validation_prompt` must be set if `--validation_image` is set")
+
+    if (
+        args.validation_image is not None
+        and args.validation_prompt is not None
+        and len(args.validation_image) != 1
+        and len(args.validation_prompt) != 1
+        and len(args.validation_image) != len(args.validation_prompt)
+    ):
+        raise ValueError(
+            "Must provide either 1 `--validation_image`, 1 `--validation_prompt`,"
+            " or the same number of `--validation_prompt`s and `--validation_image`s"
+        )
+
+    if args.resolution % 8 != 0:
+        raise ValueError(
+            "`--resolution` must be divisible by 8 for consistently sized encoded images between the VAE and the controlnet encoder."
+        )
+
+    return args

examples/boft_controlnet/utils/dataset.py

@@ -0,0 +1,207 @@
+import random
+
+import numpy as np
+import torch
+import wandb
+from datasets import load_dataset
+from diffusers import DDIMScheduler
+from PIL import Image
+from torchvision import transforms
+from utils.pipeline_controlnet import LightControlNetPipeline
+
+
+def image_grid(imgs, rows, cols):
+    assert len(imgs) == rows * cols
+
+    w, h = imgs[0].size
+    grid = Image.new("RGB", size=(cols * w, rows * h))
+
+    for i, img in enumerate(imgs):
+        grid.paste(img, box=(i % cols * w, i // cols * h))
+    return grid
+
+
+def log_validation(val_dataset, text_encoder, unet, controlnet, args, accelerator):
+    pipeline = LightControlNetPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        controlnet=accelerator.unwrap_model(controlnet, keep_fp32_wrapper=True),
+        unet=accelerator.unwrap_model(unet, keep_fp32_wrapper=True).model,
+        text_encoder=accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=True),
+        safety_checker=None,
+        revision=args.revision,
+    )
+
+    pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
+    pipeline = pipeline.to(accelerator.device)
+
+    pipeline.set_progress_bar_config(disable=True)
+
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)
+
+    image_logs = []
+
+    for idx in range(args.num_validation_images):
+        data = val_dataset[idx]
+        validation_prompt = data["text"]
+        validation_image = data["conditioning_pixel_values"]
+
+        image = pipeline(
+            validation_prompt,
+            [validation_image],
+            num_inference_steps=50,
+            generator=generator,
+        )[0][0]
+
+        image_logs.append(
+            {
+                "validation_image": validation_image,
+                "image": image,
+                "validation_prompt": validation_prompt,
+            }
+        )
+
+    for tracker in accelerator.trackers:
+        formatted_images = []
+
+        for log in image_logs:
+            image = log["image"]
+            validation_prompt = log["validation_prompt"]
+            validation_image = log["validation_image"]
+
+            formatted_images.append(wandb.Image(validation_image, caption="Controlnet conditioning"))
+
+            image = wandb.Image(image, caption=validation_prompt)
+            formatted_images.append(image)
+
+        tracker.log({"validation": formatted_images})
+
+    del pipeline
+    torch.cuda.empty_cache()
+
+
+def make_dataset(args, tokenizer, accelerator, split="train"):
+    # Get the datasets: you can either provide your own training and evaluation files (see below)
+    # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
+
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        dataset = load_dataset(
+            args.dataset_name,
+            args.dataset_config_name,
+            cache_dir=args.cache_dir,
+        )
+    else:
+        if args.train_data_dir is not None:
+            dataset = load_dataset(
+                args.train_data_dir,
+                cache_dir=args.cache_dir,
+            )
+        # See more about loading custom images at
+        # https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script
+
+    # Preprocessing the datasets.
+    # We need to tokenize inputs and targets.
+    column_names = dataset[split].column_names
+
+    # Get the column names for input/target.
+    if args.image_column is None:
+        image_column = column_names[0]
+    else:
+        image_column = args.image_column
+        if image_column not in column_names:
+            raise ValueError(
+                f"`--image_column` value '{args.image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+            )
+
+    if args.caption_column is None:
+        caption_column = column_names[1]
+    else:
+        caption_column = args.caption_column
+        if caption_column not in column_names:
+            raise ValueError(
+                f"`--caption_column` value '{args.caption_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+            )
+
+    if args.conditioning_image_column is None:
+        conditioning_image_column = column_names[2]
+    else:
+        conditioning_image_column = args.conditioning_image_column
+        if conditioning_image_column not in column_names:
+            raise ValueError(
+                f"`--conditioning_image_column` value '{args.conditioning_image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+            )
+
+    def tokenize_captions(examples, is_train=True):
+        captions = []
+        for caption in examples[caption_column]:
+            if random.random() < args.proportion_empty_prompts:
+                captions.append("")
+            elif isinstance(caption, str):
+                captions.append(caption)
+            elif isinstance(caption, (list, np.ndarray)):
+                # take a random caption if there are multiple
+                captions.append(random.choice(caption) if is_train else caption[0])
+            else:
+                raise ValueError(
+                    f"Caption column `{caption_column}` should contain either strings or lists of strings."
+                )
+        inputs = tokenizer(
+            captions, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
+        )
+        return inputs.input_ids
+
+    image_transforms = transforms.Compose(
+        [
+            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.CenterCrop(args.resolution),
+            transforms.ToTensor(),
+            transforms.Normalize([0.5], [0.5]),
+        ]
+    )
+
+    conditioning_image_transforms = transforms.Compose(
+        [
+            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.CenterCrop(args.resolution),
+            transforms.ToTensor(),
+        ]
+    )
+
+    def preprocess_train(examples):
+        images = [image.convert("RGB") for image in examples[image_column]]
+        images = [image_transforms(image) for image in images]
+
+        conditioning_images = [image.convert("RGB") for image in examples[conditioning_image_column]]
+        conditioning_images = [conditioning_image_transforms(image) for image in conditioning_images]
+
+        examples["pixel_values"] = images
+        examples["conditioning_pixel_values"] = conditioning_images
+        examples["input_ids"] = tokenize_captions(examples)
+
+        return examples
+
+    with accelerator.main_process_first():
+        if args.max_train_samples is not None:
+            dataset[split] = dataset[split].shuffle(seed=args.seed).select(range(args.max_train_samples))
+        # Set the training transforms
+        split_dataset = dataset[split].with_transform(preprocess_train)
+
+    return split_dataset
+
+
+def collate_fn(examples):
+    pixel_values = torch.stack([example["pixel_values"] for example in examples])
+    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+
+    conditioning_pixel_values = torch.stack([example["conditioning_pixel_values"] for example in examples])
+    conditioning_pixel_values = conditioning_pixel_values.to(memory_format=torch.contiguous_format).float()
+
+    input_ids = torch.stack([example["input_ids"] for example in examples])
+
+    return {
+        "pixel_values": pixel_values,
+        "conditioning_pixel_values": conditioning_pixel_values,
+        "input_ids": input_ids,
+    }

examples/boft_controlnet/utils/light_controlnet.py

@@ -0,0 +1,263 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.attention_processor import AttentionProcessor, AttnProcessor
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.unet_2d_blocks import (
+    CrossAttnDownBlock2D,
+    DownBlock2D,
+)
+from diffusers.utils import BaseOutput, logging
+from torch import nn
+from torch.nn import functional as F
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class ControlNetOutput(BaseOutput):
+    down_block_res_samples: Tuple[torch.Tensor]
+    mid_block_res_sample: torch.Tensor
+
+
+class ControlNetConditioningEmbedding(nn.Module):
+    """
+    Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
+    [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
+    training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
+    convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
+    (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
+    model) to encode image-space conditions ... into feature maps ..."
+    """
+
+    def __init__(
+        self,
+        conditioning_embedding_channels: int,
+        conditioning_channels: int = 3,
+        block_out_channels: Tuple[int] = (16, 32, 96, 256),
+    ):
+        super().__init__()
+
+        self.conv_in = nn.Conv2d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
+
+        self.blocks = nn.ModuleList([])
+
+        for i in range(len(block_out_channels) - 1):
+            channel_in = block_out_channels[i]
+            channel_out = block_out_channels[i + 1]
+            self.blocks.append(nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1))
+            self.blocks.append(nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=2))
+
+        self.conv_out = zero_module(
+            nn.Conv2d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
+        )
+
+    def forward(self, conditioning):
+        embedding = self.conv_in(conditioning)
+        embedding = F.silu(embedding)
+
+        for block in self.blocks:
+            embedding = block(embedding)
+            embedding = F.silu(embedding)
+
+        embedding = self.conv_out(embedding)
+
+        return embedding
+
+
+class ControlNetModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 4,
+        out_channels: int = 320,
+        controlnet_conditioning_channel_order: str = "rgb",
+        conditioning_embedding_out_channels: Optional[Tuple[int]] = (16, 32, 96, 256),
+    ):
+        super().__init__()
+
+        # for control image
+        self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
+            conditioning_embedding_channels=out_channels,
+            block_out_channels=conditioning_embedding_out_channels,
+        )
+
+    @property
+    # Copied from diffusers.models.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, "set_processor"):
+                processors[f"{name}.processor"] = module.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.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Parameters:
+            `processor (`dict` of `AttentionProcessor` or `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                of **all** `Attention` layers.
+            In case `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.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        self.set_attn_processor(AttnProcessor())
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                `"max"`, maximum amount of memory will be saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        controlnet_cond: torch.FloatTensor,
+    ) -> Union[ControlNetOutput, Tuple]:
+        # check channel order
+        channel_order = self.config.controlnet_conditioning_channel_order
+
+        if channel_order == "rgb":
+            # in rgb order by default
+            ...
+        elif channel_order == "bgr":
+            controlnet_cond = torch.flip(controlnet_cond, dims=[1])
+        else:
+            raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
+
+        # 2. pre-process
+
+        controlnet_cond = self.controlnet_cond_embedding(controlnet_cond)
+
+        return controlnet_cond
+
+
+def zero_module(module):
+    for p in module.parameters():
+        nn.init.zeros_(p)
+    return module

examples/boft_controlnet/utils/pipeline_controlnet.py

@@ -0,0 +1,452 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
+from diffusers.pipelines.controlnet.pipeline_controlnet import StableDiffusionControlNetPipeline
+from diffusers.utils import BaseOutput, is_compiled_module, logging
+from torch.nn import functional as F
+from utils.light_controlnet import ControlNetModel
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class LightControlNetPipelineOutput(BaseOutput):
+    """
+    Output class for Stable Diffusion pipelines.
+
+    Args:
+        images (`List[PIL.Image.Image]` or `np.ndarray`)
+            List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+            num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+        nsfw_content_detected (`List[bool]`)
+            List of flags denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, or `None` if safety checking could not be performed.
+    """
+
+    images: Union[List[PIL.Image.Image], np.ndarray]
+    nsfw_content_detected: Optional[List[bool]]
+
+
+class LightControlNetPipeline(StableDiffusionControlNetPipeline):
+    _optional_components = ["safety_checker", "feature_extractor"]
+
+    def check_inputs(
+        self,
+        prompt,
+        image,
+        callback_steps,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        controlnet_conditioning_scale=1.0,
+    ):
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        # `prompt` needs more sophisticated handling when there are multiple
+        # conditionings.
+        if isinstance(self.controlnet, MultiControlNetModel):
+            if isinstance(prompt, list):
+                logger.warning(
+                    f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
+                    " prompts. The conditionings will be fixed across the prompts."
+                )
+
+        # Check `image`
+        is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+            self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+        )
+
+        if (
+            isinstance(self.controlnet, ControlNetModel)
+            or is_compiled
+            and isinstance(self.controlnet._orig_mod, ControlNetModel)
+        ):
+            self.check_image(image, prompt, prompt_embeds)
+        elif (
+            isinstance(self.controlnet, MultiControlNetModel)
+            or is_compiled
+            and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+        ):
+            if not isinstance(image, list):
+                raise TypeError("For multiple controlnets: `image` must be type `list`")
+
+            # When `image` is a nested list:
+            # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
+            elif any(isinstance(i, list) for i in image):
+                raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+            elif len(image) != len(self.controlnet.nets):
+                raise ValueError(
+                    "For multiple controlnets: `image` must have the same length as the number of controlnets."
+                )
+
+            for image_ in image:
+                self.check_image(image_, prompt, prompt_embeds)
+        else:
+            assert False
+
+        # Check `controlnet_conditioning_scale`
+        if (
+            isinstance(self.controlnet, ControlNetModel)
+            or is_compiled
+            and isinstance(self.controlnet._orig_mod, ControlNetModel)
+        ):
+            if not isinstance(controlnet_conditioning_scale, float):
+                raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+        elif (
+            isinstance(self.controlnet, MultiControlNetModel)
+            or is_compiled
+            and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+        ):
+            if isinstance(controlnet_conditioning_scale, list):
+                if any(isinstance(i, list) for i in controlnet_conditioning_scale):
+                    raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+            elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
+                self.controlnet.nets
+            ):
+                raise ValueError(
+                    "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
+                    " the same length as the number of controlnets"
+                )
+        else:
+            assert False
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        image: Union[
+            torch.FloatTensor,
+            PIL.Image.Image,
+            np.ndarray,
+            List[torch.FloatTensor],
+            List[PIL.Image.Image],
+            List[np.ndarray],
+        ] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+        guess_mode: bool = False,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+                    `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+                The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. If
+                the type is specified as `Torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can
+                also be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If
+                height and/or width are passed, `image` is resized according to them. If multiple ControlNets are
+                specified in init, images must be passed as a list such that each element of the list can be correctly
+                batched for input to a single controlnet.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added
+                to the residual in the original unet. If multiple ControlNets are specified in init, you can set the
+                corresponding scale as a list.
+            guess_mode (`bool`, *optional*, defaults to `False`):
+                In this mode, the ControlNet encoder will try best to recognize the content of the input image even if
+                you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            image,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            controlnet_conditioning_scale,
+        )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        # 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`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+        if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
+            controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        prompt_embeds = self._encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+        )
+
+        # 4. Prepare image
+        if isinstance(controlnet, ControlNetModel):
+            image = self.prepare_image(
+                image=image,
+                width=width,
+                height=height,
+                batch_size=batch_size * num_images_per_prompt,
+                num_images_per_prompt=num_images_per_prompt,
+                device=device,
+                dtype=controlnet.dtype,
+                do_classifier_free_guidance=do_classifier_free_guidance,
+                guess_mode=guess_mode,
+            )
+            height, width = image.shape[-2:]
+        elif isinstance(controlnet, MultiControlNetModel):
+            images = []
+
+            for image_ in image:
+                image_ = self.prepare_image(
+                    image=image_,
+                    width=width,
+                    height=height,
+                    batch_size=batch_size * num_images_per_prompt,
+                    num_images_per_prompt=num_images_per_prompt,
+                    device=device,
+                    dtype=controlnet.dtype,
+                    do_classifier_free_guidance=do_classifier_free_guidance,
+                    guess_mode=guess_mode,
+                )
+
+                images.append(image_)
+
+            image = images
+            height, width = image[0].shape[-2:]
+        else:
+            assert False
+
+        # 5. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 6. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 8. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # controlnet(s) inference
+                if guess_mode and do_classifier_free_guidance:
+                    # Infer ControlNet only for the conditional batch.
+                    control_model_input = latents
+                    control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+                else:
+                    control_model_input = latent_model_input
+
+                # Get the guided hint for the UNet (320 dim)
+                guided_hint = self.controlnet(
+                    controlnet_cond=image,
+                )
+
+                # Predict the noise residual
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    guided_hint=guided_hint,
+                    encoder_hidden_states=prompt_embeds,
+                )[0]
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        callback(i, t, latents)
+
+        # If we do sequential model offloading, let's offload unet and controlnet
+        # manually for max memory savings
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.unet.to("cpu")
+            self.controlnet.to("cpu")
+            torch.cuda.empty_cache()
+
+        if not output_type == "latent":
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+        else:
+            image = latents
+            has_nsfw_concept = None
+
+        if has_nsfw_concept is None:
+            do_denormalize = [True] * image.shape[0]
+        else:
+            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+        # Offload last model to CPU
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.final_offload_hook.offload()
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return LightControlNetPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

examples/boft_controlnet/utils/tracemalloc.py

@@ -0,0 +1,58 @@
+import gc
+import threading
+
+import psutil
+import torch
+
+
+# Converting Bytes to Megabytes
+def b2mb(x):
+    return int(x / 2**20)
+
+
+# This context manager is used to track the peak memory usage of the process
+class TorchTracemalloc:
+    def __enter__(self):
+        gc.collect()
+        torch.cuda.empty_cache()
+        torch.cuda.reset_max_memory_allocated()  # reset the peak gauge to zero
+        self.begin = torch.cuda.memory_allocated()
+        self.process = psutil.Process()
+
+        self.cpu_begin = self.cpu_mem_used()
+        self.peak_monitoring = True
+        peak_monitor_thread = threading.Thread(target=self.peak_monitor_func)
+        peak_monitor_thread.daemon = True
+        peak_monitor_thread.start()
+        return self
+
+    def cpu_mem_used(self):
+        """get resident set size memory for the current process"""
+        return self.process.memory_info().rss
+
+    def peak_monitor_func(self):
+        self.cpu_peak = -1
+
+        while True:
+            self.cpu_peak = max(self.cpu_mem_used(), self.cpu_peak)
+
+            # can't sleep or will not catch the peak right (this comment is here on purpose)
+            # time.sleep(0.001) # 1msec
+
+            if not self.peak_monitoring:
+                break
+
+    def __exit__(self, *exc):
+        self.peak_monitoring = False
+
+        gc.collect()
+        torch.cuda.empty_cache()
+        self.end = torch.cuda.memory_allocated()
+        self.peak = torch.cuda.max_memory_allocated()
+        self.used = b2mb(self.end - self.begin)
+        self.peaked = b2mb(self.peak - self.begin)
+
+        self.cpu_end = self.cpu_mem_used()
+        self.cpu_used = b2mb(self.cpu_end - self.cpu_begin)
+        self.cpu_peaked = b2mb(self.cpu_peak - self.cpu_begin)
+        # print(f"delta used/peak {self.used:4d}/{self.peaked:4d}")

examples/boft_controlnet/utils/unet_2d_condition.py

@@ -0,0 +1,277 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+from diffusers.models import UNet2DConditionModel
+from diffusers.utils import BaseOutput, logging
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Hidden states conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor
+
+
+class UNet2DConditionNewModel(UNet2DConditionModel):
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        guided_hint: Optional[torch.Tensor] = None,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
+            timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
+            encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
+            encoder_attention_mask (`torch.Tensor`):
+                (batch, sequence_length) cross-attention mask, applied to encoder_hidden_states. True = keep, False =
+                discard. Mask will be converted into a bias, which adds large negative values to attention scores
+                corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+            added_cond_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified includes additonal conditions that can be used for additonal time
+                embeddings or encoder hidden states projections. See the configurations `encoder_hid_dim_type` and
+                `addition_embed_type` for more information.
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+            emb = emb + aug_emb
+        elif self.config.addition_embed_type == "text_image":
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+
+            aug_emb = self.add_embedding(text_embs, image_embs)
+            emb = emb + aug_emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+
+        # 2. pre-process and insert conditioning (ControlNet)
+        # Note: the added "guided_hint" is the only difference between this implementation and the original UNet2DConditionModel
+        sample = self.conv_in(sample)
+        sample = guided_hint + sample if guided_hint is not None else sample
+
+        # 3. down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+            down_block_res_samples += res_samples
+
+        if down_block_additional_residuals is not None:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            sample = self.mid_block(
+                sample,
+                emb,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=attention_mask,
+                cross_attention_kwargs=cross_attention_kwargs,
+                encoder_attention_mask=encoder_attention_mask,
+            )
+
+        if mid_block_additional_residual is not None:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
+                )
+
+        # 6. post-process
+        if self.conv_norm_out:
+            sample = self.conv_norm_out(sample)
+            sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

examples/boft_dreambooth/.gitignore

@@ -0,0 +1 @@
+data/

examples/boft_dreambooth/boft_dreambooth.md

@@ -0,0 +1,165 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# DreamBooth fine-tuning with BOFT
+
+This guide demonstrates how to use BOFT, an orthogonal fine-tuning method, to fine-tune Dreambooth with either `stabilityai/stable-diffusion-2-1` or `runwayml/stable-diffusion-v1-5` model.
+
+By using BOFT from 🤗 PEFT, we can significantly reduce the number of trainable parameters while still achieving impressive results in various fine-tuning tasks across different foundation models. BOFT enhances model efficiency by integrating full-rank orthogonal matrices with a butterfly structure into specific model blocks, such as attention blocks, mirroring the approach used in LoRA. During fine-tuning, only these inserted matrices are trained, leaving the original model parameters untouched. During inference, the trainable BOFT paramteres can be merged into the original model, eliminating any additional computational costs.
+
+As a member of the **orthogonal finetuning** class, BOFT presents a systematic and principled method for fine-tuning. It possesses several unique properties and has demonstrated superior performance compared to LoRA in a variety of scenarios. For further details on BOFT, please consult the [PEFT's GitHub repo's concept guide OFT](https://https://huggingface.co/docs/peft/index), the [original BOFT paper](https://arxiv.org/abs/2311.06243) and the [original OFT paper](https://arxiv.org/abs/2306.07280).
+
+In this guide we provide a Dreambooth fine-tuning script that is available in [PEFT's GitHub repo examples](https://github.com/huggingface/peft/tree/main/examples/boft_dreambooth). This implementation is adapted from [peft's lora_dreambooth](https://github.com/huggingface/peft/tree/main/examples/lora_dreambooth). You can try it out and finetune on your custom images.
+
+## Set up your environment
+
+Start by cloning the PEFT repository:
+
+```bash
+git clone --recursive https://github.com/huggingface/peft
+```
+
+Navigate to the directory containing the training scripts for fine-tuning Dreambooth with BOFT:
+
+```bash
+cd peft/examples/boft_dreambooth
+```
+
+Set up your environment: install PEFT, and all the required libraries. At the time of writing this guide we recommend installing PEFT from source. The following environment setup should work on A100 and H100:
+
+```bash
+conda create --name peft python=3.10
+conda activate peft
+conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=11.8 -c pytorch -c nvidia
+conda install xformers -c xformers
+pip install -r requirements.txt
+pip install git+https://github.com/huggingface/peft
+```
+
+## Download the data
+
+[dreambooth](https://github.com/google/dreambooth) dataset should have been automatically cloned in the following structure when running the training script.
+
+```
+boft_dreambooth
+├── data
+│   ├── data_dir
+│   └── dreambooth
+│       └── data
+│           ├── backpack
+│           └── backpack_dog
+│           ...
+```
+
+You can also put your custom images into `boft_dreambooth/data/dreambooth`.
+
+## Finetune Dreambooth with BOFT
+
+```bash
+./train_dreambooth.sh
+```
+
+or using the following script arguments:
+
+```bash
+export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+export INSTANCE_DIR="path-to-instance-images"
+export CLASS_DIR="path-to-class-images"
+export OUTPUT_DIR="path-to-save-model"
+```
+
+Here:
+
+- `INSTANCE_DIR`: The directory containing the images that you intend to use for training your model.
+- `CLASS_DIR`: The directory containing class-specific images. In this example, we use prior preservation to avoid overfitting and language-drift. For prior preservation, you need other images of the same class as part of the training process. However, these images can be generated and the training script will save them to a local path you specify here.
+- `OUTPUT_DIR`: The destination folder for storing the trained model's weights.
+
+To learn more about DreamBooth fine-tuning with prior-preserving loss, check out the [Diffusers documentation](https://huggingface.co/docs/diffusers/training/dreambooth#finetuning-with-priorpreserving-loss).
+
+Launch the training script with `accelerate` and pass hyperparameters, as well as LoRa-specific arguments to it such as:
+
+- `use_boft`: Enables BOFT in the training script.
+- `boft_block_size`: the BOFT matrix block size across different layers, expressed in `int`. Smaller block size results in sparser update matrices with fewer trainable paramters. **Note**, please choose it to be dividable to most layer `in_features` dimension, e.g., 4, 8, 16. Also, you can only specify either `boft_block_size` or `boft_block_num`, but not both simultaneously, because `boft_block_size` x `boft_block_num` = layer dimension.
+- `boft_block_num`: the number of BOFT matrix blocks across different layers, expressed in `int`. Fewer blocks result in sparser update matrices with fewer trainable paramters. **Note**, please choose it to be dividable to most layer `in_features` dimension, e.g., 4, 8, 16. Also, you can only specify either `boft_block_size` or `boft_block_num`, but not both simultaneously, because `boft_block_size` x `boft_block_num` = layer dimension.
+- `boft_n_butterfly_factor`: the number of butterfly factors. **Note**, for `boft_n_butterfly_factor=1`, BOFT is the same as vanilla OFT, for `boft_n_butterfly_factor=2`, the effective block size of OFT becomes twice as big and the number of blocks become half.
+- `bias`: specify if the `bias` paramteres should be traind. Can be `none`, `all` or `boft_only`.
+- `boft_dropout`: specify the probability of multiplicative dropout.
+
+Here's what the full set of script arguments may look like:
+
+```bash
+PEFT_TYPE="boft"
+BLOCK_NUM=8
+BLOCK_SIZE=0
+N_BUTTERFLY_FACTOR=1
+
+VALIDATION_PROMPT=${PROMPT_LIST[@]}
+INSTANCE_PROMPT="a photo of ${UNIQUE_TOKEN} ${CLASS_TOKEN}"
+CLASS_PROMPT="a photo of ${CLASS_TOKEN}"
+
+export MODEL_NAME="stabilityai/stable-diffusion-2-1"
+# export MODEL_NAME="runwayml/stable-diffusion-v1-5"
+export PROJECT_NAME="dreambooth_${PEFT_TYPE}"
+export RUN_NAME="${SELECTED_SUBJECT}_${PEFT_TYPE}_${BLOCK_NUM}${BLOCK_SIZE}${N_BUTTERFLY_FACTOR}"
+export INSTANCE_DIR="./data/dreambooth/dataset/${SELECTED_SUBJECT}"
+export CLASS_DIR="./data/class_data/${CLASS_TOKEN}"
+export OUTPUT_DIR="./data/output/${PEFT_TYPE}"
+
+
+accelerate launch train_dreambooth.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --class_data_dir="$CLASS_DIR" \
+  --output_dir=$OUTPUT_DIR \
+  --wandb_project_name=$PROJECT_NAME \
+  --wandb_run_name=$RUN_NAME \
+  --with_prior_preservation --prior_loss_weight=1.0 \
+  --instance_prompt="$INSTANCE_PROMPT" \
+  --validation_prompt="$VALIDATION_PROMPT" \
+  --class_prompt="$CLASS_PROMPT" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --num_dataloader_workers=2 \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --num_class_images=200 \
+  --use_boft \
+  --boft_block_num=$BLOCK_NUM \
+  --boft_block_size=$BLOCK_SIZE \
+  --boft_n_butterfly_factor=$N_BUTTERFLY_FACTOR \
+  --boft_dropout=0.1 \
+  --boft_bias="boft_only" \
+  --learning_rate=3e-5 \
+  --max_train_steps=1010 \
+  --checkpointing_steps=200 \
+  --validation_steps=200 \
+  --enable_xformers_memory_efficient_attention \
+  --report_to="wandb" \
+```
+
+or use this training script:
+
+```bash
+./train_dreambooth.sh $idx
+```
+
+with the `$idx` corresponds to different subjects.
+
+If you are running this script on Windows, you may need to set the `--num_dataloader_workers` to 0.
+
+## Inference with a single adapter
+
+To run inference with the fine-tuned model, simply run the jupyter notebook `dreambooth_inference.ipynb` for visualization with `jupyter notebook` under `./examples/boft_dreambooth`.

examples/boft_dreambooth/dreambooth_inference.ipynb

@@ -0,0 +1,186 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "acab479f",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "\n",
+    "import torch\n",
+    "from accelerate.logging import get_logger\n",
+    "from diffusers import StableDiffusionPipeline\n",
+    "from diffusers.utils import check_min_version\n",
+    "\n",
+    "from peft import PeftModel\n",
+    "\n",
+    "# Will error if the minimal version of diffusers is not installed. Remove at your own risks.\n",
+    "check_min_version(\"0.10.0.dev0\")\n",
+    "\n",
+    "logger = get_logger(__name__)\n",
+    "\n",
+    "MODEL_NAME = \"stabilityai/stable-diffusion-2-1\"\n",
+    "# MODEL_NAME=\"runwayml/stable-diffusion-v1-5\"\n",
+    "\n",
+    "PEFT_TYPE=\"boft\"\n",
+    "BLOCK_NUM=8\n",
+    "BLOCK_SIZE=0\n",
+    "N_BUTTERFLY_FACTOR=1\n",
+    "SELECTED_SUBJECT=\"backpack\"\n",
+    "EPOCH_IDX = 200\n",
+    "\n",
+    "PROJECT_NAME=f\"dreambooth_{PEFT_TYPE}\"\n",
+    "RUN_NAME=f\"{SELECTED_SUBJECT}_{PEFT_TYPE}_{BLOCK_NUM}{BLOCK_SIZE}{N_BUTTERFLY_FACTOR}\"\n",
+    "OUTPUT_DIR=f\"./data/output/{PEFT_TYPE}\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "06cfd506",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def get_boft_sd_pipeline(\n",
+    "    ckpt_dir, base_model_name_or_path=None, epoch=int, dtype=torch.float32, device=\"cuda\", adapter_name=\"default\"\n",
+    "):\n",
+    "\n",
+    "    if base_model_name_or_path is None:\n",
+    "        raise ValueError(\"Please specify the base model name or path\")\n",
+    "\n",
+    "    pipe = StableDiffusionPipeline.from_pretrained(\n",
+    "        base_model_name_or_path, torch_dtype=dtype, requires_safety_checker=False\n",
+    "    ).to(device)\n",
+    "    \n",
+    "    load_adapter(pipe, ckpt_dir, epoch, adapter_name)\n",
+    "\n",
+    "    if dtype in (torch.float16, torch.bfloat16):\n",
+    "        pipe.unet.half()\n",
+    "        pipe.text_encoder.half()\n",
+    "\n",
+    "    pipe.to(device)\n",
+    "    return pipe\n",
+    "\n",
+    "\n",
+    "def load_adapter(pipe, ckpt_dir, epoch, adapter_name=\"default\"):\n",
+    "    \n",
+    "    unet_sub_dir = os.path.join(ckpt_dir, f\"unet/{epoch}\", adapter_name)\n",
+    "    text_encoder_sub_dir = os.path.join(ckpt_dir, f\"text_encoder/{epoch}\", adapter_name)\n",
+    "    \n",
+    "    if isinstance(pipe.unet, PeftModel):\n",
+    "        pipe.unet.load_adapter(unet_sub_dir, adapter_name=adapter_name)\n",
+    "    else:\n",
+    "        pipe.unet = PeftModel.from_pretrained(pipe.unet, unet_sub_dir, adapter_name=adapter_name)\n",
+    "        \n",
+    "    if os.path.exists(text_encoder_sub_dir):\n",
+    "        if isinstance(pipe.text_encoder, PeftModel):\n",
+    "            pipe.text_encoder.load_adapter(text_encoder_sub_dir, adapter_name=adapter_name)\n",
+    "        else:\n",
+    "            pipe.text_encoder = PeftModel.from_pretrained(pipe.text_encoder, text_encoder_sub_dir, adapter_name=adapter_name)\n",
+    "    \n",
+    "\n",
+    "def set_adapter(pipe, adapter_name):\n",
+    "    pipe.unet.set_adapter(adapter_name)\n",
+    "    if isinstance(pipe.text_encoder, PeftModel):\n",
+    "        pipe.text_encoder.set_adapter(adapter_name)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "98a0d8ac",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "prompt = \"a photo of sks backpack on a wooden floor\"\n",
+    "negative_prompt = \"low quality, blurry, unfinished\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d4e888d2",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%time\n",
+    "pipe = get_boft_sd_pipeline(OUTPUT_DIR, MODEL_NAME, EPOCH_IDX, adapter_name=RUN_NAME)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f1c1a1c0",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%time\n",
+    "image = pipe(prompt, num_inference_steps=50, guidance_scale=7, negative_prompt=negative_prompt).images[0]\n",
+    "image"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "3a1aafdf-8cf7-4e47-9471-26478034245e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# load and reset another adapter\n",
+    "# WARNING: requires training DreamBooth with `boft_bias=None`\n",
+    "\n",
+    "SELECTED_SUBJECT=\"dog\"\n",
+    "EPOCH_IDX = 200\n",
+    "RUN_NAME=f\"{SELECTED_SUBJECT}_{PEFT_TYPE}_{BLOCK_NUM}{BLOCK_SIZE}{N_BUTTERFLY_FACTOR}\"\n",
+    "\n",
+    "load_adapter(pipe, OUTPUT_DIR, epoch=EPOCH_IDX, adapter_name=RUN_NAME)\n",
+    "set_adapter(pipe, adapter_name=RUN_NAME)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c7091ad0-2005-4528-afc1-4f9d70a9a535",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%time\n",
+    "prompt = \"a photo of sks dog running on the beach\"\n",
+    "negative_prompt = \"low quality, blurry, unfinished\"\n",
+    "image = pipe(prompt, num_inference_steps=50, guidance_scale=7, negative_prompt=negative_prompt).images[0]\n",
+    "image"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f534eca2-94a4-432b-b092-7149ac44b12f",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python [conda env:peft] *",
+   "language": "python",
+   "name": "conda-env-peft-py"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.13"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}