Release: v1.3.0

This fixes tests/test_data_loader.py::StatefulDataLoaderTester tests which
started to fail after 828aae4:
```
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_dataloader_dispatcher_state_dict_num_workers_0 - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_dataloader_dispatcher_state_dict_num_workers_2 - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_dataloader_inheritance - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_dataloader_state_dict_num_workers_0 - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_dataloader_state_dict_num_workers_2 - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_decoupled_stateful_dataloader_adapter_equivalent_to_torchdata_stateful_dataloader_num_workers_0 - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_decoupled_stateful_dataloader_adapter_equivalent_to_torchdata_stateful_dataloader_num_workers_2 - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_end_of_dataloader - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_end_of_dataloader_dispatcher - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_skip_data_loader - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_stateful_dataloader_adapter_equivalent_to_torchdata_stateful_dataloader_num_workers_0 - KeyError: 'in_order'
FAILED tests/test_data_loader.py::StatefulDataLoaderTester::test_stateful_dataloader_adapter_equivalent_to_torchdata_stateful_dataloader_num_workers_2 - KeyError: 'in_order'
```

The reason for the failure is that "in_order" is added only if data loader
is created with `prepare_data_loader` or `skip_first_batches()`. Tests in
`tests/test_data_loader.py::StatefulDataLoaderTester` however are creating
data loaders directly as classes and "in_order" was not added. Hence the
issue.

Fixes: 828aae4 ("add torchdata version check to avoid in_order error (#3344)")

Signed-off-by: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>

.github/PULL_REQUEST_TEMPLATE.md

@@ -37,11 +37,11 @@ members/contributors who may be interested in your PR.
  If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**.
 
 - Big modeling: @SunMarc
-- Fully-Sharded Data Parallism: @pacman100
-- DeepSpeed: @pacman100
+- Fully-Sharded Data Parallism: @muellerzr
+- DeepSpeed: @muellerzr
 - Command Line Interface: @muellerzr
 - Documentation: @muellerzr
-- Core parts of the library: @muellerzr @BenjaminBossan
-- Maintained examples: @muellerzr or @pacman100
+- Core parts of the library: @muellerzr @BenjaminBossan @SunMarc
+- Maintained examples: @muellerzr or @SunMarc
 
  -->

.github/workflows/build-docker-images-release.yml

@@ -15,13 +15,14 @@ jobs:
     outputs:
       version: ${{ steps.step1.outputs.version }}
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v3.1.0
       - id: step1
         run: echo "version=$(python setup.py --version)" >> $GITHUB_OUTPUT
 
   version-cpu:
     name: "Latest Accelerate CPU [version]"
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
+    runs-on:
+      group: aws-general-8-plus
     needs: get-version
     steps:
       - name: Set up Docker Buildx
@@ -37,11 +38,12 @@ jobs:
         with:
           file: docker/accelerate-cpu/Dockerfile
           push: true
-          tags: huggingface/accelerate-cpu:${{needs.get-version.outputs.version}}
+          tags: huggingface/accelerate:cpu-release-${{ needs.get-version.outputs.version }}
 
   version-cuda:
     name: "Latest Accelerate GPU [version]"
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
+    runs-on:
+      group: aws-g6-4xlarge-plus
     needs: get-version
     steps:
       - name: Set up Docker Buildx
@@ -57,4 +59,46 @@ jobs:
         with:
           file: docker/accelerate-gpu/Dockerfile
           push: true
-          tags: huggingface/accelerate-gpu:${{needs.get-version.outputs.version}}
+          tags: huggingface/accelerate:gpu-release-${{needs.get-version.outputs.version}}
+
+  version-cuda-deepspeed:
+    name: "Latest Accelerate GPU DeepSpeed [version]"
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    needs: get-version
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v2
+      - name: Login to DockerHub
+        uses: docker/login-action@v2
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v4
+        with:
+          file: docker/accelerate-gpu-deepspeed/Dockerfile
+          push: true
+          tags: huggingface/accelerate:gpu-deepspeed-release-${{needs.get-version.outputs.version}}
+
+  version-cuda-fp8-transformerengine:
+    name: "Latest Accelerate GPU FP8 TransformerEngine [version]"
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    needs: get-version
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v2
+      - name: Login to DockerHub
+        uses: docker/login-action@v2
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v4
+        with:
+          file: docker/accelerate-gpu/Dockerfile
+          push: true
+          tags: huggingface/accelerate:gpu-fp8-transformerengine-release-${{needs.get-version.outputs.version}}

.github/workflows/build_and_run_tests.yml

@@ -22,7 +22,7 @@ jobs:
       
       - name: Get changed files
         id: changed-files
-        uses: tj-actions/changed-files@v22.2
+        uses: tj-actions/changed-files@v41
       
       - name: Was setup changed 
         id: was_changed
@@ -45,6 +45,6 @@ jobs:
     uses: ./.github/workflows/run_merge_tests.yml
 
   run-integration-tests:
-    needs: run-merge-tests
+    needs: build-docker-containers
     if: always()
     uses: ./.github/workflows/self_hosted_integration_tests.yml

.github/workflows/build_docker_images.yml

@@ -11,19 +11,10 @@ concurrency:
   cancel-in-progress: false
 
 jobs:
-  clean-storage:
-    name: "Clean docker image storage"
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
-    steps:
-      - name: Clean storage
-        run: |
-          docker image prune --all -f --filter "until=48h"
-          docker system prune --all -f --filter "until=48h"
-
   latest-cpu:
     name: "Latest Accelerate CPU [dev]"
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
-    needs: clean-storage
+    runs-on:
+      group: aws-general-8-plus
     steps:
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v2
@@ -32,17 +23,23 @@ jobs:
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
+      - name: Get current date
+        id: date
+        run: |
+          echo "date=$(date '+%Y-%m-%d')" >> $GITHUB_ENV
       - name: Build and Push CPU
         uses: docker/build-push-action@v4
         with:
-          file: docker/accelerate-cpu/Dockerfile          
+          file: docker/accelerate-cpu/Dockerfile
           push: true
-          tags: huggingface/accelerate-cpu
+          tags: |
+            huggingface/accelerate:cpu-nightly
+            huggingface/accelerate:cpu-nightly-${{ env.date }}
 
   latest-cuda:
     name: "Latest Accelerate GPU [dev]"
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
-    needs: clean-storage
+    runs-on:
+      group: aws-g6-4xlarge-plus
     steps:
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v2
@@ -51,10 +48,63 @@ jobs:
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
-
+      - name: Get current date
+        id: date
+        run: |
+          echo "date=$(date '+%Y-%m-%d')" >> $GITHUB_ENV
       - name: Build and Push GPU
         uses: docker/build-push-action@v4
         with:
-          file: docker/accelerate-gpu/Dockerfile          
+          file: docker/accelerate-gpu/Dockerfile
           push: true
-          tags: huggingface/accelerate-gpu
+          tags: |
+            huggingface/accelerate:gpu-nightly
+            huggingface/accelerate:gpu-nightly-${{ env.date }}
+
+  latest-cuda-deepspeed:
+    name: "Latest Accelerate GPU DeepSpeed [dev]"
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v2
+      - name: Login to DockerHub
+        uses: docker/login-action@v2
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+      - name: Get current date
+        id: date
+        run: |
+          echo "date=$(date '+%Y-%m-%d')" >> $GITHUB_ENV
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v4
+        with:
+          file: docker/accelerate-gpu-deepspeed/Dockerfile
+          push: true
+          tags: |
+            huggingface/accelerate:gpu-deepspeed-nightly
+            huggingface/accelerate:gpu-deepspeed-nightly-${{ env.date }}
+
+  latest-cuda-fp8-transformerengine:
+    name: "Latest Accelerate GPU FP8 TransformerEngine [dev]"
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v2
+      - name: Login to DockerHub
+        uses: docker/login-action@v2
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+      - name: Get current date
+        id: date
+        run: |
+          echo "date=$(date '+%Y-%m-%d')" >> $GITHUB_ENV
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v4
+        with:
+          file: benchmarks/fp8/transformer_engine/Dockerfile
+          push: true
+          tags: huggingface/accelerate:gpu-fp8-transformerengine-nightly-${{ env.date }}

.github/workflows/build_documentation.yml

@@ -13,6 +13,6 @@ jobs:
     with:
       commit_sha: ${{ github.sha }}
       package: accelerate
+      custom_container: huggingface/transformers-doc-builder
     secrets:
-      token: ${{ secrets.HUGGINGFACE_PUSH }}
       hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/build_pr_documentation.yml

@@ -14,3 +14,4 @@ jobs:
       commit_sha: ${{ github.event.pull_request.head.sha }}
       pr_number: ${{ github.event.number }}
       package: accelerate
+      custom_container: huggingface/transformers-doc-builder

.github/workflows/delete_doc_comment.yml

@@ -1,14 +0,0 @@
-name: Delete doc comment
-
-on:
-  workflow_run:
-    workflows: ["Delete doc comment trigger"]
-    types:
-      - completed
-
-
-jobs:
-  delete:
-    uses: huggingface/doc-builder/.github/workflows/delete_doc_comment.yml@main
-    secrets:
-      comment_bot_token: ${{ secrets.COMMENT_BOT_TOKEN }}

.github/workflows/delete_doc_comment_trigger.yml

@@ -1,12 +0,0 @@
-name: Delete doc comment trigger
-
-on:
-  pull_request:
-    types: [ closed ]
-
-
-jobs:
-  delete:
-    uses: huggingface/doc-builder/.github/workflows/delete_doc_comment_trigger.yml@main
-    with:
-      pr_number: ${{ github.event.number }}

.github/workflows/integration_tests.yml

@@ -25,17 +25,14 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       fail-fast: false
-      matrix:
-        transformers-version: [
-          pypi,
-          github
-        ]
     steps:
     - uses: actions/checkout@v3.1.0
-    - name: Set up python 3.8
+    - name: Set up python 3.9
       uses: actions/setup-python@v3
       with:
-        python-version: 3.8
+        python-version: 3.9
+        cache: 'pip'
+        cache-dependency-path: 'setup.py'
 
     - name: Install Accelerate from source
       run: |
@@ -47,9 +44,6 @@ jobs:
         cd ..
         git clone https://github.com/huggingface/transformers
         cd transformers
-        if [[ ${{ matrix.transformers-version }} = pypi ]]; then 
-          git checkout $(git describe --tags `git rev-list --tags --max-count=1`)
-        fi
         pip install .[torch,testing]
 
     - name: Show installed libraries

.github/workflows/nightly.yml

@@ -12,67 +12,138 @@ env:
 
 
 jobs:
-  run_all_tests_single_gpu:
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
+  run_core_tests_single_gpu:
+    runs-on:
+      group: aws-g6-4xlarge-plus
     env:
       CUDA_VISIBLE_DEVICES: "0"
       TEST_TYPE: "single_gpu"
     container:
-      image: huggingface/accelerate-gpu:latest
+      image: huggingface/accelerate:gpu-nightly
       options: --gpus all --shm-size "16gb"
     defaults:
       run:
-        working-directory: accelerate/
         shell: bash
     steps:
       - name: Update clone & pip install
         run: |
           source activate accelerate
-          git config --global --add safe.directory '*'
-          git fetch && git checkout ${{ github.sha }} 
+          git clone https://github.com/huggingface/accelerate;
+          cd accelerate;
+          git checkout ${{ github.sha }};
           pip install -e . --no-deps
           pip install pytest-reportlog tabulate
 
+      - name: Show installed libraries
+        run: |
+          source activate accelerate;
+          pip freeze
+
       - name: Run test on GPUs
+        working-directory: accelerate
         run: |
           source activate accelerate
           make test
-          
+
       - name: Run examples on GPUs
+        working-directory: accelerate
         if: always()
         run: |
           source activate accelerate
           pip uninstall comet_ml -y
           make test_examples
-          
+
       - name: Generate Report
+        working-directory: accelerate
         if: always()
         run: |
           pip install slack_sdk tabulate
           python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
-  run_all_tests_multi_gpu:
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
+  run_deepspeed_tests_single_gpu:
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      CUDA_VISIBLE_DEVICES: "0"
+      TEST_TYPE: "single_gpu_deepspeed"
+    container:
+      image: huggingface/accelerate:gpu-deepspeed-nightly
+      options: --gpus all --shm-size "16gb"
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - name: Update clone & pip install
+        run: |
+          source activate accelerate
+          git clone https://github.com/huggingface/accelerate;
+          cd accelerate;
+          git checkout ${{ github.sha }};
+          pip install -e . --no-deps
+          pip install pytest-reportlog tabulate
+
+      - name: Show installed libraries
+        run: |
+          source activate accelerate;
+          pip freeze
+
+      - name: Run test on GPUs
+        working-directory: accelerate
+        run: |
+          source activate accelerate
+          make test_deepspeed
+
+      - name: Run Integration tests on GPUs
+        working-directory: accelerate
+        if: always()
+        run: |
+          source activate accelerate
+          make test_integrations
+
+      - name: Run examples on GPUs
+        working-directory: accelerate
+        if: always()
+        run: |
+          source activate accelerate
+          pip uninstall comet_ml -y
+          make test_examples
+
+      - name: Generate Report
+        working-directory: accelerate
+        if: always()
+        run: |
+          pip install slack_sdk tabulate
+          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+
+  run_core_tests_multi_gpu:
+    runs-on:
+      group: aws-g6-12xlarge-plus
     env:
       CUDA_VISIBLE_DEVICES: "0,1"
       TEST_TYPE: "multi_gpu"
     container:
-      image: huggingface/accelerate-gpu:latest
+      image: huggingface/accelerate:gpu-nightly
       options: --gpus all --shm-size "16gb"
     defaults:
       run:
-        working-directory: accelerate/
         shell: bash
     steps:
       - name: Update clone
         run: |
           source activate accelerate
-          git config --global --add safe.directory '*'
-          git fetch && git checkout ${{ github.sha }}
+          git clone https://github.com/huggingface/accelerate;
+          cd accelerate;
+          git checkout ${{ github.sha }};
           pip install -e . --no-deps
           pip install pytest-reportlog tabulate
 
+      - name: Show installed libraries
+        run: |
+          source activate accelerate;
+          pip freeze
+
       - name: Run core and big modeling tests on GPUs
+        working-directory: accelerate
         run: |
           source activate accelerate
           make test_core
@@ -80,12 +151,14 @@ jobs:
           make test_cli
 
       - name: Run Integration tests on GPUs
+        working-directory: accelerate
         if: always()
         run: |
           source activate accelerate
           make test_integrations
 
       - name: Run examples on GPUs
+        working-directory: accelerate
         if: always()
         run: |
           source activate accelerate
@@ -93,13 +166,68 @@ jobs:
           make test_examples
 
       - name: Generate Report
+        working-directory: accelerate
         if: always()
         run: |
           pip install slack_sdk tabulate
           python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
-  
+  run_deepspeed_tests_multi_gpu:
+    runs-on:
+      group: aws-g6-12xlarge-plus
+    env:
+      CUDA_VISIBLE_DEVICES: "0,1"
+      TEST_TYPE: "multi_gpu_deepspeed"
+    container:
+      image: huggingface/accelerate:gpu-deepspeed-nightly
+      options: --gpus all --shm-size "16gb"
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - name: Update clone
+        run: |
+          source activate accelerate
+          git clone https://github.com/huggingface/accelerate;
+          cd accelerate;
+          git checkout ${{ github.sha }};
+          pip install -e . --no-deps
+          pip install pytest-reportlog tabulate
+
+      - name: Show installed libraries
+        run: |
+          source activate accelerate;
+          pip freeze
+
+      - name: Run DeepSpeed tests
+        working-directory: accelerate
+        run: |
+          source activate accelerate
+          make test_deepspeed
+
+      - name: Run Integration tests on GPUs
+        working-directory: accelerate
+        if: always()
+        run: |
+          source activate accelerate
+          make test_integrations
+
+      - name: Run examples on GPUs
+        working-directory: accelerate
+        if: always()
+        run: |
+          source activate accelerate
+          pip uninstall comet_ml -y
+          make test_examples
+
+      - name: Generate Report
+        working-directory: accelerate
+        if: always()
+        run: |
+          pip install slack_sdk tabulate
+          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+
+
   run-integration-tests:
-    needs: [run_all_tests_single_gpu, run_all_tests_multi_gpu]
     if: always()
-    uses: ./.github/workflows/self_hosted_integration_tests.yml
+    uses: ./.github/workflows/self_hosted_integration_tests.yml

.github/workflows/quality.yml

@@ -6,11 +6,13 @@ jobs:
   quality:
     runs-on: ubuntu-latest
     steps:
-    - uses: actions/checkout@v2
-    - name: Set up Python 3.8
+    - uses: actions/checkout@v3.1.0
+    - name: Set up Python 3.9
       uses: actions/setup-python@v3
       with:
-        python-version: 3.8
+        python-version: 3.9
+        cache: 'pip'
+        cache-dependency-path: 'setup.py'
     - name: Install Python dependencies
       run: pip install -e .[quality]
     - name: Run Quality check

.github/workflows/run_merge_tests.yml

@@ -9,81 +9,180 @@ env:
   IS_GITHUB_CI: "1"
 
 jobs:
-  run_all_tests_single_gpu:
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
+  run_core_tests_single_gpu:
+    runs-on:
+      group: aws-g6-4xlarge-plus
     env:
       CUDA_VISIBLE_DEVICES: "0"
     container:
-      image: huggingface/accelerate-gpu:latest
+      image: huggingface/accelerate:gpu-nightly
       options: --gpus all --shm-size "16gb"
     defaults:
       run:
-        working-directory: accelerate/
         shell: bash
     steps:
-      - name: Update clone & pip install
+      - name: Install accelerate
         run: |
-          source activate accelerate
-          git config --global --add safe.directory '*'
-          git fetch && git checkout ${{ github.sha }}
-          pip install -e .[testing,test_trackers] -U
-          pip install pytest-reportlog tabulate
+          source activate accelerate;
+          git clone https://github.com/huggingface/accelerate;
+          cd accelerate;
+          git checkout ${{ github.sha }};
+          pip install -e .[testing,test_trackers] -U;
+          pip install pytest-reportlog tabulate  ;
 
-      - name: Run CLI tests
+      - name: Show installed libraries
         run: |
-          source activate accelerate
+          source activate accelerate;
+          pip freeze
+
+      - name: Run CLI tests (use make cli)
+        working-directory: accelerate
+        run: |
+          source activate accelerate;
           make test_cli
-          
+
       - name: Run test on GPUs
+        working-directory: accelerate
         if: always()
         run: |
-          source activate accelerate
+          source activate accelerate;
           make test
       - name: Run examples on GPUs
+        working-directory: accelerate
         if: always()
         run: |
-          source activate accelerate
-          pip uninstall comet_ml -y
+          source activate accelerate;
+          pip uninstall comet_ml -y;
           make test_examples
 
       - name: Generate Report
+        working-directory: accelerate
         if: always()
         run: |
-          pip install tabulate
+          pip install tabulate;
           python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
-  run_all_tests_multi_gpu:
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
+  run_deepspeed_tests_single_gpu:
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      CUDA_VISIBLE_DEVICES: "0"
     container:
-      image: huggingface/accelerate-gpu:latest
+      image: huggingface/accelerate:gpu-deepspeed-nightly
+      options: --gpus all --shm-size "16gb"
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - name: Install accelerate
+        run: |
+          source activate accelerate;
+          git clone https://github.com/huggingface/accelerate;
+          cd accelerate;
+          git checkout ${{ github.sha }};
+          pip install -e .[testing,test_trackers] -U;
+          pip install pytest-reportlog tabulate  ;
+
+      - name: Show installed libraries
+        run: |
+          source activate accelerate;
+          pip freeze
+
+      - name: Run test on GPUs
+        working-directory: accelerate
+        if: always()
+        run: |
+          source activate accelerate;
+          make test_deepspeed
+
+      - name: Generate Report
+        working-directory: accelerate
+        if: always()
+        run: |
+          pip install tabulate;
+          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+
+  run_core_tests_multi_gpu:
+    runs-on:
+      group: aws-g6-12xlarge-plus
+    env:
+      CUDA_VISIBLE_DEVICES: 0,1
+    container:
+      image: huggingface/accelerate:gpu-nightly
       options: --gpus all --shm-size "16gb"
     defaults:
       run:
-        working-directory: accelerate/
         shell: bash
     steps:
       - name: Update clone
         run: |
-          source activate accelerate
-          git config --global --add safe.directory '*'
-          git fetch && git checkout ${{ github.sha }}
-          pip install -e .[testing,test_trackers] -U
+          source activate accelerate;
+          git clone https://github.com/huggingface/accelerate;
+          cd accelerate;
+          git checkout ${{ github.sha }};
+          pip install -e .[testing,test_trackers] -U;
           pip install pytest-reportlog tabulate
 
-      - name: Run test on GPUs
+      - name: Show installed libraries
         run: |
-          source activate accelerate
+          source activate accelerate;
+          pip freeze
+
+      - name: Run test on GPUs
+        working-directory: accelerate
+        run: |
+          source activate accelerate;
           make test
 
       - name: Run examples on GPUs
+        working-directory: accelerate
         if: always()
         run: |
-          source activate accelerate
-          pip uninstall comet_ml -y
+          source activate accelerate;
+          pip uninstall comet_ml -y;
           make test_examples
 
       - name: Generate Report
+        working-directory: accelerate
         if: always()
         run: |
-          pip install tabulate
-          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+          source activate accelerate;
+          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+
+  run_deepspeed_tests_multi_gpu:
+    runs-on:
+      group: aws-g6-12xlarge-plus
+    container:
+      image: huggingface/accelerate:gpu-deepspeed-nightly
+      options: --gpus all --shm-size "16gb"
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - name: Install accelerate
+        run: |
+          source activate accelerate;
+          git clone https://github.com/huggingface/accelerate;
+          cd accelerate;
+          git checkout ${{ github.sha }};
+          pip install -e .[testing,test_trackers] -U;
+          pip install pytest-reportlog tabulate  ;
+
+      - name: Show installed libraries
+        run: |
+          source activate accelerate;
+          pip freeze
+
+      - name: Run test on GPUs
+        working-directory: accelerate
+        if: always()
+        run: |
+          source activate accelerate;
+          make test_deepspeed
+
+      - name: Generate Report
+        working-directory: accelerate
+        if: always()
+        run: |
+          pip install tabulate;
+          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY

.github/workflows/self_hosted_integration_tests.yml

@@ -1,7 +1,7 @@
 # CI for specifically ensuring integrations work fine (`transformers` mainly) on GPUs
 # Useful tips:
 #  - `working-directory` should be set to the root of the repo, which is cloned on the actual CI runner.
-#    It follows the directory structure of `actions-runner/_work/{repo_name}/{repo_name}/{cloned_repo} on 
+#    It follows the directory structure of `actions-runner/_work/{repo_name}/{repo_name}/{cloned_repo} on
 #    prem, but in Actions setting `working-directory` looks just in the `{repo_name}` level.
 #  - New integrations to test should have its own job, and follow a strategy method where we check both
 #    the pypi and github versions.
@@ -23,40 +23,36 @@ defaults:
 jobs:
   run-trainer-tests:
     container:
-      image: huggingface/accelerate-gpu:latest
+      image: huggingface/accelerate:gpu-deepspeed-nightly
       options: --gpus all --shm-size "16gb"
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
+    runs-on:
+      group: aws-g6-12xlarge-plus
     strategy:
       fail-fast: false
       matrix:
-        transformers-version: [
-          pypi,
-          github
-        ]
         cuda_visible_devices: [
-          "0", 
+          "0",
           "0,1"
         ]
     steps:
-      - name: Update accelerate clone and pip install
-        working-directory: accelerate/
-        run: 
-          source activate accelerate;
-          git config --global --add safe.directory '*';
-          git checkout main && git fetch && git checkout ${{ github.sha }};
-          pip install -e .;
-
-      - name: Update transformers clone & pip install
-        working-directory: transformers/
+      - name: Install transformers
         run: |
-          source activate accelerate
-          git config --global --add safe.directory '*'
-          git checkout main && git pull
-          if [[ ${{ matrix.transformers-version }} = pypi ]]; then 
-            git checkout $(git describe --tags `git rev-list --tags --max-count=1`)
-          fi
-          pip install .[torch,deepspeed-testing]
-      
+          source activate accelerate;
+          git clone https://github.com/huggingface/transformers --depth 1;
+          cd transformers;
+          pip install .[torch,deepspeed-testing];
+          cd ..;
+
+      - name: Install accelerate
+        run: |
+          source activate accelerate;
+          git clone https://github.com/huggingface/accelerate;
+          cd accelerate;
+          git checkout ${{ github.sha }} ;
+          pip install -e .[testing];
+          pip uninstall comet_ml wandb dvclive -y
+          cd ..;
+
       - name: Show installed libraries
         run: |
           source activate accelerate;
@@ -81,36 +77,41 @@ jobs:
           source activate accelerate;
           pytest -sv tests/deepspeed
 
-  run-skorch-tests:
-    container:
-      image: huggingface/accelerate-gpu:latest
-      options: --gpus all --shm-size "16gb"
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
-    strategy:
-      fail-fast: false
-      matrix:
-        skorch-version: [
-          pypi,
-          github
-        ]
-    steps:
-      - name: Update accelerate clone and pip install
-        working-directory: accelerate/
-        run: 
-          source activate accelerate;
-          git config --global --add safe.directory '*';
-          git checkout main && git fetch && git checkout ${{ github.sha }};
-          pip install -e .;
-
-      - name: Update skorch clone & pip install
-        working-directory: skorch/
+      - name: Run transformers examples tests
+        working-directory: transformers/
+        env:
+          CUDA_VISIBLE_DEVICES: ${{ matrix.cuda_visible_devices }}
+          WANDB_DISABLED: true
         run: |
           source activate accelerate
+          pip install -r examples/pytorch/_tests_requirements.txt
+          pytest -sv examples/pytorch/test_accelerate_examples.py examples/pytorch/test_pytorch_examples.py
+
+  run-skorch-tests:
+    container:
+      image: huggingface/accelerate:gpu-nightly
+      options: --gpus all --shm-size "16gb"
+    runs-on:
+      group: aws-g6-12xlarge-plus
+    strategy:
+      fail-fast: false
+    steps:
+      - name: Install accelerate
+        run:
+          source activate accelerate;
+          git clone https://github.com/huggingface/accelerate;
+          cd accelerate;
+          git checkout ${{ github.sha }};
+          pip install -e .[testing];
+          cd ..
+
+      - name: Install skorch
+        run: |
+          source activate accelerate
+          git clone https://github.com/skorch-dev/skorch;
+          cd skorch;
           git config --global --add safe.directory '*'
           git checkout master && git pull
-          if [[ ${{ matrix.skorch-version }} = pypi ]]; then 
-            git checkout $(git describe --tags `git rev-list --tags --max-count=1`)
-          fi
           pip install .[testing]
           pip install flaky
 
@@ -123,4 +124,4 @@ jobs:
         working-directory: skorch/
         run: |
           source activate accelerate;
-          pytest -sv -k TestAccelerate
+          pytest -sv -k TestAccelerate

.github/workflows/stale.yml

@@ -10,19 +10,24 @@ jobs:
     name: Close Stale Issues
     if: github.repository == 'huggingface/accelerate'
     runs-on: ubuntu-latest
+    permissions:
+      issues: write
+      pull-requests: write
     env:
       GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
     steps:
-    - uses: actions/checkout@v2
+    - uses: actions/checkout@v3.1.0
     
     - name: Setup Python
-      uses: actions/setup-python@v1
+      uses: actions/setup-python@v3
       with:
-        python-version: 3.8
+        python-version: 3.9
+        cache: 'pip'
+        cache-dependency-path: 'setup.py'
     
     - name: Install requirements
       run: |
         pip install PyGithub
     - name: Close stale issues
       run: |
-        python utils/stale.py
+        python utils/stale.py

.github/workflows/test.yml

@@ -39,27 +39,24 @@ jobs:
         ]
     steps:
     - uses: actions/checkout@v3.1.0
-    - name: Set up python 3.8
+    - name: Set up python 3.9
       uses: actions/setup-python@v3
       with:
-        python-version: 3.8
-    
-    - name: Activate python cache
-      uses: actions/cache@v3
-      with:
-        path: |
-          ${{ env.pythonLocation }}
-          ${{ env.HF_HOME }}
-        key: ${{ env.pythonLocation }}-${{ matrix.pytorch-version }}-${{ matrix.test-kind }}-${{ hashFiles('setup.py') }}
+        python-version: 3.9
+        cache: 'pip'
+        cache-dependency-path: 'setup.py'
     
     - name: Install the library
       run: |
-        pip install --upgrade pip
         if [[ ${{ matrix.test-kind }} = test_prod ]]; then pip install -e .[test_prod]; fi
         if [[ ${{ matrix.test-kind }} != test_prod ]]; then pip install -e .[testing,test_trackers]; fi
         if [[ ${{ matrix.test-kind }} = test_rest ]]; then pip uninstall comet_ml -y; fi
-        if [[ ${{ matrix.test-kind }} = minimum ]]; then pip install torch==1.10.0; fi
-        pip install pytest-reportlog tabulate
+        if [[ ${{ matrix.pytorch-version }} = minimum ]]; then pip install torchvision==0.18.1 torch==2.3.1; fi
+        pip install pytest-reportlog tabulate setuptools
+
+    - name: Show installed libraries
+      run: |
+        pip freeze
     
     - name: Run Tests
       env: 
@@ -70,4 +67,4 @@ jobs:
     - name: Generate Report
       if: always()
       run: |
-        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY

.github/workflows/test_imports.yml

@@ -0,0 +1,55 @@
+name: Run Import Tests
+
+on:
+  pull_request:
+    paths:
+      - "src/**"
+      - "tests/**"
+      - ".github/**"
+      - "examples/**"
+      - "setup.py"
+    types: [opened, synchronize, reopened]
+
+env:
+  HF_HOME: ~/hf_cache
+  TESTING_MOCKED_DATALOADERS: "1"
+  IS_GITHUB_CI: "1"
+
+jobs:
+  run-tests:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        pytorch-version: [
+          latest,
+          minimum,
+        ]
+    steps:
+    - uses: actions/checkout@v3.1.0
+    - name: Set up python 3.9
+      uses: actions/setup-python@v3
+      with:
+        python-version: 3.9
+        cache: 'pip'
+        cache-dependency-path: 'setup.py'
+    
+    - name: Install the library
+      run: |
+        pip install -e .
+        pip install pytest-reportlog tabulate setuptools git+https://github.com/muellerzr/import-timer
+
+    - name: Show installed libraries
+      run: |
+        pip freeze
+    
+    - name: Run Import Tests
+      env: 
+        PYTORCH_VERSION: ${{ matrix.pytorch-version }}
+      run: |
+        pytest -sv tests/test_imports.py
+
+    - name: Generate Report
+      if: always()
+      run: |
+        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY

.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

.pre-commit-config.yaml

@@ -0,0 +1,13 @@
+repos:
+  - repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.2.1
+    hooks:
+      - id: ruff
+        args:
+          - --fix
+      - id: ruff-format
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v4.5.0
+    hooks:
+      - id: check-merge-conflict
+      - id: check-yaml

CONTRIBUTING.md

@@ -123,12 +123,15 @@ Follow these steps to start contributing:
 4. Set up a development environment by running the following command in a conda or a virtual environment you've created for working on this library:
 
    ```bash
-   $ pip install -e ".[quality]"
+   $ pip install -e ".[dev]"
    ```
+   
+   This will install all testing and linting/code quality dependencies for the library (see `quality`, `test_dev`, 
+   `test_prod` targets in [`setup.py`](./setup.py)).
 
    (If accelerate was already installed in the virtual environment, remove
    it with `pip uninstall accelerate` before reinstalling it in editable
-   mode with the `-e` flag.)
+   mode with the `-e` flag).
 
    Alternatively, if you are using [Visual Studio Code](https://code.visualstudio.com/Download), the fastest way to get set up is by using
    the provided Dev Container. Documentation on how to get started with dev containers is available [here](https://code.visualstudio.com/docs/remote/containers).
@@ -152,7 +155,7 @@ Follow these steps to start contributing:
    $ make test
    ```
 
-   `accelerate` relies on `black` and `ruff` to format its source code
+   `accelerate` relies on `ruff` to format its source code
    consistently. After you make changes, apply automatic style corrections and code verifications
    that can't be automated in one go with:
 
@@ -172,6 +175,14 @@ Follow these steps to start contributing:
    $ make quality
    ```
 
+   You can also set up [`pre-commit`](https://pre-commit.com/) to run these checks
+   automatically as Git commit hooks.
+
+   ```bash
+   $ pip install pre-commit
+   $ pre-commit install
+   ```
+
    Once you're happy with your changes, add changed files using `git add` and
    make a commit with `git commit` to record your changes locally:
 
@@ -235,4 +246,4 @@ $ python -m pytest -sv ./tests
 In fact, that's how `make test` is implemented (sans the `pip install` line)!
 
 You can specify a smaller set of tests in order to test only the feature
-you're working on.
+you're working on.

Makefile

@@ -1,6 +1,6 @@
 .PHONY: quality style test docs utils
 
-check_dirs := tests src examples benchmarks utils
+check_dirs := .
 
 # Check that source code meets quality standards
 
@@ -12,20 +12,17 @@ extra_quality_checks:
 
 # this target runs checks on all files
 quality:
-	black --required-version 23 --check $(check_dirs)
-	ruff $(check_dirs)
+	ruff check $(check_dirs)
+	ruff format --check $(check_dirs)
 	doc-builder style src/accelerate 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 --required-version 23 $(check_dirs)
-	ruff $(check_dirs) --fix
+	ruff check $(check_dirs) --fix
+	ruff format $(check_dirs)
 	doc-builder style src/accelerate docs/source --max_len 119
 	
 # Run tests for the library
-test:
-	python -m pytest -s -v ./tests/ --ignore=./tests/test_examples.py $(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_all.log",)
-
 test_big_modeling:
 	python -m pytest -s -v ./tests/test_big_modeling.py ./tests/test_modeling_utils.py $(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_big_modeling.log",)
 
@@ -42,6 +39,15 @@ test_deepspeed:
 test_fsdp:
 	python -m pytest -s -v ./tests/fsdp $(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_fsdp.log",)
 
+# Since the new version of pytest will *change* how things are collected, we need `deepspeed` to 
+# run after test_core and test_cli
+test:
+	$(MAKE) test_core
+	$(MAKE) test_cli
+	$(MAKE) test_big_modeling
+	$(MAKE) test_deepspeed
+	$(MAKE) test_fsdp
+
 test_examples:
 	python -m pytest -s -v ./tests/test_examples.py $(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_examples.log",)
 

README.md

@@ -22,22 +22,12 @@ limitations under the License.
 
 <p align="center">
     <!-- Uncomment when CircleCI is set up
-    <a href="https://circleci.com/gh/huggingface/accelerate">
-        <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/master">
-    </a>
+    <a href="https://circleci.com/gh/huggingface/accelerate"><img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/master"></a>
     -->
-    <a href="https://github.com/huggingface/accelerate/blob/main/LICENSE">
-        <img alt="License" src="https://img.shields.io/github/license/huggingface/accelerate.svg?color=blue">
-    </a>
-    <a href="https://huggingface.co/docs/accelerate/index.html">
-        <img alt="Documentation" src="https://img.shields.io/website/http/huggingface.co/docs/accelerate/index.html.svg?down_color=red&down_message=offline&up_message=online">
-    </a>
-    <a href="https://github.com/huggingface/accelerate/releases">
-        <img alt="GitHub release" src="https://img.shields.io/github/release/huggingface/accelerate.svg">
-    </a>
-    <a href="https://github.com/huggingface/accelerate/blob/main/CODE_OF_CONDUCT.md">
-        <img alt="Contributor Covenant" src="https://img.shields.io/badge/Contributor%20Covenant-v2.0%20adopted-ff69b4.svg">
-    </a>
+    <a href="https://github.com/huggingface/accelerate/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/github/license/huggingface/accelerate.svg?color=blue"></a>
+    <a href="https://huggingface.co/docs/accelerate/index.html"><img alt="Documentation" src="https://img.shields.io/website/http/huggingface.co/docs/accelerate/index.html.svg?down_color=red&down_message=offline&up_message=online"></a>
+    <a href="https://github.com/huggingface/accelerate/releases"><img alt="GitHub release" src="https://img.shields.io/github/release/huggingface/accelerate.svg"></a>
+    <a href="https://github.com/huggingface/accelerate/blob/main/CODE_OF_CONDUCT.md"><img alt="Contributor Covenant" src="https://img.shields.io/badge/Contributor%20Covenant-v2.0%20adopted-ff69b4.svg"></a>
 </p>
 
 <h3 align="center">
@@ -167,11 +157,21 @@ accelerate launch --multi_gpu --num_processes 2 examples/nlp_example.py
 
 To learn more, check the CLI documentation available [here](https://huggingface.co/docs/accelerate/package_reference/cli).
 
+Or view the configuration zoo [here](https://github.com/huggingface/accelerate/blob/main/examples/config_yaml_templates/)
+
 ## Launching multi-CPU run using MPI
 
 🤗 Here is another way to launch multi-CPU run using MPI. You can learn how to install Open MPI on [this page](https://www.open-mpi.org/faq/?category=building#easy-build). You can use Intel MPI or MVAPICH as well.
 Once you have MPI setup on your cluster, just run:
-
+```bash
+accelerate config
+```
+Answer the questions that are asked, selecting to run using multi-CPU, and answer "yes" when asked if you want accelerate to launch mpirun.
+Then, use `accelerate launch` with your script like:
+```bash
+accelerate launch examples/nlp_example.py
+```
+Alternatively, you can use mpirun directly, without using the CLI like:
 ```bash
 mpirun -np 2 python examples/nlp_example.py
 ```
@@ -220,6 +220,7 @@ You shouldn't use 🤗 Accelerate if you don't want to write a training loop you
 
 If you like the simplicity of 🤗 Accelerate but would prefer a higher-level abstraction around its capabilities, some frameworks and libraries that are built on top of 🤗 Accelerate are listed below:
 
+* [Amphion](https://github.com/open-mmlab/Amphion) is a toolkit for Audio, Music, and Speech Generation. Its purpose is to support reproducible research and help junior researchers and engineers get started in the field of audio, music, and speech generation research and development.
 * [Animus](https://github.com/Scitator/animus) is a minimalistic framework to run machine learning experiments. Animus highlights common "breakpoints" in ML experiments and provides a unified interface for them within [IExperiment](https://github.com/Scitator/animus/blob/main/animus/core.py#L76).
 * [Catalyst](https://github.com/catalyst-team/catalyst#getting-started) is a PyTorch framework for Deep Learning Research and Development. It focuses on reproducibility, rapid experimentation, and codebase reuse so you can create something new rather than write yet another train loop. Catalyst provides a [Runner](https://catalyst-team.github.io/catalyst/api/core.html#runner) to connect all parts of the experiment: hardware backend, data transformations, model training, and inference logic.
 * [fastai](https://github.com/fastai/fastai#installing) is a PyTorch framework for Deep Learning that simplifies training fast and accurate neural nets using modern best practices. fastai provides a [Learner](https://docs.fast.ai/learner.html#Learner) to handle the training, fine-tuning, and inference of deep learning algorithms.
@@ -257,7 +258,7 @@ pip install accelerate
 - multi-GPU on several nodes (machines)
 - TPU
 - FP16/BFloat16 mixed precision
-- FP8 mixed precision with [Transformer Engine](https://github.com/NVIDIA/TransformerEngine)
+- FP8 mixed precision with [Transformer Engine](https://github.com/NVIDIA/TransformerEngine) or [MS-AMP](https://github.com/Azure/MS-AMP/)
 - DeepSpeed support (Experimental)
 - PyTorch Fully Sharded Data Parallel (FSDP) support (Experimental)
 - Megatron-LM support (Experimental)
@@ -269,7 +270,7 @@ If you use 🤗 Accelerate in your publication, please cite it by using the foll
 ```bibtex
 @Misc{accelerate,
   title =        {Accelerate: Training and inference at scale made simple, efficient and adaptable.},
-  author =       {Sylvain Gugger, Lysandre Debut, Thomas Wolf, Philipp Schmid, Zachary Mueller, Sourab Mangrulkar, Marc Sun, Benjamin Bossan},
+  author =       {Sylvain Gugger and Lysandre Debut and Thomas Wolf and Philipp Schmid and Zachary Mueller and Sourab Mangrulkar and Marc Sun and Benjamin Bossan},
   howpublished = {\url{https://github.com/huggingface/accelerate}},
   year =         {2022}
 }

benchmarks/README.md

@@ -1,46 +1,5 @@
-# Big model inference benchmarks
+# Benchmarks
 
-Running inference with Accelerate on big models.
+The folders below contain suites to test various functionalities in Accelerate.
 
-## Setup
-
-These benchmarks use the `transformers` library:
-
-```bash
-pip install transformers
-```
-
-To reproduce or test a new setup, run
-
-```py
-python inference_acc.py model_name
-```
-
-This script supports `gpt-j-6b`, `gpt-neox`, `opt` (30B version) and `T0pp` out of the box, but you can specify any valid checkpoint for `model_name`.
-
-To force a different `torch_dtype` than the one in the config: `--torch_dtype xxx`.
-
-If you get an error linked to disk offload, you need to add the option `--disk-offload`
-
-## Results
-
-On a setup with two Titan RTXs (24GB of RAM) and 32GB of RAM, we get the following benchmarks (T0pp does not run in float16, which is why it's not included).
-
-| Model | Model load time | Generation time | dtype | GPU 0 use | GPU 1 use | CPU use | Disk offload |
-|:-----:|:---------------:|:---------------:|:-----:|:---------:|:---------:|:-------:|:------------:|
-| GPT-J-6B | 8.7s | 0.05s per token | float16 | 11.7GB | 0GB | 0GB | no |
-| GPT-J-6B | 12.4s | 0.06s per token | float32 | 21.9GB | 1.5GB | 0GB | no |
-| GPT-Neo-X-20B | 30.9s | 0.08s per token | float16 | 21.5GB | 18GB | 0GB | no |
-| GPT-Neo-X-20B | 78.2s | 10.72s per token | float32 | 20.3GB | 22.7 GB | 24.4GB | yes |
-| T0pp (11B) | 29.4s | 0.05s per token | float32 | 21.1GB | 21.3GB | 0GB | no |
-| OPT-30B | 34.5s | 2.37s per token | float16 | 20.7GB | 22.3GB | 14.1GB | no |
-| OPT-30B | 112.3s | 33.9s per token | float32 | 20.2GB | 21.2GB | 23.5GB | yes |
-
-Note on the results:
-- using two GPUs instead of one does not slow down generation
-- using CPU offload slows down a bit (see OPT-30b)
-- using disk offload slows down a lot (need to implement prefetching)
-
-You will also note that Accelerate does not use anymore GPU and CPU RAM than necessary:
-- peak GPU memory is exactly the size of the model put on a given GPU
-- peak CPU memory is either the size of the biggest checkpoint shard or the part of the model offloaded on CPU, whichever is bigger.
+See their relevant README.md's for more information.

benchmarks/big_model_inference/README.md

@@ -0,0 +1,46 @@
+# Big model inference benchmarks
+
+Running inference with Accelerate on big models.
+
+## Setup
+
+These benchmarks use the `transformers` library:
+
+```bash
+pip install transformers
+```
+
+To reproduce or test a new setup, run
+
+```py
+python inference_acc.py model_name
+```
+
+This script supports `gpt-j-6b`, `gpt-neox`, `opt` (30B version) and `T0pp` out of the box, but you can specify any valid checkpoint for `model_name`.
+
+To force a different `torch_dtype` than the one in the config: `--torch_dtype xxx`.
+
+If you get an error linked to disk offload, you need to add the option `--disk-offload`
+
+## Results
+
+On a setup with two Titan RTXs (24GB of RAM) and 32GB of RAM, we get the following benchmarks (T0pp does not run in float16, which is why it's not included).
+
+| Model | Model load time | Generation time | dtype | GPU 0 use | GPU 1 use | CPU use | Disk offload |
+|:-----:|:---------------:|:---------------:|:-----:|:---------:|:---------:|:-------:|:------------:|
+| GPT-J-6B | 8.7s | 0.05s per token | float16 | 11.7GB | 0GB | 0GB | no |
+| GPT-J-6B | 12.4s | 0.06s per token | float32 | 21.9GB | 1.5GB | 0GB | no |
+| GPT-Neo-X-20B | 30.9s | 0.08s per token | float16 | 21.5GB | 18GB | 0GB | no |
+| GPT-Neo-X-20B | 78.2s | 10.72s per token | float32 | 20.3GB | 22.7 GB | 24.4GB | yes |
+| T0pp (11B) | 29.4s | 0.05s per token | float32 | 21.1GB | 21.3GB | 0GB | no |
+| OPT-30B | 34.5s | 2.37s per token | float16 | 20.7GB | 22.3GB | 14.1GB | no |
+| OPT-30B | 112.3s | 33.9s per token | float32 | 20.2GB | 21.2GB | 23.5GB | yes |
+
+Note on the results:
+- using two GPUs instead of one does not slow down generation
+- using CPU offload slows down a bit (see OPT-30b)
+- using disk offload slows down a lot (need to implement prefetching)
+
+You will also note that Accelerate does not use anymore GPU and CPU RAM than necessary:
+- peak GPU memory is exactly the size of the model put on a given GPU
+- peak CPU memory is either the size of the biggest checkpoint shard or the part of the model offloaded on CPU, whichever is bigger.

benchmarks/big_model_inference/measures_util.py

@@ -1,3 +1,16 @@
+# 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.
 import gc
 import threading
 import time

benchmarks/fp8/ms_amp/Dockerfile

@@ -0,0 +1,12 @@
+FROM ghcr.io/azure/msamp
+
+RUN pip install transformers evaluate datasets
+RUN git clone https://github.com/huggingface/accelerate
+
+RUN cd accelerate && \
+    pip install -e . && \
+    cd benchmarks/fp8
+
+CMD ["bash"]
+
+

benchmarks/fp8/ms_amp/ddp.py

@@ -0,0 +1,123 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script tests to ensure that `accelerate` performs at the same level as raw `MS-AMP`.
+
+This particular script verifies this for DDP training.
+"""
+
+import evaluate
+import msamp
+import torch
+from fp8_utils import evaluate_model, get_training_utilities
+from torch.nn.parallel import DistributedDataParallel as DDP
+
+from accelerate import Accelerator
+from accelerate.state import AcceleratorState
+from accelerate.utils import FP8RecipeKwargs, get_grad_scaler, set_seed
+
+
+MODEL_NAME = "bert-base-cased"
+METRIC = evaluate.load("glue", "mrpc")
+
+
+def train_baseline(opt_level="O2"):
+    set_seed(42)
+    scaler = get_grad_scaler()
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(MODEL_NAME)
+    accelerator = Accelerator()
+    device = accelerator.device
+
+    model, optimizer = msamp.initialize(model, optimizer, opt_level=opt_level)
+
+    model.to(device)
+
+    # Convert the model to DDP
+    device_ids, output_device = [accelerator.local_process_index], accelerator.local_process_index
+    model = DDP(model, device_ids=device_ids, output_device=output_device)
+
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.train()
+
+    for i, batch in enumerate(train_dataloader):
+        with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+            outputs = model(**batch)
+            loss = outputs.loss
+        scaler.scale(loss).backward()
+        optimizer.step()
+        optimizer.zero_grad()
+        lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+def train_integration(opt_level="O2"):
+    kwargs_handlers = [FP8RecipeKwargs(backend="msamp", opt_level=opt_level)]
+    AcceleratorState()._reset_state(True)
+    accelerator = Accelerator(mixed_precision="fp8", kwargs_handlers=kwargs_handlers)
+    set_seed(42)
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(
+        MODEL_NAME, accelerator=accelerator
+    )
+
+    model, optimizer = accelerator.prepare(model, optimizer)
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.train()
+    for i, batch in enumerate(train_dataloader):
+        with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+            outputs = model(**batch)
+            loss = outputs.loss
+        accelerator.backward(loss)
+        optimizer.step()
+        optimizer.zero_grad()
+        lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+if __name__ == "__main__":
+    for opt_level in ["O1", "O2"]:
+        baseline_not_trained, baseline_trained = train_baseline(opt_level)
+        accelerator_not_trained, accelerator_trained = train_integration(opt_level)
+        assert (
+            baseline_not_trained["accuracy"] == accelerator_not_trained["accuracy"]
+        ), f'Accuracy not the same for untrained baseline and accelerator using opt_level={opt_level}: {baseline_not_trained["accuracy"]} == {accelerator_not_trained["accuracy"]}'
+        assert (
+            baseline_not_trained["f1"] == accelerator_not_trained["f1"]
+        ), f'F1 not the same for untrained baseline and accelerator using opt_level={opt_level}: {baseline_not_trained["f1"]} == {accelerator_not_trained["f1"]}'
+        assert (
+            baseline_trained["accuracy"] == accelerator_trained["accuracy"]
+        ), f'Accuracy not the same for trained baseline and accelerator using opt_level={opt_level}: {baseline_trained["accuracy"]} == {accelerator_trained["accuracy"]}'
+        assert (
+            baseline_trained["f1"] == accelerator_trained["f1"]
+        ), f'F1 not the same for trained baseline and accelerator using opt_level={opt_level}: {baseline_trained["f1"]} == {accelerator_trained["f1"]}'

benchmarks/fp8/ms_amp/distrib_deepspeed.py

@@ -0,0 +1,161 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script tests to ensure that `accelerate` performs at the same level as raw `MS-AMP`.
+
+This particular script verifies this for DeepSpeed training.
+
+NOTE: MS-AMP does *not* support ZeRO-3.
+"""
+
+# import msamp.deepspeed as msamp_deepspeed
+import evaluate
+import torch
+from fp8_utils import evaluate_model, get_training_utilities
+from msamp import deepspeed as msamp_deepspeed
+
+from accelerate import Accelerator, DeepSpeedPlugin
+from accelerate.state import AcceleratorState
+from accelerate.utils import set_seed
+
+
+MODEL_NAME = "bert-base-cased"
+METRIC = evaluate.load("glue", "mrpc")
+
+
+def train_baseline(zero_stage: int = 1, opt_level: str = "O1"):
+    set_seed(42)
+    accelerator = Accelerator()
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(
+        MODEL_NAME, accelerator=accelerator
+    )
+
+    import numpy as np
+
+    config = {
+        "train_batch_size": 32,
+        "train_micro_batch_size_per_gpu": 16,
+        "gradient_accumulation_steps": 1,
+        "zero_optimization": {
+            "stage": zero_stage,
+            "offload_optimizer": {"device": "none", "nvme_path": None},
+            "offload_param": {"device": "none", "nvme_path": None},
+        },
+        "gradient_clipping": 1.0,
+        "steps_per_print": np.inf,
+        "bf16": {"enabled": True},
+        "fp16": {"enabled": False},
+        "zero_allow_untested_optimizer": True,
+        "msamp": {
+            "enabled": True,
+            "opt_level": opt_level,
+        },
+    }
+    (
+        model,
+        optimizer,
+        _,
+        _,
+    ) = msamp_deepspeed.initialize(
+        model=model,
+        optimizer=optimizer,
+        config_params=config,
+    )
+
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.train()
+
+    for _ in range(2):
+        for batch in train_dataloader:
+            outputs = model(**batch)
+            loss = outputs.loss
+            model.backward(loss)
+            model.step()
+            for _ in range(accelerator.num_processes):
+                lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.destroy()
+    torch.cuda.empty_cache()
+    AcceleratorState()._reset_state(True)
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+def train_integration(zero_stage: int = 1, opt_level: str = "O1"):
+    set_seed(42)
+    deepspeed_plugin = DeepSpeedPlugin(
+        zero_stage=zero_stage,
+        enable_msamp=True,
+        msamp_opt_level=opt_level,
+    )
+    accelerator = Accelerator(mixed_precision="fp8", deepspeed_plugin=deepspeed_plugin)
+    accelerator.state.deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"] = 16
+
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(
+        MODEL_NAME, accelerator=accelerator
+    )
+
+    model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.train()
+    for _ in range(2):
+        for batch in train_dataloader:
+            outputs = model(**batch)
+            loss = outputs.loss
+            accelerator.backward(loss)
+            optimizer.step()
+            lr_scheduler.step()
+            optimizer.zero_grad()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.destroy()
+    torch.cuda.empty_cache()
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    AcceleratorState()._reset_state(True)
+    return base_model_results, trained_model_results
+
+
+if __name__ == "__main__":
+    for zero_stage in [1, 2]:
+        for opt_level in ["O1", "O2", "O3"]:
+            baseline_not_trained, baseline_trained = train_baseline(zero_stage, opt_level)
+            accelerator_not_trained, accelerator_trained = train_integration(zero_stage, opt_level)
+            assert (
+                baseline_not_trained["accuracy"] == accelerator_not_trained["accuracy"]
+            ), f'ZERO stage {zero_stage}, opt_level={opt_level}:\nAccuracy should be the same for the baseline and accelerator: {baseline_not_trained["accuracy"]} == {accelerator_not_trained["accuracy"]}'
+            assert (
+                baseline_not_trained["f1"] == accelerator_not_trained["f1"]
+            ), f'ZERO stage {zero_stage}, opt_level={opt_level}:\nF1 score should be the same for the baseline and accelerator: {baseline_not_trained["f1"]} == {accelerator_not_trained["f1"]}'
+            assert (
+                baseline_trained["accuracy"] == accelerator_trained["accuracy"]
+            ), f'ZERO stage {zero_stage}, opt_level={opt_level}:\nAccuracy should be the same for the baseline and accelerator: {baseline_trained["accuracy"]} == {accelerator_trained["accuracy"]}'
+            assert (
+                baseline_trained["f1"] == accelerator_trained["f1"]
+            ), f'ZERO stage {zero_stage}, opt_level={opt_level}:\nF1 score should be the same for the baseline and accelerator: {baseline_trained["f1"]} == {accelerator_trained["f1"]}'
+
+    torch.distributed.destroy_process_group()

benchmarks/fp8/ms_amp/fp8_utils.py

@@ -0,0 +1,118 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+
+def get_dataloaders(model_name: str, batch_size: int = 16):
+    from datasets import load_dataset
+    from torch.utils.data import DataLoader
+    from transformers import AutoTokenizer
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    datasets = load_dataset("glue", "mrpc")
+
+    def tokenize_function(examples):
+        # max_length=None => use the model max length (it's actually the default)
+        outputs = tokenizer(examples["sentence1"], examples["sentence2"], truncation=True, max_length=None)
+        return outputs
+
+    # Apply the method we just defined to all the examples in all the splits of the dataset
+    # starting with the main process first:
+    tokenized_datasets = datasets.map(
+        tokenize_function,
+        batched=True,
+        remove_columns=["idx", "sentence1", "sentence2"],
+    )
+
+    # We also rename the 'label' column to 'labels' which is the expected name for labels by the models of the
+    # transformers library
+    tokenized_datasets = tokenized_datasets.rename_column("label", "labels")
+
+    def collate_fn(examples):
+        return tokenizer.pad(
+            examples,
+            padding="longest",
+            pad_to_multiple_of=16,  # Specific for FP8
+            return_tensors="pt",
+        )
+
+    # Instantiate dataloaders.
+    train_dataloader = DataLoader(
+        tokenized_datasets["train"], shuffle=True, collate_fn=collate_fn, batch_size=batch_size, drop_last=True
+    )
+    eval_dataloader = DataLoader(
+        tokenized_datasets["validation"],
+        shuffle=False,
+        collate_fn=collate_fn,
+        batch_size=16,
+        drop_last=True,
+    )
+
+    return train_dataloader, eval_dataloader
+
+
+def get_training_utilities(model_name: str, batch_size: int = 16, accelerator=None):
+    """
+    Returns a tuple of:
+        - Model
+        - Optimizer
+        - Train dataloader (prepared)
+        - Eval dataloader (prepared)
+        - LR Scheduler
+    Suitable for training on the MRPC dataset
+    """
+    from torch.optim import AdamW
+    from transformers import AutoModelForSequenceClassification, get_linear_schedule_with_warmup
+
+    from accelerate import Accelerator
+
+    if accelerator is None:
+        accelerator = Accelerator()
+    model = AutoModelForSequenceClassification.from_pretrained(model_name)
+    train_dataloader, eval_dataloader = get_dataloaders(model_name, batch_size)
+    optimizer = AdamW(model.parameters(), lr=0.0001)
+    lr_scheduler = get_linear_schedule_with_warmup(
+        optimizer=optimizer,
+        num_warmup_steps=100,
+        num_training_steps=len(train_dataloader) * 2,
+    )
+    train_dataloader, eval_dataloader = accelerator.prepare(train_dataloader, eval_dataloader)
+    return model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
+
+
+def get_named_parameters(model):
+    """
+    Same thing as `Accelerator.get_named_parameters` Returns a list of the named parameters of the model (extracted
+    from parallel)
+    """
+    from accelerate.utils import extract_model_from_parallel
+
+    model = extract_model_from_parallel(model)
+    return {n: p for n, p in model.named_parameters()}
+
+
+def evaluate_model(model, dataloader, metric, accelerator=None):
+    "Turns model to .eval(), runs dataloader, calculates metric, then turns eval back on"
+    model.eval()
+    for step, batch in enumerate(dataloader):
+        with torch.no_grad():
+            # W/ MS-AMP, we need to cast while evaluating
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+                outputs = model(**batch)
+        predictions = outputs.logits.argmax(dim=-1)
+        references = batch["labels"]
+        if accelerator is not None and accelerator.num_processes > 1:
+            predictions, references = accelerator.gather_for_metrics((predictions, references))
+        metric.add_batch(predictions=predictions, references=references)
+    return metric.compute()

benchmarks/fp8/ms_amp/non_distributed.py

@@ -0,0 +1,118 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script tests to ensure that `accelerate` performs at the same level as raw `MS-AMP`.
+
+This particular script verifies this for single GPU training.
+"""
+
+import evaluate
+import msamp
+import torch
+from fp8_utils import evaluate_model, get_training_utilities
+
+from accelerate import Accelerator
+from accelerate.state import AcceleratorState
+from accelerate.utils import FP8RecipeKwargs, get_grad_scaler, set_seed
+
+
+MODEL_NAME = "bert-base-cased"
+METRIC = evaluate.load("glue", "mrpc")
+
+
+def train_baseline(opt_level="O2"):
+    set_seed(42)
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(MODEL_NAME)
+
+    model, optimizer = msamp.initialize(model, optimizer, opt_level=opt_level)
+    model.to("cuda")
+
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC)
+    model.train()
+    scaler = get_grad_scaler()
+
+    for batch in train_dataloader:
+        batch = batch.to("cuda")
+        with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+            outputs = model(**batch)
+        loss = outputs.loss
+        loss = scaler.scale(loss)
+        loss.backward()
+        optimizer.step()
+        optimizer.zero_grad()
+        lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC)
+
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+def train_integration(opt_level="O2"):
+    kwargs_handlers = [FP8RecipeKwargs(backend="msamp", opt_level=opt_level)]
+    AcceleratorState()._reset_state(True)
+    accelerator = Accelerator(mixed_precision="fp8", kwargs_handlers=kwargs_handlers)
+    set_seed(42)
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(
+        MODEL_NAME, accelerator=accelerator
+    )
+
+    model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC)
+    model.train()
+
+    for batch in train_dataloader:
+        outputs = model(**batch)
+        loss = outputs.loss
+        accelerator.backward(loss)
+        optimizer.step()
+        optimizer.zero_grad()
+        lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC)
+
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+if __name__ == "__main__":
+    for opt_level in ["O1", "O2"]:
+        baseline_not_trained, baseline_trained = train_baseline(opt_level)
+        accelerator_not_trained, accelerator_trained = train_integration(opt_level)
+
+        assert (
+            baseline_not_trained["accuracy"] == accelerator_not_trained["accuracy"]
+        ), f'Accuracy should be the same for the baseline and accelerator: {baseline_not_trained["accuracy"]} == {accelerator_not_trained["accuracy"]}'
+        assert (
+            baseline_not_trained["f1"] == accelerator_not_trained["f1"]
+        ), f'F1 score should be the same for the baseline and accelerator: {baseline_not_trained["f1"]} == {accelerator_not_trained["f1"]}'
+        assert (
+            baseline_trained["accuracy"] == accelerator_trained["accuracy"]
+        ), f'Accuracy should be the same for the baseline and accelerator: {baseline_trained["accuracy"]} == {accelerator_trained["accuracy"]}'
+        assert (
+            baseline_trained["f1"] == accelerator_trained["f1"]
+        ), f'F1 score should be the same for the baseline and accelerator: {baseline_trained["f1"]} == {accelerator_trained["f1"]}'

benchmarks/fp8/transformer_engine/Dockerfile

@@ -0,0 +1,12 @@
+FROM nvcr.io/nvidia/pytorch:24.07-py3
+
+RUN pip install transformers evaluate datasets
+RUN git clone https://github.com/huggingface/accelerate.git
+
+RUN cd accelerate && \
+    pip install -e . && \
+    cd benchmarks/fp8
+
+RUN /bin/bash
+
+

benchmarks/fp8/transformer_engine/README.md

@@ -0,0 +1,32 @@
+# FP8 Benchmarks
+
+Comparing and running [TransformerEngine](https://github.com/NVIDIA/TransformerEngine) FP8 with accelerate
+
+## Overview
+
+This repo provides scripts which compare native TransformerEngine model training against `accelerate`'s own integration. Each modeling type is segmented out via a script, supporting the following:
+
+* Single GPU training (`non_distributed.py`)
+* Multi-GPU training via DistributedDataParallelism (`ddp.py`)
+* Fully Sharded Data Parallelism (`fsdp.py`)
+* DeepSpeed ZeRO 1-3 (`deepspeed.py`)
+
+To run them, it's recommended to use a docker image (see the attached `Dockerfile`) and not install `TransformerEngine` manually.
+
+## Running:
+
+There are official Docker images located at `huggingface/accelerate:gpu-fp8-transformerengine-nightly` which can be used.
+
+You can run all scripts using the core `accelerate launch` command without any `accelerate config` being needed.
+
+For single GPU, run it via `python`:
+
+```bash
+python non_distributed.py
+```
+
+For the rest, run it via `accelerate launch`:
+
+```bash
+accelerate launch ddp.py # or distrib_deepspeed.py, ddp.py
+```

benchmarks/fp8/transformer_engine/ddp.py

@@ -0,0 +1,144 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script tests to ensure that `accelerate` performs at the same level as raw `TransformersEngine`.
+
+This particular script verifies this for DDP training.
+"""
+
+import evaluate
+import torch
+import transformer_engine.common.recipe as te_recipe
+import transformer_engine.pytorch as te
+from fp8_utils import evaluate_model, get_named_parameters, get_training_utilities
+from torch.nn.parallel import DistributedDataParallel as DDP
+from transformer_engine.common.recipe import DelayedScaling
+
+from accelerate import Accelerator
+from accelerate.state import AcceleratorState
+from accelerate.utils import FP8RecipeKwargs, set_seed
+from accelerate.utils.transformer_engine import convert_model
+
+
+MODEL_NAME = "bert-base-cased"
+METRIC = evaluate.load("glue", "mrpc")
+
+
+def train_baseline():
+    set_seed(42)
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(MODEL_NAME)
+    accelerator = Accelerator()
+    device = accelerator.device
+    model.to(device)
+
+    # Convert the model to TE
+    old_named_params = get_named_parameters(model)
+
+    with torch.no_grad():
+        convert_model(model)
+
+    FP8_RECIPE_KWARGS = {"fp8_format": te_recipe.Format.HYBRID, "amax_history_len": 32, "amax_compute_algo": "max"}
+    fp8_recipe = DelayedScaling(**FP8_RECIPE_KWARGS)
+
+    new_named_params = get_named_parameters(model)
+
+    # Convert the model to DDP
+    device_ids, output_device = [accelerator.local_process_index], accelerator.local_process_index
+    model = DDP(model, device_ids=device_ids, output_device=output_device)
+
+    mapping = {p: new_named_params[n] for n, p in old_named_params.items()}
+    for param_group in optimizer.param_groups:
+        param_group["params"] = [mapping[p] for p in param_group["params"]]
+
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.train()
+
+    for _ in range(2):
+        for batch in train_dataloader:
+            with te.fp8_autocast(enabled=True, fp8_recipe=fp8_recipe):
+                with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+                    batch = batch.to(device)
+                    outputs = model(**batch)
+            loss = outputs.loss
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+            lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+def train_integration():
+    FP8_RECIPE_KWARGS = {"fp8_format": "HYBRID", "amax_history_len": 32, "amax_compute_algo": "max"}
+    kwargs_handlers = [FP8RecipeKwargs(backend="TE", **FP8_RECIPE_KWARGS)]
+    AcceleratorState()._reset_state(True)
+    accelerator = Accelerator(mixed_precision="fp8", kwargs_handlers=kwargs_handlers)
+    set_seed(42)
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(
+        MODEL_NAME, accelerator=accelerator
+    )
+
+    model, optimizer = accelerator.prepare(model, optimizer)
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.train()
+
+    for _ in range(2):
+        for batch in train_dataloader:
+            outputs = model(**batch)
+            loss = outputs.loss
+            accelerator.backward(loss)
+            optimizer.step()
+            optimizer.zero_grad()
+            lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+if __name__ == "__main__":
+    baseline_not_trained, baseline_trained = train_baseline()
+    accelerator_not_trained, accelerator_trained = train_integration()
+
+    assert (
+        baseline_not_trained["accuracy"] == accelerator_not_trained["accuracy"]
+    ), f'Accuracy should be the same for the baseline and accelerator: {baseline_not_trained["accuracy"]} == {accelerator_not_trained["accuracy"]}'
+    assert (
+        baseline_not_trained["f1"] == accelerator_not_trained["f1"]
+    ), f'F1 score should be the same for the baseline and accelerator: {baseline_not_trained["f1"]} == {accelerator_not_trained["f1"]}'
+    assert (
+        baseline_trained["accuracy"] == accelerator_trained["accuracy"]
+    ), f'Accuracy should be the same for the baseline and accelerator: {baseline_trained["accuracy"]} == {accelerator_trained["accuracy"]}'
+    assert (
+        baseline_trained["f1"] == accelerator_trained["f1"]
+    ), f'F1 score should be the same for the baseline and accelerator: {baseline_trained["f1"]} == {accelerator_trained["f1"]}'
+
+    torch.distributed.destroy_process_group()

benchmarks/fp8/transformer_engine/distrib_deepspeed.py

@@ -0,0 +1,190 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script tests to ensure that `accelerate` performs at the same level as raw `TransformersEngine`.
+
+This particular script verifies this for DDP training.
+"""
+
+from unittest.mock import patch
+
+import deepspeed
+import evaluate
+import torch
+import transformer_engine.common.recipe as te_recipe
+import transformer_engine.pytorch as te
+from fp8_utils import evaluate_model, get_named_parameters, get_training_utilities
+from transformer_engine.common.recipe import DelayedScaling
+
+from accelerate import Accelerator, DeepSpeedPlugin
+from accelerate.state import AcceleratorState
+from accelerate.utils import FP8RecipeKwargs, set_seed
+from accelerate.utils.transformer_engine import convert_model
+
+
+MODEL_NAME = "bert-base-cased"
+METRIC = evaluate.load("glue", "mrpc")
+
+
+def train_baseline(zero_stage: int = 1):
+    # This forces transformers to think Zero-3 Init should be used
+    with patch("transformers.integrations.deepspeed.is_deepspeed_zero3_enabled") as mock:
+        mock.return_value = zero_stage == 3
+    set_seed(42)
+
+    accelerator = Accelerator()
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(
+        MODEL_NAME, accelerator=accelerator
+    )
+
+    # Convert the model to TE
+    old_named_params = get_named_parameters(model)
+
+    with torch.no_grad():
+        convert_model(model)
+    new_named_params = get_named_parameters(model)
+
+    mapping = {p: new_named_params[n] for n, p in old_named_params.items()}
+    for param_group in optimizer.param_groups:
+        param_group["params"] = [mapping[p] for p in param_group["params"]]
+
+    FP8_RECIPE_KWARGS = {"fp8_format": te_recipe.Format.HYBRID, "amax_history_len": 32, "amax_compute_algo": "max"}
+    fp8_recipe = DelayedScaling(**FP8_RECIPE_KWARGS)
+
+    import numpy as np
+
+    config = {
+        "train_batch_size": 32,
+        "train_micro_batch_size_per_gpu": 16,
+        "gradient_accumulation_steps": 1,
+        "zero_optimization": {
+            "stage": zero_stage,
+            "offload_optimizer": {"device": "none", "nvme_path": None},
+            "offload_param": {"device": "none", "nvme_path": None},
+            "stage3_gather_16bit_weights_on_model_save": False,
+        },
+        "gradient_clipping": 1.0,
+        "steps_per_print": np.inf,
+        "bf16": {"enabled": True},
+        "fp16": {"enabled": False},
+        "zero_allow_untested_optimizer": True,
+    }
+
+    (
+        model,
+        optimizer,
+        _,
+        _,
+    ) = deepspeed.initialize(
+        model=model,
+        optimizer=optimizer,
+        config_params=config,
+    )
+
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.train()
+
+    model_outputs = []
+    data = []
+
+    for _ in range(2):
+        for batch in train_dataloader:
+            with te.fp8_autocast(enabled=True, fp8_recipe=fp8_recipe):
+                outputs = model(**batch)
+                data.append(batch.to("cpu"))
+            model_outputs.append(outputs.logits.to("cpu"))
+            loss = outputs.loss
+            model.backward(loss)
+            model.step()
+            for _ in range(accelerator.num_processes):
+                lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.destroy()
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results, model_outputs, data
+
+
+def train_integration(zero_stage: int = 1):
+    set_seed(42)
+    FP8_RECIPE_KWARGS = {"fp8_format": "HYBRID", "amax_history_len": 32, "amax_compute_algo": "max"}
+    kwargs_handlers = [FP8RecipeKwargs(backend="TE", **FP8_RECIPE_KWARGS)]
+    AcceleratorState()._reset_state(True)
+    deepspeed_plugin = DeepSpeedPlugin(
+        zero_stage=zero_stage,
+        zero3_init_flag=zero_stage == 3,
+    )
+    accelerator = Accelerator(
+        mixed_precision="fp8", kwargs_handlers=kwargs_handlers, deepspeed_plugin=deepspeed_plugin
+    )
+    accelerator.state.deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"] = 16
+
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(
+        MODEL_NAME, accelerator=accelerator
+    )
+
+    model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.train()
+    model_outputs = []
+    data = []
+    for _ in range(2):
+        for batch in train_dataloader:
+            outputs = model(**batch)
+            data.append(batch.to("cpu"))
+            model_outputs.append(outputs.logits.to("cpu"))
+            loss = outputs.loss
+            accelerator.backward(loss)
+            optimizer.step()
+            lr_scheduler.step()
+            optimizer.zero_grad()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.destroy()
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results, model_outputs, data
+
+
+if __name__ == "__main__":
+    # for zero_stage in [1, 2, 3]:
+    zero_stage = 1
+    baseline_not_trained, baseline_trained, baseline_outputs, baseline_data = train_baseline(zero_stage)
+    accelerator_not_trained, accelerator_trained, accelerator_outputs, accelerator_data = train_integration(zero_stage)
+    assert (
+        baseline_not_trained["accuracy"] == accelerator_not_trained["accuracy"]
+    ), f'ZERO stage {zero_stage}: Accuracy should be the same for the baseline and accelerator: {baseline_not_trained["accuracy"]} == {accelerator_not_trained["accuracy"]}'
+    assert (
+        baseline_not_trained["f1"] == accelerator_not_trained["f1"]
+    ), f'ZERO stage {zero_stage}: F1 score should be the same for the baseline and accelerator: {baseline_not_trained["f1"]} == {accelerator_not_trained["f1"]}'
+    assert (
+        baseline_trained["accuracy"] == accelerator_trained["accuracy"]
+    ), f'ZERO stage {zero_stage}: Accuracy should be the same for the baseline and accelerator: {baseline_trained["accuracy"]} == {accelerator_trained["accuracy"]}'
+    assert (
+        baseline_trained["f1"] == accelerator_trained["f1"]
+    ), f'ZERO stage {zero_stage}: F1 score should be the same for the baseline and accelerator: {baseline_trained["f1"]} == {accelerator_trained["f1"]}'
+
+    torch.distributed.destroy_process_group()

benchmarks/fp8/transformer_engine/fp8_utils.py

@@ -0,0 +1,116 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+
+def get_dataloaders(model_name: str, batch_size: int = 16):
+    from datasets import load_dataset
+    from torch.utils.data import DataLoader
+    from transformers import AutoTokenizer
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    datasets = load_dataset("glue", "mrpc")
+
+    def tokenize_function(examples):
+        # max_length=None => use the model max length (it's actually the default)
+        outputs = tokenizer(examples["sentence1"], examples["sentence2"], truncation=True, max_length=None)
+        return outputs
+
+    # Apply the method we just defined to all the examples in all the splits of the dataset
+    # starting with the main process first:
+    tokenized_datasets = datasets.map(
+        tokenize_function,
+        batched=True,
+        remove_columns=["idx", "sentence1", "sentence2"],
+    )
+
+    # We also rename the 'label' column to 'labels' which is the expected name for labels by the models of the
+    # transformers library
+    tokenized_datasets = tokenized_datasets.rename_column("label", "labels")
+
+    def collate_fn(examples):
+        return tokenizer.pad(
+            examples,
+            padding="longest",
+            pad_to_multiple_of=16,  # Specific for FP8
+            return_tensors="pt",
+        )
+
+    # Instantiate dataloaders.
+    train_dataloader = DataLoader(
+        tokenized_datasets["train"], shuffle=True, collate_fn=collate_fn, batch_size=batch_size, drop_last=True
+    )
+    eval_dataloader = DataLoader(
+        tokenized_datasets["validation"],
+        shuffle=False,
+        collate_fn=collate_fn,
+        batch_size=16,
+        drop_last=True,
+    )
+
+    return train_dataloader, eval_dataloader
+
+
+def get_training_utilities(model_name: str, batch_size: int = 16, accelerator=None):
+    """
+    Returns a tuple of:
+        - Model
+        - Optimizer
+        - Train dataloader (prepared)
+        - Eval dataloader (prepared)
+        - LR Scheduler
+    Suitable for training on the MRPC dataset
+    """
+    from torch.optim import AdamW
+    from transformers import AutoModelForSequenceClassification, get_linear_schedule_with_warmup
+
+    from accelerate import Accelerator
+
+    if accelerator is None:
+        accelerator = Accelerator()
+    model = AutoModelForSequenceClassification.from_pretrained(model_name)
+    train_dataloader, eval_dataloader = get_dataloaders(model_name, batch_size)
+    optimizer = AdamW(model.parameters(), lr=0.0001)
+    lr_scheduler = get_linear_schedule_with_warmup(
+        optimizer=optimizer,
+        num_warmup_steps=100,
+        num_training_steps=len(train_dataloader) * 2,
+    )
+    train_dataloader, eval_dataloader = accelerator.prepare(train_dataloader, eval_dataloader)
+    return model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
+
+
+def get_named_parameters(model):
+    """
+    Same thing as `Accelerator.get_named_parameters` Returns a list of the named parameters of the model (extracted
+    from parallel)
+    """
+    from accelerate.utils import extract_model_from_parallel
+
+    model = extract_model_from_parallel(model)
+    return {n: p for n, p in model.named_parameters()}
+
+
+def evaluate_model(model, dataloader, metric, accelerator=None):
+    "Turns model to .eval(), runs dataloader, calculates metric, then turns eval back on"
+    model.eval()
+    for step, batch in enumerate(dataloader):
+        with torch.no_grad():
+            outputs = model(**batch)
+        predictions = outputs.logits.argmax(dim=-1)
+        references = batch["labels"]
+        if accelerator is not None and accelerator.num_processes > 1:
+            predictions, references = accelerator.gather_for_metrics((predictions, references))
+        metric.add_batch(predictions=predictions, references=references)
+    return metric.compute()

benchmarks/fp8/transformer_engine/fsdp.py

@@ -0,0 +1,161 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script tests to ensure that `accelerate` performs at the same level as raw `TransformersEngine`.
+
+This particular script verifies this for FSDP training.
+"""
+
+from functools import partial
+
+import evaluate
+import torch
+import transformer_engine.common.recipe as te_recipe
+import transformer_engine.pytorch as te
+from fp8_utils import evaluate_model, get_named_parameters, get_training_utilities
+from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+from torch.distributed.fsdp import MixedPrecision
+from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy
+from transformer_engine.common.recipe import DelayedScaling
+from transformers.models.bert import BertLayer
+
+from accelerate import Accelerator
+from accelerate import FullyShardedDataParallelPlugin as FSDPPlugin
+from accelerate.state import AcceleratorState
+from accelerate.utils import FP8RecipeKwargs, set_seed
+from accelerate.utils.transformer_engine import convert_model
+
+
+MODEL_NAME = "bert-base-cased"
+METRIC = evaluate.load("glue", "mrpc")
+
+FSDP_WRAP_POLICY = partial(transformer_auto_wrap_policy, transformer_layer_cls={BertLayer})
+
+
+def train_baseline():
+    set_seed(42)
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(MODEL_NAME)
+    accelerator = Accelerator()
+    device = accelerator.device
+    model.to(device)
+
+    # Convert the model to TE
+    old_named_params = get_named_parameters(model)
+
+    with torch.no_grad():
+        convert_model(model)
+
+    FP8_RECIPE_KWARGS = {"fp8_format": te_recipe.Format.HYBRID, "amax_history_len": 32, "amax_compute_algo": "max"}
+    fp8_recipe = DelayedScaling(**FP8_RECIPE_KWARGS)
+
+    new_named_params = get_named_parameters(model)
+
+    # Convert the model to FSDP
+    model = FSDP(
+        model,
+        use_orig_params=True,
+        mixed_precision=MixedPrecision(param_dtype=torch.bfloat16, reduce_dtype=torch.float32),
+        auto_wrap_policy=FSDP_WRAP_POLICY,
+    )
+
+    mapping = {p: new_named_params[n] for n, p in old_named_params.items()}
+    for param_group in optimizer.param_groups:
+        param_group["params"] = [mapping[p] for p in param_group["params"]]
+
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.train()
+
+    for _ in range(2):
+        for batch in train_dataloader:
+            with te.fp8_autocast(enabled=True, fp8_recipe=fp8_recipe):
+                with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+                    batch = batch.to(device)
+                    outputs = model(**batch)
+            loss = outputs.loss
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+            lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+def train_integration():
+    FP8_RECIPE_KWARGS = {"fp8_format": "HYBRID", "amax_history_len": 32, "amax_compute_algo": "max"}
+    kwargs_handlers = [FP8RecipeKwargs(backend="TE", **FP8_RECIPE_KWARGS)]
+    AcceleratorState()._reset_state(True)
+    fsdp_plugin = FSDPPlugin(
+        auto_wrap_policy=FSDP_WRAP_POLICY,
+        use_orig_params=True,
+        mixed_precision_policy=MixedPrecision(param_dtype=torch.bfloat16, reduce_dtype=torch.float32),
+    )
+    accelerator = Accelerator(mixed_precision="fp8", fsdp_plugin=fsdp_plugin, kwargs_handlers=kwargs_handlers)
+    set_seed(42)
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(
+        MODEL_NAME, accelerator=accelerator
+    )
+
+    model, optimizer = accelerator.prepare(model, optimizer)
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+    model.train()
+
+    for _ in range(2):
+        for batch in train_dataloader:
+            outputs = model(**batch)
+            loss = outputs.loss
+            accelerator.backward(loss)
+            optimizer.step()
+            optimizer.zero_grad()
+            lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
+
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+if __name__ == "__main__":
+    baseline_not_trained, baseline_trained = train_baseline()
+    accelerator_not_trained, accelerator_trained = train_integration()
+
+    assert (
+        baseline_not_trained["accuracy"] == accelerator_not_trained["accuracy"]
+    ), f'Accuracy should be the same for the baseline and accelerator: {baseline_not_trained["accuracy"]} == {accelerator_not_trained["accuracy"]}'
+    assert (
+        baseline_not_trained["f1"] == accelerator_not_trained["f1"]
+    ), f'F1 score should be the same for the baseline and accelerator: {baseline_not_trained["f1"]} == {accelerator_not_trained["f1"]}'
+    assert (
+        baseline_trained["accuracy"] == accelerator_trained["accuracy"]
+    ), f'Accuracy should be the same for the baseline and accelerator: {baseline_trained["accuracy"]} == {accelerator_trained["accuracy"]}'
+    assert (
+        baseline_trained["f1"] == accelerator_trained["f1"]
+    ), f'F1 score should be the same for the baseline and accelerator: {baseline_trained["f1"]} == {accelerator_trained["f1"]}'
+
+    torch.distributed.destroy_process_group()

benchmarks/fp8/transformer_engine/non_distributed.py

@@ -0,0 +1,132 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script tests to ensure that `accelerate` performs at the same level as raw `TransformersEngine`.
+
+This particular script verifies this for single GPU training.
+"""
+
+import evaluate
+import torch
+import transformer_engine.common.recipe as te_recipe
+import transformer_engine.pytorch as te
+from fp8_utils import evaluate_model, get_named_parameters, get_training_utilities
+from transformer_engine.common.recipe import DelayedScaling
+
+from accelerate import Accelerator
+from accelerate.state import AcceleratorState
+from accelerate.utils import FP8RecipeKwargs, set_seed
+from accelerate.utils.transformer_engine import convert_model
+
+
+MODEL_NAME = "bert-base-cased"
+METRIC = evaluate.load("glue", "mrpc")
+
+
+def train_baseline():
+    set_seed(42)
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(MODEL_NAME)
+
+    # Convert the model to TE
+    old_named_params = get_named_parameters(model)
+
+    with torch.no_grad():
+        convert_model(model)
+
+    new_named_params = get_named_parameters(model)
+    mapping = {p: new_named_params[n] for n, p in old_named_params.items()}
+    for param_group in optimizer.param_groups:
+        param_group["params"] = [mapping[p] for p in param_group["params"]]
+
+    FP8_RECIPE_KWARGS = {"fp8_format": te_recipe.Format.HYBRID, "amax_history_len": 32, "amax_compute_algo": "max"}
+    fp8_recipe = DelayedScaling(**FP8_RECIPE_KWARGS)
+
+    model.to("cuda")
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC)
+    model.train()
+
+    for batch in train_dataloader:
+        with te.fp8_autocast(enabled=True, fp8_recipe=fp8_recipe):
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+                batch = batch.to("cuda")
+                outputs = model(**batch)
+        loss = outputs.loss
+        loss.backward()
+        optimizer.step()
+        optimizer.zero_grad()
+        lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC)
+
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+def train_integration():
+    FP8_RECIPE_KWARGS = {"fp8_format": "HYBRID", "amax_history_len": 32, "amax_compute_algo": "max"}
+    kwargs_handlers = [FP8RecipeKwargs(backend="TE", **FP8_RECIPE_KWARGS)]
+    AcceleratorState()._reset_state(True)
+    accelerator = Accelerator(mixed_precision="fp8", kwargs_handlers=kwargs_handlers)
+    set_seed(42)
+    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(
+        MODEL_NAME, accelerator=accelerator
+    )
+
+    model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)
+    base_model_results = evaluate_model(model, eval_dataloader, METRIC)
+    model.train()
+
+    for batch in train_dataloader:
+        outputs = model(**batch)
+        loss = outputs.loss
+        accelerator.backward(loss)
+        optimizer.step()
+        optimizer.zero_grad()
+        lr_scheduler.step()
+
+    trained_model_results = evaluate_model(model, eval_dataloader, METRIC)
+
+    assert (
+        trained_model_results["accuracy"] > base_model_results["accuracy"]
+    ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}'
+    assert (
+        trained_model_results["f1"] > base_model_results["f1"]
+    ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}'
+
+    return base_model_results, trained_model_results
+
+
+if __name__ == "__main__":
+    baseline_not_trained, baseline_trained = train_baseline()
+    accelerator_not_trained, accelerator_trained = train_integration()
+
+    assert (
+        baseline_not_trained["accuracy"] == accelerator_not_trained["accuracy"]
+    ), f'Accuracy should be the same for the baseline and accelerator: {baseline_not_trained["accuracy"]} == {accelerator_not_trained["accuracy"]}'
+    assert (
+        baseline_not_trained["f1"] == accelerator_not_trained["f1"]
+    ), f'F1 score should be the same for the baseline and accelerator: {baseline_not_trained["f1"]} == {accelerator_not_trained["f1"]}'
+    assert (
+        baseline_trained["accuracy"] == accelerator_trained["accuracy"]
+    ), f'Accuracy should be the same for the baseline and accelerator: {baseline_trained["accuracy"]} == {accelerator_trained["accuracy"]}'
+    assert (
+        baseline_trained["f1"] == accelerator_trained["f1"]
+    ), f'F1 score should be the same for the baseline and accelerator: {baseline_trained["f1"]} == {accelerator_trained["f1"]}'

docker/README.md

@@ -0,0 +1,74 @@
+<!---
+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.
+-->
+
+# Official Hugging Face Accelerate Docker Images
+
+Accelerate publishes a variety of docker versions as part of our CI that users can also use. These are stable images that Accelerate can run off of which comes with a variety of different setup configurations, all of which are officially hosted on [Docker Hub](https://hub.docker.com/r/huggingface/accelerate).
+
+A breakdown of each are given below
+
+## Naming Conventions
+
+Accelerate docker images follow a tagging convention of:
+
+```bash
+huggingface/accelerate:{accelerator}-{nightly,release}
+```
+
+`accelerator` in this instance is one of many applical pre-configured backend supports:
+* `gpu`: Comes compiled off of the `nvidia/cuda` image and includes core parts like `bitsandbytes`. Runs off python 3.9.
+* `cpu`: Comes compiled off of `python:3.9-slim` and is designed for non-CUDA based workloads.
+* More to come soon
+* `gpu-deepspeed`: Comes compiled off of the `nvidia/cuda` image and includes core parts like `bitsandbytes` as well as the latest `deepspeed` version. Runs off python 3.10.
+* `gpu-fp8-transformerengine`: Comes compiled off of `nvcr.io/nvidia/pytorch` and is specifically for running the `benchmarks/fp8` scripts on devices which support FP8 operations using the `TransformerEngine` library (RTX 4090, H100, etc)
+
+## Nightlies vs Releases
+
+Each release a new build is pushed with a version number included in the name. For a GPU-supported image of version 0.28.0 for instance, it would look like the following:
+
+```bash
+huggingface/accelerate:gpu-release-0.28.0
+```
+
+Nightlies contain two different image tags. There is a general `nightly` tag which is built each night, and a `nightly-YYYY-MM-DD` which corresponds to a build from a particular date.
+
+For instance, here is an example nightly CPU image from 3/14/2024
+
+```bash
+huggingface/accelerate:cpu-nightly-2024-03-14
+```
+
+## Running the images
+
+Each image comes compiled with `conda` and an `accelerate` environment contains all of the installed dependencies. 
+
+To pull down the latest nightly run:
+
+```bash
+docker pull huggingface/accelerate:gpu-nightly
+```
+
+To then run it in interactive mode with GPU-memory available, run:
+
+```bash
+docker container run --gpus all -it huggingface/accelerate:gpu-nightly
+```
+
+## DEPRECATED IMAGES
+
+CPU and GPU docker images were hosted at `huggingface/accelerate-gpu` and `huggingface/accelerate-cpu`. These builds are now outdated and will not receive updates. 
+
+The builds at the corresponding `huggingface/accelerate:{gpu,cpu}` contain the same `Dockerfile`, so it's as simple as changing the docker image to the desired ones from above. We will not be deleting these images for posterity, but they will not be receiving updates going forward.

docker/accelerate-cpu/Dockerfile

@@ -1,7 +1,7 @@
 # Builds CPU-only Docker image of PyTorch
 # Uses multi-staged approach to reduce size
 # Stage 1
-FROM python:3.8-slim as compile-image
+FROM python:3.9-slim as compile-image
 
 ARG DEBIAN_FRONTEND=noninteractive
 
@@ -25,7 +25,7 @@ RUN python3 -m pip install --no-cache-dir \
     --extra-index-url https://download.pytorch.org/whl/cpu
     
 # Stage 2
-FROM python:3.8-slim AS build-image
+FROM python:3.9-slim AS build-image
 COPY --from=compile-image /opt/venv /opt/venv
 RUN useradd -ms /bin/bash user
 USER user

docker/accelerate-gpu-deepspeed/Dockerfile

@@ -0,0 +1,46 @@
+# Builds GPU docker image of PyTorch specifically
+# 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
+# Note: DeepSpeed beyond v0.12.6 requires py 3.10
+ENV PYTHON_VERSION=3.10
+# Install apt libs
+RUN apt-get update && \
+    apt-get install -y curl git wget && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists*
+
+# Create our conda env
+RUN conda create --name accelerate python=${PYTHON_VERSION} ipython jupyter pip
+# We don't install pytorch here yet since CUDA isn't available
+# instead we use the direct torch wheel
+ENV PATH /opt/conda/envs/accelerate/bin:$PATH
+# Activate our bash shell
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+# Activate the conda env, install mpy4pi, and install torch + accelerate
+RUN source activate accelerate && conda install -c conda-forge mpi4py
+RUN source activate accelerate && \
+    python3 -m pip install --no-cache-dir \
+    git+https://github.com/huggingface/accelerate#egg=accelerate[testing,test_trackers,deepspeed] \
+    --extra-index-url https://download.pytorch.org/whl/cu117
+
+RUN python3 -m pip install --no-cache-dir bitsandbytes
+
+# Stage 2
+FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu20.04 AS build-image
+COPY --from=compile-image /opt/conda /opt/conda
+ENV PATH /opt/conda/bin:$PATH
+
+# 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 accelerate" >> ~/.profile
+
+# Activate the virtualenv
+CMD ["/bin/bash"]

docker/accelerate-gpu/Dockerfile

@@ -1,10 +1,10 @@
-# Builds GPU docker image of PyTorch
+# Builds GPU docker image of PyTorch specifically
 # 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
+ENV PYTHON_VERSION=3.9
 # Install apt libs
 RUN apt-get update && \
     apt-get install -y curl git wget && \
@@ -19,7 +19,8 @@ ENV PATH /opt/conda/envs/accelerate/bin:$PATH
 # Activate our bash shell
 RUN chsh -s /bin/bash
 SHELL ["/bin/bash", "-c"]
-# Activate the conda env and install torch + accelerate
+# Activate the conda env, install mpy4pi, and install torch + accelerate
+RUN source activate accelerate && conda install -c conda-forge mpi4py
 RUN source activate accelerate && \
     python3 -m pip install --no-cache-dir \
     git+https://github.com/huggingface/accelerate#egg=accelerate[testing,test_trackers] \
@@ -28,7 +29,7 @@ RUN source activate accelerate && \
 RUN python3 -m pip install --no-cache-dir bitsandbytes
 
 # Stage 2
-FROM nvidia/cuda:11.2.2-cudnn8-devel-ubuntu20.04 AS build-image
+FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu20.04 AS build-image
 COPY --from=compile-image /opt/conda /opt/conda
 ENV PATH /opt/conda/bin:$PATH
 

docs/source/_toctree.yml

@@ -10,51 +10,72 @@
   - local: basic_tutorials/overview
     title: Overview
   - local: basic_tutorials/migration
-    title: Migrating to 🤗 Accelerate
+    title: Add Accelerate to your code
+  - local: basic_tutorials/execution
+    title: Execution process
+  - local: basic_tutorials/tpu
+    title: TPU training
   - local: basic_tutorials/launch
-    title: Launching distributed code
+    title: Launching Accelerate scripts
   - local: basic_tutorials/notebook
     title: Launching distributed training from Jupyter Notebooks
   title: Tutorials
 - sections:
-  - local: usage_guides/explore
-    title: Start Here!
-  - local: usage_guides/training_zoo
-    title: Example Zoo
-  - local: usage_guides/big_modeling
-    title: How to perform inference on large models with small resources
-  - local: usage_guides/model_size_estimator
-    title: Knowing how big of a model you can fit into memory
-  - local: usage_guides/quantization
-    title: How to quantize model 
-  - local: usage_guides/distributed_inference
-    title: How to perform distributed inference with normal resources
-  - local: usage_guides/gradient_accumulation
-    title: Performing gradient accumulation
-  - local: usage_guides/local_sgd
-    title: Accelerating training with local SGD
-  - local: usage_guides/checkpoint
-    title: Saving and loading training states
-  - local: usage_guides/tracking
-    title: Using experiment trackers
-  - local: usage_guides/debug
-    title: Debugging timeout errors
-  - local: usage_guides/memory
-    title: How to avoid CUDA Out-of-Memory
-  - local: usage_guides/mps
-    title: How to use Apple Silicon M1 GPUs
-  - local: usage_guides/deepspeed
-    title: How to use DeepSpeed
-  - local: usage_guides/fsdp
-    title: How to use Fully Sharded Data Parallelism
-  - local: usage_guides/megatron_lm
-    title: How to use Megatron-LM
-  - local: usage_guides/sagemaker
-    title: How to use 🤗 Accelerate with SageMaker
-  - local: usage_guides/ipex
-    title: How to use 🤗 Accelerate with Intel® Extension for PyTorch for cpu
-  title: How-To Guides
+  - isExpanded: true
+    sections:
+    - local: usage_guides/explore
+      title: Start Here!
+    - local: usage_guides/model_size_estimator
+      title: Model memory estimator
+    - local: usage_guides/quantization
+      title: Model quantization
+    - local: usage_guides/tracking
+      title: Experiment trackers
+    - local: usage_guides/profiler
+      title: Profiler
+    - local: usage_guides/checkpoint
+      title: Checkpointing
+    - local: basic_tutorials/troubleshooting
+      title: Troubleshoot
+    - local: usage_guides/training_zoo
+      title: Example Zoo
+    title: Accelerate
+  - isExpanded: true
+    sections:
+    - local: usage_guides/gradient_accumulation
+      title: Gradient accumulation
+    - local: usage_guides/local_sgd
+      title: Local SGD
+    - local: usage_guides/low_precision_training
+      title: Low precision (FP8) training
+    - local: usage_guides/deepspeed
+      title: DeepSpeed
+    - local: usage_guides/deepspeed_multiple_model
+      title: Using multiple models with DeepSpeed
+    - local: usage_guides/ddp_comm_hook
+      title: DDP Communication Hooks
+    - local: usage_guides/fsdp
+      title: Fully Sharded Data Parallel
+    - local: usage_guides/megatron_lm
+      title: Megatron-LM
+    - local: usage_guides/sagemaker
+      title: Amazon SageMaker
+    - local: usage_guides/mps
+      title: Apple M1 GPUs
+    - local: usage_guides/ipex
+      title: IPEX training with CPU
+    title: Training
+  - isExpanded: true
+    sections:
+    - local: usage_guides/big_modeling
+      title: Big Model Inference
+    - local: usage_guides/distributed_inference
+      title: Distributed inference
+    title: Inference
+  title: How to guides
 - sections:
+  - local: concept_guides/internal_mechanism
+    title: Accelerate's internal mechanism
   - local: concept_guides/big_model_inference
     title: Loading big models into memory
   - local: concept_guides/performance
@@ -63,34 +84,42 @@
     title: Executing and deferring jobs
   - local: concept_guides/gradient_synchronization
     title: Gradient synchronization
+  - local: concept_guides/fsdp_and_deepspeed
+    title: FSDP vs DeepSpeed
+  - local: concept_guides/low_precision_training
+    title: Low precision training methods
   - local: concept_guides/training_tpu
-    title: TPU best practices
+    title: Training on TPUs
   title: Concepts and fundamentals
 - sections: 
   - local: package_reference/accelerator
-    title: Main Accelerator class
+    title: Accelerator
   - local: package_reference/state
-    title: Stateful configuration classes
+    title: Stateful classes
   - local: package_reference/cli
     title: The Command Line
   - local: package_reference/torch_wrappers
-    title: Torch wrapper classes
+    title: DataLoaders, Optimizers, Schedulers
   - local: package_reference/tracking
     title: Experiment trackers
   - local: package_reference/launchers
-    title: Distributed launchers
+    title: Launchers
   - local: package_reference/deepspeed
     title: DeepSpeed utilities
   - local: package_reference/logging
     title: Logging
   - local: package_reference/big_modeling
     title: Working with large models
+  - local: package_reference/inference
+    title: Pipeline parallelism
   - local: package_reference/kwargs
     title: Kwargs handlers
+  - local: package_reference/fp8
+    title: FP8
   - local: package_reference/utilities
     title: Utility functions and classes
   - local: package_reference/megatron_lm
-    title: Megatron-LM Utilities
+    title: Megatron-LM utilities
   - local: package_reference/fsdp
-    title: Fully Sharded Data Parallelism Utilities
+    title: Fully Sharded Data Parallel utilities
   title: "Reference"

docs/source/basic_tutorials/execution.md

@@ -0,0 +1,128 @@
+<!--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.
+-->
+
+# Execution process
+
+When working with distributed training systems, it is important to manage how and when processes are executed across GPUs. Some processes are completed faster than others, and some processes shouldn't begin if others haven't finished yet. Accelerate provides tools for orchestrating when processes are executed to ensure everything remains synchronized across all devices.
+
+This tutorial will teach you how to execute a process on only one machine and how to delay execution until all processes have reached a certain point.
+
+## Execute on one process
+
+Certain code only needs to be run once on a given machine, such as printing a log statement or only displaying one progress bar on the local main process.
+
+<hfoptions id="local-execution">
+<hfoption id="statements">
+
+You should use `accelerator.is_local_main_process` to indicate code that should only be executed once.
+
+```py
+from tqdm.auto import tqdm
+
+progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
+```
+
+You could also wrap a statement with `accelerator.is_local_main_process`.
+
+> [!TIP]
+> For standalone `print` statements that aren't wrapped in `accelerator.is_local_main_process`, replace `print` with Accelerate's [`~Accelerator.print`] method to only print once per process.
+
+```py
+if accelerator.is_local_main_process:
+    print("Accelerate is the best")
+```
+
+</hfoption>
+<hfoption id="function">
+
+For a function that should only be executed once, use [`~Accelerator.on_local_main_process`].
+
+```py
+@accelerator.on_local_main_process
+def do_my_thing():
+    "Something done once per server"
+    do_thing_once_per_server()
+```
+
+</hfoption>
+</hfoptions>
+
+You could also direct Accelerate to execute code once across *all processes* regardless of the number of machines. This is useful if you're uploading a final model to the Hub.
+
+<hfoptions id="main-execution">
+<hfoption id="statement">
+
+You should use `accelerator.is_main_process` to indicate code that should only be executed once across all processes.
+
+```py
+if accelerator.is_main_process:
+    repo.push_to_hub()
+```
+
+</hfoption>
+<hfoption id="function">
+
+For a function that should only be executed once across all processes, use [`~Accelerator.on_main_process`].
+
+```py
+@accelerator.on_main_process
+def do_my_thing():
+    "Something done once per server"
+    do_thing_once()
+```
+
+</hfoption>
+</hfoptions>
+
+## Execute on a specific process
+
+Accelerate can also help you execute functions that should only be executed on a specific process or a local process index.
+
+<hfoptions id="specific-execution">
+<hfoption id="specific process">
+
+Use the [`~Accelerator.on_process`] method and specify the process index to execute a function on.
+
+```py
+@accelerator.on_process(process_index=0)
+def do_my_thing():
+    "Something done on process index 0"
+    do_thing_on_index_zero()
+```
+
+</hfoption>
+<hfoption id="local process">
+
+Use the [`~Accelerator.on_local_process`] method and specify the local process index to execute a function on.
+
+```py
+@accelerator.on_local_process(local_process_idx=0)
+def do_my_thing():
+    "Something done on process index 0 on each server"
+    do_thing_on_index_zero_on_each_server()
+```
+
+</hfoption>
+</hfoptions>
+
+## Defer execution
+
+When you run your script on several GPUs at the same time, some code may be executed faster than others. You might need to wait for all processes to reach a certain point before executing the next set of instructions. For instance, you shouldn’t save a model before making sure every process is done with training.
+
+To do this, add [`~Accelerator.wait_for_everyone`] in your code. This blocks all processes that have finished first from continuing until all remaining processes have reached the same point (this has no effect if you're running on a single GPU or CPU).
+
+```py
+accelerator.wait_for_everyone()
+```

docs/source/basic_tutorials/install.md

@@ -13,31 +13,29 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Installation and Configuration
+# Installation
 
-Before you start, you will need to setup your environment, install the appropriate packages, and configure 🤗 Accelerate. 🤗 Accelerate is tested on **Python 3.8+**.
+Before you start, you will need to setup your environment, install the appropriate packages, and configure Accelerate. Accelerate is tested on **Python 3.8+**.
 
-## Installing 🤗 Accelerate
+Accelerate is available on pypi and conda, as well as on GitHub. Details to install from each are below:
 
-🤗 Accelerate is available on pypi and conda, as well as on GitHub. Details to install from each are below:
+## pip
 
-### pip 
-
-To install 🤗 Accelerate from pypi, perform:
+To install Accelerate from pypi, perform:
 
 ```bash
 pip install accelerate
 ```
 
-### conda
+## conda
 
-🤗 Accelerate can also be installed with conda with:
+Accelerate can also be installed with conda with:
 
 ```bash
 conda install -c conda-forge accelerate
 ```
 
-### Source
+## Source
 
 New features are added every day that haven't been released yet. To try them out yourself, install
 from the GitHub repository:
@@ -56,9 +54,9 @@ cd accelerate
 pip install -e .
 ```
 
-## Configuring 🤗 Accelerate
+## Configuration
 
-After installing, you need to configure 🤗 Accelerate for how the current system is setup for training. 
+After installing, you need to configure Accelerate for how the current system is setup for training. 
 To do so run the following and answer the questions prompted to you:
 
 ```bash
@@ -70,7 +68,8 @@ To write a barebones configuration that doesn't include options such as DeepSpee
 ```bash
 python -c "from accelerate.utils import write_basic_config; write_basic_config(mixed_precision='fp16')"
 ```
-🤗 Accelerate will automatically utilize the maximum number of GPUs available and set the mixed precision mode.
+
+Accelerate will automatically utilize the maximum number of GPUs available and set the mixed precision mode.
 
 To check that your configuration looks fine, run:
 
@@ -80,23 +79,36 @@ accelerate env
 
 An example output is shown below, which describes two GPUs on a single machine with no mixed precision being used:
 
+
 ```bash
-- `Accelerate` version: 0.11.0.dev0
-- Platform: Linux-5.10.0-15-cloud-amd64-x86_64-with-debian-11.3
-- Python version: 3.7.12
-- Numpy version: 1.19.5
-- PyTorch version (GPU?): 1.12.0+cu102 (True)
+- `Accelerate` version: 1.2.0.dev0
+- Platform: Linux-6.8.0-47-generic-x86_64-with-glibc2.35
+- `accelerate` bash location: /home/zach/miniconda3/envs/accelerate/bin/accelerate
+- Python version: 3.10.13
+- Numpy version: 1.26.4
+- PyTorch version (GPU?): 2.5.1+cu124 (True)
+- PyTorch XPU available: False
+- PyTorch NPU available: False
+- PyTorch MLU available: False
+- PyTorch MUSA available: False
+- System RAM: 187.91 GB
+- GPU type: NVIDIA GeForce RTX 4090
 - `Accelerate` default config:
         - compute_environment: LOCAL_MACHINE
         - distributed_type: MULTI_GPU
         - mixed_precision: no
         - use_cpu: False
+        - debug: False
         - num_processes: 2
         - machine_rank: 0
         - num_machines: 1
-        - main_process_ip: None
-        - main_process_port: None
+        - gpu_ids: all
+        - rdzv_backend: static
+        - same_network: True
         - main_training_function: main
-        - deepspeed_config: {}
-        - fsdp_config: {}
-```
+        - enable_cpu_affinity: False
+        - downcast_bf16: no
+        - tpu_use_cluster: False
+        - tpu_use_sudo: False
+        - tpu_env: []
+```

docs/source/basic_tutorials/launch.md

@@ -13,9 +13,9 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Launching your 🤗 Accelerate scripts
+# Launching Accelerate scripts
 
-In the previous tutorial, you were introduced to how to modify your current training script to use 🤗 Accelerate.
+In the previous tutorial, you were introduced to how to modify your current training script to use Accelerate.
 The final version of that code is shown below:
 
 ```python
@@ -69,14 +69,14 @@ Next, you need to launch it with `accelerate launch`.
 <Tip warning={true}>
 
   It's recommended you run `accelerate config` before using `accelerate launch` to configure your environment to your liking. 
-  Otherwise 🤗 Accelerate will use very basic defaults depending on your system setup.
+  Otherwise Accelerate will use very basic defaults depending on your system setup.
 
 </Tip>
 
 
 ## Using accelerate launch
 
-🤗 Accelerate has a special CLI command to help you launch your code in your system through `accelerate launch`.
+Accelerate has a special CLI command to help you launch your code in your system through `accelerate launch`.
 This command wraps around all of the different commands needed to launch your script on various platforms, without you having to remember what each of them is.
 
 <Tip>
@@ -97,11 +97,14 @@ Since this runs the various torch spawn methods, all of the expected environment
 For example, here is how to use `accelerate launch` with a single GPU:
 
 ```bash
+# for cuda device:
 CUDA_VISIBLE_DEVICES="0" accelerate launch {script_name.py} --arg1 --arg2 ...
+# for xpu device:
+ZE_AFFINITY_MASK="0" accelerate launch {script_name.py} --arg1 --arg2 ...
 ```
 
 You can also use `accelerate launch` without performing `accelerate config` first, but you may need to manually pass in the right configuration parameters.
-In this case, 🤗 Accelerate will make some hyperparameter decisions for you, e.g., if GPUs are available, it will use all of them by default without the mixed precision.
+In this case, Accelerate will make some hyperparameter decisions for you, e.g., if GPUs are available, it will use all of them by default without the mixed precision.
 Here is how you would use all GPUs and train with mixed precision disabled:
 
 ```bash
@@ -129,14 +132,14 @@ accelerate launch -h
 
 <Tip>
 
-  Even if you are not using 🤗 Accelerate in your code, you can still use the launcher for starting your scripts!
+  Even if you are not using Accelerate in your code, you can still use the launcher for starting your scripts!
 
 </Tip>
 
 For a visualization of this difference, that earlier `accelerate launch` on multi-gpu would look something like so with `torchrun`:
 
 ```bash
-MIXED_PRECISION="fp16" torchrun --nproc_per_node=2 --num_machines=1 {script_name.py} {--arg1} {--arg2} ...
+MIXED_PRECISION="fp16" torchrun --nproc_per_node=2 --nnodes=1 {script_name.py} {--arg1} {--arg2} ...
 ```
 
 You can also launch your script utilizing the launch CLI as a python module itself, enabling the ability to pass in other python-specific
@@ -153,6 +156,15 @@ the below example enabling unbuffered stdout and stderr:
 python -u -m accelerate.commands.launch --num_processes=2 {script_name.py} {--arg1} {--arg2}
 ```
 
+<Tip>
+
+  You can run your code on CPU as well! This is helpful for debugging and testing purposes on toy models and datasets. 
+
+```bash
+accelerate launch --cpu {script_name.py} {--arg1} {--arg2}
+```  
+
+</Tip>
 
 ## Why you should always use `accelerate config`
 
@@ -169,7 +181,7 @@ accelerate launch {script_name.py} {--arg1} {--arg2} ...
 ## Custom Configurations
 
 As briefly mentioned earlier, `accelerate launch` should be mostly used through combining set configurations 
-made with the `accelerate config` command. These configs are saved to a `default_config.yaml` file in your cache folder for 🤗 Accelerate. 
+made with the `accelerate config` command. These configs are saved to a `default_config.yaml` file in your cache folder for Accelerate. 
 This cache folder is located at (with decreasing order of priority):
 
 - The content of your environment variable `HF_HOME` suffixed with `accelerate`.
@@ -200,3 +212,24 @@ Launching a script from the location of that custom yaml file looks like the fol
 ```bash
 accelerate launch --config_file {path/to/config/my_config_file.yaml} {script_name.py} {--arg1} {--arg2} ...
 ```
+
+## Multi-node training
+Multi-node training with Accelerate is similar to [multi-node training with torchrun](https://pytorch.org/tutorials/intermediate/ddp_series_multinode.html). The simplest way to launch a multi-node training run is to do the following:
+
+- Copy your codebase and data to all nodes. (or place them on a shared filesystem)
+- Setup your python packages on all nodes.
+- Run `accelerate config` on the main single node first. After specifying the number of nodes, you will be asked to specify the rank of each node (this will be 0 for the main/master node), along with the IP address and port for the main process. This is required for the worker nodes to communicate with the main process. Afterwards, you can copy or send this config file across all of your nodes, changing the `machine_rank` to 1, 2,3, etc. to avoid having to run the command (or just follow their directions directly for launching with `torchrun` as well)
+
+Once you have done this, you can start your multi-node training run by running `accelerate launch` (or `torchrun`) on all nodes.
+
+<Tip>
+    It is required that the command be ran on all nodes for everything to start, not just running it from the main node. You can use something like SLURM or a different process executor to wrap around this requirement and call everything from a single command.
+</Tip>
+
+<Tip>
+
+ It is recommended to use the intranet IP of your main node over the public IP for better latency. This is the `192.168.x.x` or the `172.x.x.x` address you see when you run `hostname -I` on the main node.
+
+</Tip>
+
+To get a better idea about multi-node training, check out our example for [multi-node training with FSDP](https://huggingface.co/blog/ram-efficient-pytorch-fsdp).

docs/source/basic_tutorials/migration.md

@@ -13,21 +13,11 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Migrating your code to 🤗 Accelerate
+# Add Accelerate to your code
 
-This tutorial will detail how to easily convert existing PyTorch code to use 🤗 Accelerate!
-You'll see that by just changing a few lines of code, 🤗 Accelerate can perform its magic and get you on 
-your way toward running your code on distributed systems with ease!
+Each distributed training framework has their own way of doing things which can require writing a lot of custom code to adapt it to your PyTorch training code and training environment. Accelerate offers a friendly way to interface with these distributed training frameworks without having to learn the specific details of each one. Accelerate takes care of those details for you, so you can focus on the training code and scale it to any distributed training environment.
 
-## The base training loop
-
-To begin, write out a very basic PyTorch training loop. 
-
-<Tip>
-
-    We are under the presumption that `training_dataloader`, `model`, `optimizer`, `scheduler`, and `loss_function` have been defined beforehand.
-
-</Tip>
+In this tutorial, you'll learn how to adapt your existing PyTorch code with Accelerate and get you on your way toward training on distributed systems with ease! You'll start with a basic PyTorch training loop (it assumes all the training objects like `model` and `optimizer` have been setup already) and progressively integrate Accelerate into it.
 
 ```python
 device = "cuda"
@@ -45,50 +35,44 @@ for batch in training_dataloader:
     scheduler.step()
 ```
 
-## Add in 🤗 Accelerate
+## Accelerator
+
+The [`Accelerator`] is the main class for adapting your code to work with Accelerate. It knows about the distributed setup you're using such as the number of different processes and your hardware type. This class also provides access to many of the necessary methods for enabling your PyTorch code to work in any distributed training environment and for managing and executing processes across devices.
+
+That's why you should always start by importing and creating an [`Accelerator`] instance in your script.
 
-To start using 🤗 Accelerate, first import and create an [`Accelerator`] instance:
 ```python
 from accelerate import Accelerator
 
 accelerator = Accelerator()
 ```
-[`Accelerator`] is the main force behind utilizing all the possible options for distributed training!
 
-### Setting the right device
-
-The [`Accelerator`] class knows the right device to move any PyTorch object to at any time, so you should
-change the definition of `device` to come from [`Accelerator`]:
+The [`Accelerator`] also knows which device to move your PyTorch objects to, so it is recommended to let Accelerate handle this for you.
 
 ```diff
-- device = 'cuda'
+- device = "cuda"
 + device = accelerator.device
   model.to(device)
 ```
 
-### Preparing your objects
+## Prepare PyTorch objects
 
-Next, you need to pass all of the important objects related to training into [`~Accelerator.prepare`]. 🤗 Accelerate will
-make sure everything is setup in the current environment for you to start training:
+Next, you need to prepare your PyTorch objects (model, optimizer, scheduler, etc.) for distributed training. The [`~Accelerator.prepare`] method takes care of placing your model in the appropriate container (like single GPU or multi-GPU) for your training setup, adapting the optimizer and scheduler to use Accelerate's [`~optimizer.AcceleratedOptimizer`] and [`~scheduler.AcceleratedScheduler`], and creating a new dataloader that can be sharded across processes.
 
-```
+> [!TIP]
+> Accelerate only prepares objects that inherit from their respective PyTorch classes such as `torch.optim.Optimizer`.
+
+The PyTorch objects are returned in the same order they're sent.
+
+```py
 model, optimizer, training_dataloader, scheduler = accelerator.prepare(
     model, optimizer, training_dataloader, scheduler
 )
 ```
-These objects are returned in the same order they were sent in. By default when using `device_placement=True`, all of the objects that can be sent to the right device will be.
-If you need to work with data that isn't passed to [~Accelerator.prepare] but should be on the active device, you should pass in the `device` you made earlier. 
 
-<Tip warning={true}>
+## Training loop
 
-    Accelerate will only prepare objects that inherit from their respective PyTorch classes (such as `torch.optim.Optimizer`).
-
-</Tip>
-
-### Modifying the training loop
-
-Finally, three lines of code need to be changed in the training loop. 🤗 Accelerate's DataLoader classes will automatically handle the device placement by default,
-and [`~Accelerator.backward`] should be used for performing the backward pass:
+Finally, remove the `to(device)` calls to the inputs and targets in the training loop because Accelerate's DataLoader classes automatically places them on the right device. You should also replace the usual `backward()` pass with Accelerate's [`~Accelerator.backward`] method which scales the gradients for you and uses the appropriate `backward()` method depending on your distributed setup (for example, DeepSpeed or Megatron).
 
 ```diff
 -   inputs = inputs.to(device)
@@ -99,17 +83,13 @@ and [`~Accelerator.backward`] should be used for performing the backward pass:
 +   accelerator.backward(loss)
 ```
 
-With that, your training loop is now ready to use 🤗 Accelerate!
-
-## The finished code
-
-Below is the final version of the converted code: 
+Put everything together and your new Accelerate training loop should now look like this!
 
 ```python
 from accelerate import Accelerator
-
 accelerator = Accelerator()
 
+device = accelerator.device
 model, optimizer, training_dataloader, scheduler = accelerator.prepare(
     model, optimizer, training_dataloader, scheduler
 )
@@ -124,6 +104,121 @@ for batch in training_dataloader:
     scheduler.step()
 ```
 
-## More Resources
+## Training features
 
-To check out more ways on how to migrate to 🤗 Accelerate, check out our [interactive migration tutorial](https://huggingface.co/docs/accelerate/usage_guides/explore) which showcases other items that need to be watched for when using Accelerate and how to do so quickly.
+Accelerate offers additional features - like gradient accumulation, gradient clipping, mixed precision training and more - you can add to your script to improve your training run. Let's explore these three features.
+
+### Gradient accumulation
+
+Gradient accumulation enables you to train on larger batch sizes by accumulating the gradients over multiple batches before updating the weights. This can be useful for getting around memory limitations. To enable this feature in Accelerate, specify the `gradient_accumulation_steps` parameter in the [`Accelerator`] class and add the [`~Accelerator.accumulate`] context manager to your script.
+
+```diff
++ accelerator = Accelerator(gradient_accumulation_steps=2)
+  model, optimizer, training_dataloader = accelerator.prepare(model, optimizer, training_dataloader)
+
+  for input, label in training_dataloader:
++     with accelerator.accumulate(model):
+          predictions = model(input)
+          loss = loss_function(predictions, label)
+          accelerator.backward(loss)
+          optimizer.step()
+          scheduler.step()
+          optimizer.zero_grad()
+```
+
+### Gradient clipping
+
+Gradient clipping is a technique to prevent "exploding gradients", and Accelerate offers:
+
+* [`~Accelerator.clip_grad_value_`] to clip gradients to a minimum and maximum value
+* [`~Accelerator.clip_grad_norm_`] for normalizing gradients to a certain value
+
+### Mixed precision
+
+Mixed precision accelerates training by using a lower precision data type like fp16 (half-precision) to calculate the gradients. For the best performance with Accelerate, the loss should be computed inside your model (like in Transformers models) because computations outside of the model are computed in full precision.
+
+Set the mixed precision type to use in the [`Accelerator`], and then use the [`~Accelerator.autocast`] context manager to automatically cast the values to the specified data type.
+
+> [!WARNING]
+> Accelerate enables automatic mixed precision, so [`~Accelerator.autocast`] is only needed if there are other mixed precision operations besides those performed on loss by [`~Accelerator.backward`] which already handles the scaling.
+
+```diff
++ accelerator = Accelerator(mixed_precision="fp16")
++ with accelerator.autocast():
+      loss = complex_loss_function(outputs, target)
+```
+
+## Save and load
+
+Accelerate can also save and load a *model* once training is complete or you can also save the model and optimizer *state* which could be useful for resuming training.
+
+### Model
+
+Once all processes are complete, unwrap the model with the [`~Accelerator.unwrap_model`] method before saving it because the [`~Accelerator.prepare`] method wrapped your model into the proper interface for distributed training. If you don't unwrap the model, saving the model state dictionary also saves any potential extra layers from the larger model and you won't be able to load the weights back into your base model.
+
+You should use the [`~Accelerator.save_model`] method to unwrap and save the model state dictionary. This method can also save a model into sharded checkpoints or into the [safetensors](https://hf.co/docs/safetensors/index) format.
+
+<hfoptions id="save">
+<hfoption id="single checkpoint">
+
+```py
+accelerator.wait_for_everyone()
+accelerator.save_model(model, save_directory)
+```
+
+<Tip>
+
+For models from the [Transformers](https://hf.co/docs/transformers/index) library, save the model with the [`~transformers.PreTrainedModel.save_pretrained`] method so that it can be reloaded with the [`~transformers.PreTrainedModel.from_pretrained`] method.
+
+```py
+from transformers import AutoModel
+
+unwrapped_model = accelerator.unwrap_model(model)
+unwrapped_model.save_pretrained(
+    "path/to/my_model_directory",
+    is_main_process=accelerator.is_main_process,
+    save_function=accelerator.save,
+)
+
+model = AutoModel.from_pretrained("path/to/my_model_directory")
+```
+
+</Tip>
+
+To load your weights, use the [`~Accelerator.unwrap_model`] method to unwrap the model first before loading the weights. All model parameters are references to tensors, so this loads your weights inside `model`.
+
+```py
+unwrapped_model = accelerator.unwrap_model(model)
+path_to_checkpoint = os.path.join(save_directory,"pytorch_model.bin")
+unwrapped_model.load_state_dict(torch.load(path_to_checkpoint))
+```
+
+</hfoption>
+<hfoption id="sharded checkpoint">
+
+Set `safe_serialization=True` to save the model in the safetensor format.
+
+```py
+accelerator.wait_for_everyone()
+accelerator.save_model(model, save_directory, max_shard_size="1GB", safe_serialization=True)
+```
+
+To load a sharded checkpoint or a safetensor formatted checkpoint, use the [`~accelerate.load_checkpoint_in_model`] method. This method allows you to load a checkpoint onto a specific device.
+
+```py
+load_checkpoint_in_model(unwrapped_model, save_directory, device_map={"":device})
+```
+
+</hfoption>
+</hfoptions>
+
+### State
+
+During training, you may want to save the current state of the model, optimizer, random generators, and potentially learning rate schedulers so they can be restored in the *same script*. You should add the [`~Accelerator.save_state`] and [`~Accelerator.load_state`] methods to your script to save and load states.
+
+To further customize where and how states are saved through [`~Accelerator.save_state`], use the [`~utils.ProjectConfiguration`] class. For example, if `automatic_checkpoint_naming` is enabled, each saved checkpoint is stored at `Accelerator.project_dir/checkpoints/checkpoint_{checkpoint_number}`.
+
+Any other stateful items to be stored should be registered with the [`~Accelerator.register_for_checkpointing`] method so they can be saved and loaded. Every object passed to this method to be stored must have a `load_state_dict` and `state_dict` function.
+
+> [!TIP]
+> If you have [`torchdata>=0.8.0`](https://github.com/pytorch/data/tree/main) installed, you can additionally pass `use_stateful_dataloader=True` into your [`~utils.DataLoaderConfiguration`]. This extends Accelerate's DataLoader classes with a `load_state_dict` and `state_dict` function, and makes it so `Accelerator.save_state` and `Accelerator.load_state` also track how far into the training dataset it has read when persisting the model.

docs/source/basic_tutorials/notebook.md

@@ -13,7 +13,7 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Launching Multi-GPU Training from a Jupyter Environment
+# Launching distributed training from Jupyter Notebooks
 
 This tutorial teaches you how to fine tune a computer vision model with 🤗 Accelerate from a Jupyter Notebook on a distributed system.
 You will also learn how to setup a few requirements needed for ensuring your environment is configured properly, your data has been prepared properly, and finally how to launch training.
@@ -26,13 +26,13 @@ You will also learn how to setup a few requirements needed for ensuring your env
 
 ## Configuring the Environment
 
-Before any training can be performed, a 🤗 Accelerate config file must exist in the system. Usually this can be done by running the following in a terminal and answering the prompts:
+Before any training can be performed, a Accelerate config file must exist in the system. Usually this can be done by running the following in a terminal and answering the prompts:
 
 ```bash
 accelerate config
 ```
 
-However, if general defaults are fine and you are *not* running on a TPU, 🤗Accelerate has a utility to quickly write your GPU configuration into a config file via [`utils.write_basic_config`].
+However, if general defaults are fine and you are *not* running on a TPU, Accelerate has a utility to quickly write your GPU configuration into a config file via [`utils.write_basic_config`].
 
 The following code will restart Jupyter after writing the configuration, as CUDA code was called to perform this. 
 
@@ -186,7 +186,7 @@ Here is a basic training loop for the animal classification problem:
 
 <Tip>
 
-    The code has been split up to allow for explainations on each section. A full version that can be copy and pasted will be available at the end
+    The code has been split up to allow for explanations on each section. A full version that can be copy and pasted will be available at the end
 
 </Tip>
 
@@ -327,7 +327,7 @@ def training_loop(mixed_precision="fp16", seed: int = 42, batch_size: int = 64):
     # Build dataloaders
     train_dataloader, eval_dataloader = get_dataloaders(batch_size)
 
-    # Instantiate the model (you build the model here so that the seed also controls new weight initaliziations)
+    # Instantiate the model (you build the model here so that the seed also controls new weight initializations)
     model = create_model("resnet50d", pretrained=True, num_classes=len(label_to_id))
 
     # Freeze the base model
@@ -344,7 +344,7 @@ def training_loop(mixed_precision="fp16", seed: int = 42, batch_size: int = 64):
     mean = mean.to(accelerator.device)
     std = std.to(accelerator.device)
 
-    # Intantiate the optimizer
+    # Instantiate the optimizer
     optimizer = torch.optim.Adam(params=model.parameters(), lr=3e-2 / 25)
 
     # Instantiate the learning rate scheduler
@@ -430,6 +430,17 @@ args = (model, "fp16", 42, 64)
 notebook_launcher(training_loop, args, num_processes=8)
 ```
 
+To launch the training process with elasticity, enabling fault tolerance, you can use the `elastic_launch` feature provided by PyTorch. This requires setting additional parameters such as `rdzv_backend` and `max_restarts`. Here is an example of how to use `notebook_launcher` with elastic capabilities:
+
+```python
+notebook_launcher(
+    training_loop,
+    args,
+    num_processes=2,
+    max_restarts=3
+)
+```
+
 As it's running it will print the progress as well as state how many devices you ran on. This tutorial was ran with two GPUs:
 
 ```python out
@@ -443,6 +454,12 @@ epoch 4: 94.71
 
 And that's it!
 
+Please note that [`notebook_launcher`] ignores the Accelerate config file, to launch based on the config use:
+
+```bash
+accelerate launch
+```
+
 ## Debugging 
 
 A common issue when running the `notebook_launcher` is receiving a CUDA has already been initialized issue. This usually stems

docs/source/basic_tutorials/overview.md

@@ -15,10 +15,10 @@ rendered properly in your Markdown viewer.
 
 # Overview
 
-Welcome to the 🤗 Accelerate tutorials! These introductory guides will help catch you up to speed on working with 🤗 Accelerate.
+Welcome to the Accelerate tutorials! These introductory guides will help catch you up to speed on working with Accelerate.
 You'll learn how to modify your code to have it work with the API seamlessly, how to launch your script properly,
 and more!
 
 These tutorials assume some basic knowledge of Python and familiarity with the PyTorch framework.
 
-If you have any questions about 🤗 Accelerate, feel free to join and ask the community on our [forum](https://discuss.huggingface.co/c/accelerate/18).
+If you have any questions about Accelerate, feel free to join and ask the community on our [forum](https://discuss.huggingface.co/c/accelerate/18).

docs/source/basic_tutorials/tpu.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.
+-->
+
+# TPU training
+
+A [TPU (Tensor Processing Unit)](https://cloud.google.com/tpu/docs/intro-to-tpu) is a type of hardware specifically designed for training models efficiently. Accelerate supports TPU training, but there are a few things you should be aware of, namely graph compilation. This tutorial briefly discusses compilation, and for more details, take a look at the [Training on TPUs with Accelerate](../concept_guides/training_tpu) guide.
+
+## Compilation
+
+A TPU creates a graph of all the operations in the training step such as the forward pass, backward pass and optimizer step. This is why the first training step always takes a while because building and compiling this graph takes time. But once compilation is complete, it is cached and all subsequent steps are much faster.
+
+The key is to avoid compiling your code again or else training is super slow. This means all your operations must be exactly the same:
+
+* all tensors in your batches must have the same length (for example, no dynamic padding for NLP tasks)
+* your code must be static (for example, no layers with for loops that have different lengths depending on the input such as a LSTM)
+
+## Weight tying
+
+A common language model design is to tie the weights of the embedding and softmax layers. However, moving the model to a TPU (either yourself or passing it to the [`~Accelerator.prepare`] method) breaks the weight tying and you'll need to retie the weights.
+
+To add special behavior (like weight tying) in your script for TPUs, set [`~Accelerator.distributed_type`] to `DistributedType.TPU` first. Then you can use the [`~transformers.PreTrainedModel.tie_weights`] method to tie the weights.
+
+```py
+if accelerator.distributed_type == DistributedType.TPU:
+    model.tie_weights()
+```

docs/source/basic_tutorials/troubleshooting.md

@@ -0,0 +1,211 @@
+<!--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.
+-->
+
+# Troubleshoot
+
+This guide provides solutions to some issues you might encounter when using Accelerate. Not all errors are covered because Accelerate is an active library that is continuously evolving and there are many different use cases and distributed training setups. If the solutions described here don't help with your specific error, please take a look at the [Ask for help](#ask-for-help) section to learn where and how to get help.
+
+## Logging
+
+Logging can help you identify where an error is coming from. In a distributed setup with multiple processes, logging can be a challenge, but Accelerate provides the [`~accelerate.logging`] utility to ensure logs are synchronized.
+
+To troubleshoot an issue, use [`~accelerate.logging`] instead of the standard Python [`logging`](https://docs.python.org/3/library/logging.html#module-logging) module. Set the verbosity level (`INFO`, `DEBUG`, `WARNING`, `ERROR`, `CRITICAL`) with the `log_level` parameter, and then you can either:
+
+1. Export the `log_level` as the `ACCELERATE_LOG_LEVEL` environment variable.
+2. Pass the `log_level` directly to `get_logger`.
+
+For example, to set `log_level="INFO"`:
+
+```py
+from accelerate.logging import get_logger
+
+logger = get_logger(__name__, log_level="DEBUG")
+```
+
+By default, the log is called on main processes only. To call it on all processes, pass `main_process_only=False`.
+If a log should be called on all processes and in order, also pass `in_order=True`.
+
+```py
+from accelerate.logging import get_logger
+
+logger = get_logger(__name__, log_level="DEBUG")
+# log all processes
+logger.debug("thing_to_log", main_process_only=False)
+# log all processes in order
+logger.debug("thing_to_log", main_process_only=False, in_order=True)
+```
+
+## Hanging code and timeout errors
+
+There can be many reasons why your code is hanging. Let's take a look at how to solve some of the most common issues that can cause your code to hang.
+
+### Mismatched tensor shapes
+
+Mismatched tensor shapes is a common issue that can cause your code to hang for a significant amount of time on a distributed setup.
+
+When running scripts in a distributed setup, functions such as [`Accelerator.gather`] and [`Accelerator.reduce`] are necessary to grab tensors across devices to collectively perform operations on them. These (and other) functions rely on `torch.distributed` to perform a `gather` operation, which requires tensors to have the **exact same shape** across all processes. When the tensor shapes don't match, your code hangs and you'll eventually hit a timeout exception.
+
+You can use Accelerate's operational debug mode to immediately catch this issue. We recommend enabling this mode during the `accelerate config` setup, but you can also enable it from the CLI, as an environment variable, or by manually editing the `config.yaml` file.
+
+<hfoptions id="mismatch">
+<hfoption id="CLI">
+
+```bash
+accelerate launch --debug {my_script.py} --arg1 --arg2
+```
+
+</hfoption>
+<hfoption id="environment variable">
+
+If enabling debug mode as an environment variable, you don't need to call `accelerate launch`.
+
+```bash
+ACCELERATE_DEBUG_MODE="1" torchrun {my_script.py} --arg1 --arg2
+```
+
+</hfoption>
+<hfoption id="config.yaml">
+
+Add `debug: true` to your `config.yaml` file.
+
+```yaml
+compute_environment: LOCAL_MACHINE
+debug: true
+```
+
+</hfoption>
+</hfoptions>
+
+Once you enable debug mode, you should get a traceback that points to the tensor shape mismatch issue.
+
+```py
+Traceback (most recent call last):
+  File "/home/zach_mueller_huggingface_co/test.py", line 18, in <module>
+    main()
+  File "/home/zach_mueller_huggingface_co/test.py", line 15, in main
+    broadcast_tensor = broadcast(tensor)
+  File "/home/zach_mueller_huggingface_co/accelerate/src/accelerate/utils/operations.py", line 303, in wrapper
+accelerate.utils.operations.DistributedOperationException:
+
+Cannot apply desired operation due to shape mismatches. All shapes across devices must be valid.
+
+Operation: `accelerate.utils.operations.broadcast`
+Input shapes:
+  - Process 0: [1, 5]
+  - Process 1: [1, 2, 5]
+```
+
+### Early stopping
+
+For early stopping in distributed training, if each process has a specific stopping condition (e.g. validation loss), it may not be synchronized across all processes. As a result, a break can happen on process 0 but not on process 1 which will cause your code to hang indefinitely until a timeout occurs.
+
+If you have early stopping conditionals, use the `set_trigger` and `check_trigger` methods to make sure all the processes
+are ended correctly.
+
+```py
+# Assume `should_do_breakpoint` is a custom defined function that returns a conditional, 
+# and that conditional might be true only on process 1
+if should_do_breakpoint(loss):
+    accelerator.set_trigger()
+
+# Later in the training script when we need to check for the breakpoint
+if accelerator.check_trigger():
+    break
+```
+
+### Low kernel versions on Linux
+
+On Linux with kernel version < 5.5, hanging processes have been reported. To avoid this problem, upgrade your system to a later kernel version.
+
+### MPI
+
+If your distributed CPU training job using MPI is hanging, ensure that you have
+[passwordless SSH](https://www.open-mpi.org/faq/?category=rsh#ssh-keys) setup (using keys) between the nodes. This means
+that for all nodes in your hostfile, you should to be able to SSH from one node to another without being prompted for a password.
+
+Next, try to run the `mpirun` command as a sanity check. For example, the command below should print out the
+hostnames for each of the nodes.
+
+```bash
+mpirun -f hostfile -n {number of nodes} -ppn 1 hostname
+```
+
+## Out-of-Memory
+
+One of the most frustrating errors when it comes to running training scripts is hitting "Out-of-Memory" on devices like CUDA, XPU or CPU. The entire script needs to be restarted and any progress is lost.
+
+To address this problem, Accelerate provides the [`find_executable_batch_size`] utility that is heavily based on [toma](https://github.com/BlackHC/toma).
+This utility retries code that fails due to OOM (out-of-memory) conditions and automatically lowers batch sizes. For each OOM condition, the algorithm decreases the batch size by half and retries the code until it succeeds.
+
+To use [`find_executable_batch_size`], restructure your training function to include an inner function with `find_executable_batch_size` and build your dataloaders inside it. At a minimum, this only takes 4 new lines of code.
+
+<Tip warning={true}> 
+
+The inner function **must** take batch size as the first parameter, but we do not pass one to it when called. The wrapper will handles this for you. Any object (models, optimizers) that consumes device memory and is passed to the [`Accelerator`] also **must** be declared inside the inner function.
+
+</Tip>
+
+```diff
+def training_function(args):
+    accelerator = Accelerator()
+
++   @find_executable_batch_size(starting_batch_size=args.batch_size)
++   def inner_training_loop(batch_size):
++       nonlocal accelerator # Ensure they can be used in our context
++       accelerator.free_memory() # Free all lingering references
+        model = get_model()
+        model.to(accelerator.device)
+        optimizer = get_optimizer()
+        train_dataloader, eval_dataloader = get_dataloaders(accelerator, batch_size)
+        lr_scheduler = get_scheduler(
+            optimizer, 
+            num_training_steps=len(train_dataloader)*num_epochs
+        )
+        model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
+            model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
+        )
+        train(model, optimizer, train_dataloader, lr_scheduler)
+        validate(model, eval_dataloader)
++   inner_training_loop()
+```
+
+## Non-reproducible results between device setups
+
+If you changed the device setup and observe different model performance, it is likely you didn't update your script when moving from one setup to another. Even if you're using the same script with the same batch size, the results will still be different on a TPU, multi-GPU, and single GPU.
+
+For example, if you were training on a single GPU with a batch size of 16 and you move to a dual GPU setup, you need to change the batch size to 8 to have the same effective batch size. This is because when training with Accelerate, the batch size passed to the dataloader is the **batch size per GPU**.
+
+To make sure you can reproduce the results between the setups, make sure to use the same seed, adjust the batch size accordingly, and consider scaling the learning rate.
+
+For more details and a quick reference for batch sizes, check out the [Comparing performance between different device setups](../concept_guides/performance) guide.
+
+## Performance issues on different GPUs
+
+If your multi-GPU setup consists of different GPUs, you may encounter some performance issues:
+
+- There may be an imbalance in GPU memory between the GPUs. In this case, the GPU with the smaller memory will limit the batch size or the size of the model that can be loaded onto the GPUs.
+- If you are using GPUs with different performance profiles, the performance will be driven by the slowest GPU you are using because the other GPUs will have to wait for it to complete its workload.
+
+Vastly different GPUs within the same setup can lead to performance bottlenecks.
+
+## Ask for help
+
+If none of the solutions and advice here helped resolve your issue, you can always reach out to the community and Accelerate team for help.
+
+- Ask for help on the Hugging Face forums by posting your question in the [Accelerate category](https://discuss.huggingface.co/c/accelerate/18). Make sure to write a descriptive post with relevant context about your setup and reproducible code to maximize the likelihood that your problem is solved!
+
+- Post a question on [Discord](http://hf.co/join/discord), and let the team and the community help you.
+
+- Create an Issue on the Accelerate [GitHub repository](https://github.com/huggingface/accelerate/issues) if you think you've found a bug related to the library. Include context regarding the bug and details about your distributed setup to help us better figure out what's wrong and how we can fix it.

docs/source/concept_guides/big_model_inference.md

@@ -13,7 +13,7 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Handling big models for inference
+# Loading big models into memory
 
 When loading a pre-trained model in PyTorch, the usual workflow looks like this:
 
@@ -46,7 +46,7 @@ This API is quite new and still in its experimental stage. While we strive to pr
 
 ### Instantiating an empty model
 
-The first tool 🤗 Accelerate introduces to help with big models is a context manager [`init_empty_weights`] that helps you initialize a model without using any RAM so that step 1 can be done on models of any size. Here is how it works:
+The first tool Accelerate introduces to help with big models is a context manager [`init_empty_weights`] that helps you initialize a model without using any RAM so that step 1 can be done on models of any size. Here is how it works:
 
 ```py
 from accelerate import init_empty_weights
@@ -74,7 +74,7 @@ initializes an empty model with a bit more than 100B parameters. Behind the scen
 
 It's possible your model is so big that even a single copy won't fit in RAM. That doesn't mean it can't be loaded: if you have one or several GPUs, this is more memory available to store your model. In this case, it's better if your checkpoint is split into several smaller files that we call checkpoint shards.
 
-🤗 Accelerate will handle sharded checkpoints as long as you follow the following format: your checkpoint should be in a folder, with several files containing the partial state dicts, and there should be an index in the JSON format that contains a dictionary mapping parameter names to the file containing their weights. You can easily shard your model with [`~Accelerator.save_model`]. For instance, we could have a folder containing:
+Accelerate will handle sharded checkpoints as long as you follow the following format: your checkpoint should be in a folder, with several files containing the partial state dicts, and there should be an index in the JSON format that contains a dictionary mapping parameter names to the file containing their weights. You can easily shard your model with [`~Accelerator.save_model`]. For instance, we could have a folder containing:
 
 ```bash
 first_state_dict.bin
@@ -97,9 +97,9 @@ and `first_state_dict.bin` containing the weights for `"linear1.weight"` and `"l
 
 ### Loading weights
 
-The second tool 🤗 Accelerate introduces is a function [`load_checkpoint_and_dispatch`], that will allow you to load a checkpoint inside your empty model. This supports full checkpoints (a single file containing the whole state dict) as well as sharded checkpoints. It will also automatically dispatch those weights across the devices you have available (GPUs, CPU RAM), so if you are loading a sharded checkpoint, the maximum RAM usage will be the size of the biggest shard.
+The second tool Accelerate introduces is a function [`load_checkpoint_and_dispatch`], that will allow you to load a checkpoint inside your empty model. This supports full checkpoints (a single file containing the whole state dict) as well as sharded checkpoints. It will also automatically dispatch those weights across the devices you have available (GPUs, CPU RAM), so if you are loading a sharded checkpoint, the maximum RAM usage will be the size of the biggest shard.
 
-If you want to use big model inference with 🤗 Transformers models, check out this [documentation](https://huggingface.co/docs/transformers/main/en/main_classes/model#large-model-loading).
+If you want to use big model inference with Transformers models, check out this [documentation](https://huggingface.co/docs/transformers/main/en/main_classes/model#large-model-loading).
 
 Here is how we can use this to load the [GPT2-1.5B](https://huggingface.co/marcsun13/gpt2-xl-linear-sharded) model.
 
@@ -145,7 +145,7 @@ model = load_checkpoint_and_dispatch(
 )
 ```
 
-By passing `device_map="auto"`, we tell 🤗 Accelerate to determine automatically where to put each layer of the model depending on the available resources:
+By passing `device_map="auto"`, we tell Accelerate to determine automatically where to put each layer of the model depending on the available resources:
 - first, we use the maximum space available on the GPU(s)
 - if we still need space, we store the remaining weights on the CPU
 - if there is not enough RAM, we store the remaining weights on the hard drive as memory-mapped tensors
@@ -154,12 +154,12 @@ By passing `device_map="auto"`, we tell 🤗 Accelerate to determine automatical
 #### `no_split_module_classes`
 
 This parameter will indicate that some of the modules with the name `"Block"` should not be split across different devices. You should set here all blocks that 
-include a residutal connection of some kind.
+include a residual connection of some kind.
 
 
 #### The `device_map`
 
-You can see the `device_map` that 🤗 Accelerate picked by accessing the `hf_device_map` attribute of your model:
+You can see the `device_map` that Accelerate picked by accessing the `hf_device_map` attribute of your model:
 
 ```py
 model.hf_device_map
@@ -210,7 +210,7 @@ outputs = model.generate(x1, max_new_tokens=10, do_sample=False)[0]
 tokenizer.decode(outputs.cpu().squeeze())
 ```
 
-Behind the scenes, 🤗 Accelerate added hooks to the model, so that:
+Behind the scenes, Accelerate added hooks to the model, so that:
 - at each layer, the inputs are put on the right device (so even if your model is spread across several GPUs, it works)
 - for the weights offloaded on the CPU, they are put on a GPU just before the forward pass and cleaned up just after
 - for the weights offloaded on the hard drive, they are loaded in RAM then put on a GPU just before the forward pass and cleaned up just after
@@ -225,7 +225,7 @@ This way, your model can run for inference even if it doesn't fit on one of the
 
 ### Designing a device map
 
-You can let 🤗 Accelerate handle the device map computation by setting `device_map` to one of the supported options (`"auto"`, `"balanced"`, `"balanced_low_0"`, `"sequential"`) or create one yourself if you want more control over where each layer should go.
+You can let Accelerate handle the device map computation by setting `device_map` to one of the supported options (`"auto"`, `"balanced"`, `"balanced_low_0"`, `"sequential"`) or create one yourself if you want more control over where each layer should go.
 
 <Tip>
 
@@ -295,11 +295,44 @@ device_map = {"block1": 0, "block2.linear1": 1, "block2.linear2": 1}
 
 </Tip>
 
+## CPU offload only
+
+If you want to offload your model on CPU, you can use [`cpu_offload`]. As a result, all parameters of the model will be offloaded and only one copy of the state dict of the model will be kept. During the forward pass, parameters will be extracted from that state dict and put on the execution device and passed as they are needed, then offloaded again. 
+
+```python
+cpu_offload(model, execution_device)
+```
+
+You can also use [`cpu_offload_with_hook`]. This function will offloads a model on the CPU and puts it back to an execution device when executed. The difference with [`cpu_offload`] is that the model stays on the execution device after the forward and is only offloaded again when the `offload` method of the returned `hook` is called. Furthermore, [`cpu_offload_with_hook`] is more performant but less memory saving. It is useful for pipelines running a model in a loop:
+
+```python
+model_1, hook_1 = cpu_offload_with_hook(model_1, execution_device)
+model_2, hook_2 = cpu_offload_with_hook(model_2, execution_device, prev_module_hook=hook_1)
+model_3, hook_3 = cpu_offload_with_hook(model_3, execution_device, prev_module_hook=hook_2)
+
+hid_1 = model_1(input)
+for i in range(50):
+    # model1 is offloaded on the CPU at the first iteration, model 2 stays on the GPU for this whole loop.
+    hid_2 = model_2(hid_1)
+# model2 is offloaded to the CPU just before this forward.
+hid_3 = model_3(hid_3)
+
+# For model3, you need to manually call the hook offload method.
+hook_3.offload()
+```
+
+## Disk offload only
+
+To perform disk offload, you can use [`disk_offload`]. As a result, all parameters of the model will be offloaded as memory-mapped array in a given folder. During the forward pass, parameters will be accessed from that folder and put on the execution device passed as they are needed, then offloaded again.
+
+```python
+disk_offload(model, offload_dir, execution_device)
+```
+
 ## Limits and further development
 
 We are aware of the current limitations in the API:
 
-- While this could theoretically work on just one CPU with potential disk offload, you need at least one GPU to run this API. This will be fixed in further development.
 - [`infer_auto_device_map`] (or `device_map="auto"` in [`load_checkpoint_and_dispatch`]) tries to maximize GPU and CPU RAM it sees available when you execute it. While PyTorch is very good at managing GPU RAM efficiently (and giving it back when not needed), it's not entirely true with Python and CPU RAM. Therefore, an automatically computed device map might be too intense on the CPU. Move a few modules to the disk device if you get crashes due to a lack of RAM.
 - [`infer_auto_device_map`] (or `device_map="auto"` in [`load_checkpoint_and_dispatch`]) attributes devices sequentially (to avoid moving things back and forth) so if your first layer is bigger than the size of the GPU you have, it will end up with everything on the CPU/Disk.
 - [`load_checkpoint_and_dispatch`] and [`load_checkpoint_in_model`] do not perform any check on the correctness of your state dict compared to your model at the moment (this will be fixed in a future version), so you may get some weird errors if trying to load a checkpoint with mismatched or missing keys.

docs/source/concept_guides/deferring_execution.md

@@ -13,9 +13,9 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Deferring Executions
+# Executing and deferring jobs
 
-When you run your usual script, instructions are executed in order. Using 🤗 Accelerate to deploy your script on several
+When you run your usual script, instructions are executed in order. Using Accelerate to deploy your script on several
 GPUs at the same time introduces a complication: while each process executes all instructions in order, some may be
 faster than others.
 
@@ -127,4 +127,4 @@ for (x,y) in data_loader:
     # Later in the training script when we need to check for the breakpoint
     if accelerator.check_trigger():
         break
-```
+```

docs/source/concept_guides/fsdp_and_deepspeed.md

@@ -0,0 +1,192 @@
+<!--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.
+-->
+
+# FSDP vs DeepSpeed
+
+Accelerate offers flexibilty of training frameworks, by integrating two extremely powerful tools for distributed training, namely [Pytorch FSDP](../usage_guides/fsdp) and [Microsoft DeepSpeed](../usage_guides/deepspeed). The aim of this tutorial is to draw parallels, as well as to outline potential differences, to empower the user to switch seamlessly between these two frameworks.
+
+<Tip>
+
+  To switch between the frameworks, we recommend launching code `accelerate launch` passing in the correct config file with `--config_file`, or passing in the respective arguments directly for [FSDP and DeepSpeed](../package_reference/cli#accelerate-launch) .
+
+  Example Accelerate configurations can be found here for [DeepSpeed](../usage_guides/deepspeed#accelerate-deepspeed-plugin) and [FSDP](../usage_guides/fsdp#how-it-works-out-of-the-box), or in the [example zoo under "Launch Configurations"](../usage_guides/explore)
+ 
+</Tip>
+
+<Tip warning={true}>
+
+This tutorial is for single-node, multi-GPU, scenarios only.
+
+</Tip>
+
+## Configuring Functionalities
+
+Model tensors are split into different GPUs in an attempt to scale up model sizes; this is termed *sharding* in FSDP, and *partitioning* in DeepSpeed. FSDP sharding and DeepSpeed ZeRO (partitioning) stages are configured by `--fsdp_sharding_strategy`, and `--zero_stage`, respectively.  In particular, FSDP `FULL_SHARD` maps to DeepSpeed ZeRO stage `3`; see this [comprehensive mapping between FSDP sharding and DeepSpeed ZeRO settings](../usage_guides/fsdp#mapping-between-fsdp-sharding-strategies-and-deepspeed-zero-stages). The below table summarizes and groups similar settings:
+
+Group | Framework | Configuration | Example | Restrictions (if any)
+--|--|--|--|--
+sharding / partitioning | FSDP<br>DeepSpeed | `--fsdp_sharding_strategy`<br>`--zero_stage` | `1` (`FULL_SHARD`) <br>`3` | 
+offload | FSDP<br>DeepSpeed | `--fsdp_offload_params`<br>`--offload_param_device`<br>`--offload_optimizer_device` | `true`<br>`cpu`<br>`cpu` | all or nothing <br><br> 
+model loading | FSDP<br>DeepSpeed | <span style="white-space:nowrap;">`--fsdp_cpu_ram_efficient_loading`</span><br>`--zero3_init_flag` | `true`<br>`true` | <br>only ZeRO 3
+efficient checkpointing | FSDP<br>DeepSpeed | `--fsdp_state_dict_type`<br>`--zero3_save_16bit_model` |  `SHARDED_STATE_DICT`<br>`true` |  <br>only ZeRO 3
+weights prefetching | FSDP<br><br>DeepSpeed | `--fsdp_forward_prefetch`<br>`--fsdp_backward_prefetch`<br>None | `true`<br>`BACKWARD_PRE` | <br><br>
+model | FSDP<br><br>DeepSpeed |  `--fsdp_auto_wrap_policy`<br><span style="white-space:nowrap;">`--fsdp_transformer_layer_cls_to_wrap`</span><br>None | `TRANSFORMER_BASED_WRAP`<br><Layer Class> |<br>Usually not needed <br>Transparent to user.
+parameters summoning | FSDP<br>DeepSpeed | `--fsdp_use_orig_params`<br>None | `true` | required for `torch.compile`<br>Transparent to user
+parameters syncing | FSDP<br>DeepSpeed | `--fsdp_sync_module_states`<br>None | `true` | 
+training | FSDP<br>DeepSpeed | None<br>`--gradient_accumulation_steps`<br>`--gradient_clipping` | <br>`auto`<br>`auto` | Transparent to user
+
+For detailed descriptions of the above, refer to [`Accelerate` launch documentation](../package_reference/cli#accelerate-launch).
+
+<Tip>
+
+    To access other DeepSpeed configurations, such as mixed precision settings, 
+    you need to pass in a `--deepspeed_config_file`, see the [documentation](../usage_guides/deepspeed#deepspeed-config-file).  
+
+    DeepSpeed can be also configured via [`DeepSpeedPlugin`], e.g., `DeepSpeedPlugin.zero_stage` is equivalent of `--zero_stage`, and `DeepSpeedPlugin.hf_ds_config` can be used to pass `--deepeed_config_file.`
+
+</Tip>
+
+<Tip>
+
+    FSDP can be also configured via [`FullyShardedDataParallelPlugin`], e.g., `FullyShardedDataParallelPlugin.sharding_strategy` is equivalent of `--fsdp_sharding_strategy`.
+    
+</Tip>
+
+### Checkpointing
+
+Do note that while FSDP can be configured via `--fsdp_state_dict_type` to save either full / sharded checkpoints.
+
+<Tip>
+
+    For DeepSpeed Zero3, one could pass a `--zero3_save_16bit_model true`, which conveniently consolidates the model to a single rank and saves; this is the FSDP equivalent of `fsdp_state_dict_type: FULL_STATE_DICT`. 
+
+</Tip>
+
+<Tip warning={true}>
+
+    For large models, consolidating the model to a single rank can be very slow.
+
+</Tip>
+
+<Tip>
+
+    For quicker checkpointing, for FSDP use `fsdp_state_dict_type: SHARDED_STATE_DICT`, and for DeepSpeed Zero3 [use the `zero_to_fp32.py` script to post-convert sharded checkpoints](https://www.deepspeed.ai/tutorials/zero/#extracting-weights).
+
+
+</Tip>
+
+### Offloading
+
+FSDP only allows *all-or-nothing* offload (i.e., either offload parameters, gradients, and optimizer, or keep them all in GPU), but DeepSpeed can offload parameters and optimizer differently. Furthermore, DeepSpeed also supports [offloading to NVME](https://www.deepspeed.ai/docs/config-json/#parameter-offloading).
+
+### Prefetching
+
+FSDP allows two prefetching configurations `--fsdp_forward_prefetch` and `--fsdp_backward_prefetch` to improve overlap of comms / computation at a cost of extra memory, see [FSDP documentation](https://pytorch.org/docs/stable/fsdp.html). 
+For DeepSpeed, the prefetching will be turned on when needed, and it turns on depending on certain hyper-params like `stage3_param_persistence_threshold`, `stage3_max_reuse_distance`, etc, [that can be configured for Zero3](https://www.deepspeed.ai/docs/config-json/#parameter-offloading); `accelerate` may set these hyper-params automatically if you don't set those explicitly in the deepspeed config file.
+
+<Tip>
+
+    For FSDP set `fsdp_backward_prefetch: BACKWARD_PRE` for improved throughputs if memory allows.
+
+</Tip>
+
+### Model Loading
+
+While FSDP require an explicit `--fsdp_cpu_ram_efficient_loading true` to activate efficient model loading, `transformers` will activate the similar feature whenever DeepSpeed Zero3 is used.
+
+<Tip>
+
+    For FSDP, whenever setting `--fsdp_cpu_ram_efficient_loading true`, `accelerate` will automatically set `sync_module_states` to true. 
+    For RAM efficient loading the weights will be loaded only in a singe rank, and thus requires `sync_module_states` to broadcast weights to other ranks.
+
+</Tip>
+
+### Model
+
+FSDP requires an explicit `--fsdp_auto_wrap_policy` for the algorithm to decide how to schedule the all-gather and reduce-scatter operations. But for DeepSpeed this is transparent to the user.
+
+<Tip>
+
+    For FSDP, simply set `fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP`. With the latest [`transformers`] versions, we try our best to figure out the suitable `fsdp_transformer_layer_cls_to_wrap` for HF transformers models. However, if you get an error regarding it, please specify this.
+
+</Tip>
+
+### Parameters Summoning
+
+FSDP requires an explicit `--fsdp_use_orig_params` flag if using `torch.compile`, see [the pytorch documenation](https://pytorch.org/docs/stable/fsdp.html#module-torch.distributed.fsdp). For DeepSpeed this is transparent to the user.
+
+<Tip>
+
+    For FSDP, when using `torch.compile` please set `fsdp_use_orig_params: True`.
+
+</Tip>
+
+
+## Training
+
+Deepspeed requires explicit `--gradient_accumulation_steps` and `--gradient_clipping` flags. For FSDP this is transparent to the user.
+
+<Tip>
+
+    When using DeepSpeed, set `gradient_accumulation_steps: "auto"` and `gradient_clipping: "auto"` to automatically pick up values set in the [`Accelerator`] or [`TrainingArguments`] (if using `transformers`).
+
+</Tip>
+
+
+## On Differences in Data Precision Handling
+
+To discuss the how data precision is handled in both FSDP and Deepspeed, it is instructive to first give an overview of how model parameters are handled in these frameworks. Before the model / optimizer parameters are distributed across GPUs, parameter preparation is involved to first "flatten" them to  one-dimensional [`torch.Tensor`](https://pytorch.org/docs/stable/tensors.html#torch-tensor). The implementation of FSDP / DeepSpeed varies in the respect of the `dtype` in which these "flattened" parameters are stored, and there are ramifications with regards to how [`torch.Optimizer`](https://pytorch.org/docs/stable/optim.html#module-torch.optim) allocate their `dtype`s. The table below outlines the processes for both frameworks; the "Local" column indicates the process occurring at a per-gpu level, therefore any memory overheads by upcasting should be understood to be amortized by the number of gpus used.
+
+<Tip>
+
+    As a rule of thumb, for stable training with automatic mixed precision, all the trainable parameters have to be in `torch.float32`.
+
+</Tip>
+
+Process | Local | Framework | Details
+--|--|--|--
+Loading, i.e., [`AutoModel.from_pretrained(..., torch_dtype=torch_dtype)`] |  
+Preparation, i.e., creation of "flat params" | ✅ | FSDP<br>DeepSpeed | created in `torch_dtype`.<br> disregards `torch_dtype`, created in `float32`.
+Optimizer initialization | ✅ | FSDP<br>DeepSpeed  | creates parameters in `torch_dtype`<br> creates parameters in `float32`
+Training Step, i.e, forward, backward, reduction | | FSDP<br>DeepSpeed  | follows [`MixedPrecision`](https://pytorch.org/docs/stable/fsdp.html#torch.distributed.fsdp.MixedPrecision)<br> follows `deepspeed_config_file` mixed precision settings.
+Optimizer (Pre-Step) | ✅ | FSDP<br>DeepSpeed | upcasting (if any) to `torch_dtype`<br>upcasted to `float32`
+Optimizer (Actual Step) | ✅ | FSDP<br>DeepSpeed  | occurs in `torch_dtype` <br> occurs in `float32`.
+
+<Tip warning={true}>
+
+    Therefore when using DeepSpeed a small number of GPUs, be aware of potentially significant memory overheads due to the upcasting during preperation.
+
+</Tip>
+
+<Tip>
+
+    With FSDP, in the absence of mixed precision, it is possible to operate the [`torch.Optimizer`](https://pytorch.org/docs/stable/optim.html#module-torch.optim) in low precision `torch_dtype`, which may be helpful when using small number of GPUs. 
+
+</Tip>
+
+<Tip warning={true}>
+
+    With mixed precision, FSDP and DeepSpeed will upcast in the model preparation step (c.f. table above). But do note that FSDP will then save checkpoints in the upcasted precision; Deepspeed may still save low precision checkpoints if `--zero3_save_16bit_model` is specified.
+
+</Tip>
+
+
+To clarify the above table consider the concrete examples below; the optimizer pre- and actual step combined for brevity. With FSDP it is possible to operate in the two modes shown below, but DeepSpeed can only operate in one.
+
+Framework | Model Loading (`torch_dtype`) | Mixed Precision | Preparation (Local) | Training | Optimizer (Local)
+--|--|--|--|--|--
+FSDP | bf16 | default (none) | bf16 | bf16 | bf16
+FSDP | bf16 | bf16 | fp32 | bf16 | fp32
+DeepSpeed   | bf16 | bf16 | fp32 | bf16 | fp32

docs/source/concept_guides/gradient_synchronization.md

@@ -13,7 +13,7 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Gradient Synchronization
+# Gradient synchronization
 
 PyTorch's distributed module operates by communicating back and forth between all of the GPUs in your system.
 This communication takes time, and ensuring all processes know the states of each other happens at particular triggerpoints
@@ -28,7 +28,7 @@ from torch.nn.parallel import DistributedDataParallel
 model = nn.Linear(10, 10)
 ddp_model = DistributedDataParallel(model)
 ```
-In 🤗 Accelerate this conversion happens automatically when calling [`~Accelerator.prepare`] and passing in your model.
+In Accelerate this conversion happens automatically when calling [`~Accelerator.prepare`] and passing in your model.
 
 ```diff
 + from accelerate import Accelerator
@@ -55,8 +55,8 @@ their gradients computed, collated, and updated before moving on to the next
 batch of data.
 When performing gradient accumulation, you accumulate `n` loss gradients and
 skip `optimizer.step()` until `n` batches have been reached. As all training
-processes only need to sychronize by the time `optimizer.step()` is called,
-without any modification to your training step, this neededless inter-process
+processes only need to synchronize by the time `optimizer.step()` is called,
+without any modification to your training step, this needless inter-process
 communication can cause a significant slowdown.
 
  How can you avoid this overhead?
@@ -90,7 +90,7 @@ for index, batch in enumerate(dataloader):
         optimizer.step()
 ```
 
-In 🤗 Accelerate to make this an API that can be called no matter the training device (though it may not do anything if you are not in a distributed system!),
+In Accelerate to make this an API that can be called no matter the training device (though it may not do anything if you are not in a distributed system!),
 `ddp_model.no_sync` gets replaced with [`~Accelerator.no_sync`] and operates the same way:
 
 ```diff
@@ -167,3 +167,18 @@ As you can see, if you are not careful about how you set up your gradient synchr
 
 If you are worried about making sure everything is done properly, we highly recommend utilizing the [`~Accelerator.accumulate`] function and passing in
 `gradient_accumulation_steps` or `gradient_accumulation_plugin` to the [`Accelerator`] object so Accelerate can handle this for you.
+
+### `no_sync` requires additional GPU memory when using FSDP
+
+Be aware that not syncing gradients can have adverse effects while performing FSDP training. As it has been warned in `torch`, the [`no_sync` context manager for FSDP](https://pytorch.org/docs/stable/fsdp.html#torch.distributed.fsdp.FullyShardedDataParallel.no_sync) will require additional memory.
+
+Therefore in memory intensive situations while using FSDP, we recommend to set `sync_each_batch` to `True` in the [`~utils.GradientAccumulationPlugin`] to disable `no_sync`.
+
+See the example below where we fine-tune Mixtral (47B parameters) on 8 A100-80GB GPUs. We see that even for a modest `gradient_accumulation_steps=2` we quickly go out-of-memory (OOM) if `no_sync` is enabled. Again, this is due to additional memory overheads due to FSDP's `no_sync`. However, if `no_sync` is disabled via `sync_each_batch=True`, then the memory consumption for `gradient_accumulation_steps=16` reverts to that of `gradient_accumulation_steps=1`.
+
+| Model           | `no_sync` (accum=1) | `no_sync` (accum=2) | `no_sync` disabled (accum=16)
+| :-------------: | :-----------------: | :-----------------: | :-----------------: 
+mixtral 8x7B      | 69G                 | OOM                 | 69G
+
+> [!WARNING] 
+> Disabling `no_sync` means there _will be slowdown_ due the extra data syncs, as explained by the earlier sections of this guide.

docs/source/concept_guides/internal_mechanism.md

@@ -0,0 +1,74 @@
+<!--Copyright 2021 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.
+-->
+
+# Accelerate's internal mechanisms
+
+Internally, Accelerate works by first analyzing the environment in which the script is launched to determine which
+kind of distributed setup is used, how many different processes there are and which one the current script is in. All
+that information is stored in the [`~AcceleratorState`].
+
+This class is initialized the first time you instantiate an [`~Accelerator`] as well as performing any
+specific initialization your distributed setup needs. Its state is then uniquely shared through all instances of
+[`~state.AcceleratorState`]. (The same can also be done with the [`PartialState`], a more barebones version it inherits)
+
+Then, when calling [`~Accelerator.prepare`], the library:
+
+- wraps your model(s) in the container adapted for the distributed setup,
+- wraps your optimizer(s) in an [`~optimizer.AcceleratedOptimizer`],
+- wraps your scheduler(s) in an [`~scheduler.AcceleratedScheduler`]
+- creates a new version of your dataloader(s) in a [`~data_loader.DataLoaderShard`] or [`~data_loader.DataLoaderDispatcher`]
+
+While the model(s), optimizer(s), and scheduler(s) are just put in simple wrappers, the dataloader(s) are re-created. This is mostly
+because PyTorch does not let the user change the `batch_sampler` of a dataloader once it's been created and the
+library handles the sharding of your data between processes by changing that `batch_sampler` to yield every other
+`num_processes` batches (if enabled).
+
+The [`~data_loader.DataLoaderShard`] subclasses `DataLoader` to add the following functionality:
+
+- it synchronizes the appropriate random number generator of all processes at each new iteration, to ensure any
+  randomization (like shuffling) is done the exact same way across processes.
+- it puts the batches on the proper device before yielding them (unless you have opted out of
+  `device_placement=True`).
+  
+The [`~data_loader.DataLoaderDispatcher`] subclasses differs from the [`~data_loader.DataLoaderShard`] in that when iterating through the `DataLoader`, the data is all starting from process 0 and *then* split and sent off to each process rather than it happening at the dataset level.
+
+The random number generator synchronization will by default synchronize:
+
+- the `generator` attribute of a given sampler (like the PyTorch `RandomSampler`) for PyTorch >= 1.6
+- the main random number generator in PyTorch <=1.5.1
+
+You can choose which random number generator(s) to synchronize with the `rng_types` argument of the main
+[`Accelerator`]. In PyTorch >= 1.6, it is recommended to rely on a local `generator` to avoid
+setting the same seed in the main random number generator in all processes.
+
+<Tip warning={true}>
+
+    Synchronization of the main torch (or CUDA or XLA) random number generator will affect any other potential random
+    artifacts you could have in your dataset (like random data augmentation) in the sense that all processes will get
+    the same random numbers from the torch random modules (so will apply the same random data augmentation if it's
+    controlled by torch).
+
+</Tip>
+
+<Tip>
+
+    The randomization part of your custom sampler, batch sampler or iterable dataset should be done using a local
+    `torch.Generator` object (in PyTorch >= 1.6), see the traditional `RandomSampler`, as an example.
+
+</Tip>
+
+If you have [`torchdata>=0.8.0`](https://github.com/pytorch/data/tree/main) installed, and you have passed `use_stateful_dataloader=True` into your [`~utils.DataLoaderConfiguration`], these classes will directly inherit from `StatefulDataLoader` instead, and maintain a `state_dict`.
+
+For more details about the internals, see the [Internals page](package_reference/torch_wrappers).

docs/source/concept_guides/low_precision_training.md

@@ -0,0 +1,74 @@
+<!--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.
+-->
+
+# Low precision training methods
+
+The release of new kinds of hardware led to the emergence of new training paradigms that better utilize them. Currently, this is in the form of training
+in 8-bit precision using packages such as [TransformersEngine](https://github.com/NVIDIA/TransformerEngine) (TE) or [MS-AMP](https://github.com/Azure/MS-AMP/tree/main).
+
+For an introduction to the topics discussed today, we recommend reviewing the [low-precision usage guide](../usage_guides/low_precision_training) as this documentation will reference it regularly. 
+
+## A Quick Chart
+
+Below is a quick chart from the MS-AMP documentation showing the different bit-precisions for each solution during training:
+
+Optimization Level | Computation(GEMM) | Comm | Weight | Master Weight | Weight Gradient | Optimizer States
+-- | -- | -- | -- | -- | -- | --
+FP16 AMP | FP16 | FP32 | FP32 | N/A | FP32 | FP32+FP32
+Nvidia TE | FP8 | FP32 | FP32 | N/A | FP32 | FP32+FP32
+MS-AMP O1 | FP8 | FP8 | FP16 | N/A | FP8 | FP32+FP32
+MS-AMP O2 | FP8 | FP8 | FP16 | N/A | FP8 | FP8+FP16
+MS-AMP O3 | FP8 | FP8 | FP8 | FP16 | FP8 | FP8+FP16
+
+## `TransformersEngine`
+
+`TransformersEngine` is the first solution to trying to train in 8-bit floating point. It works by using drop-in replacement layers for certain ones in a model that utilizes their FP8-engine to reduce the number of bits (such as 32 to 8) without degrading the final accuracy of the model. 
+
+Specifically, Accelerate will find and replace the following layers with `TransformersEngine` versions:
+
+* `nn.LayerNorm` for `te.LayerNorm`
+* `nn.Linear` for `te.Linear`
+
+As a result we wind up with a model that has most of its layers in BF16, while some layers are in FP8 reducing some of the memory. 
+
+Anecdotally, we have noticed that performance gains don't really start showing when using `TransformerEngine` until a large majority of the layers
+in the model are made up of those two layers to replace. As a result, only larger models have shown performance improvements when the number of parameters is around and upwards of a few billion. 
+
+The `TransformerEngine` can receive many different arguments that customize how it performs FP8 calculations and what they do. A full list of the arguments is available below:
+
+* `margin`: The margin to use for the gradient scaling.
+* `interval`: The interval to use for how often the scaling factor is recomputed.
+* `fp8_format``: The format to use for the FP8 recipe. Must be one of `HYBRID` or `E4M3`. (Generally `HYBRID` for training, `E4M3` for evaluation)
+* `amax_history_len`: The length of the history to use for the scaling factor computation
+* `amax_compute_algo`: The algorithm to use for the scaling factor computation. Must be one of `max` or `most_recent`.
+* `override_linear_precision`: Whether or not to execute `fprop`, `dgrad`, and `wgrad` GEMMS in higher precision.
+
+You can customize each of these as part of [`utils.FP8RecipeKwargs`] to help optimize performance of your models.
+
+If we notice in the chart mentioned earlier, TE simply casts the computation layers into FP8, while everything else is in FP32. As a result this winds up utilizing the most memory but does so with the benefit of guaranteeing the least amount of loss in end accuracy during training. 
+
+## `MS-AMP`
+
+MS-AMP takes a different approach to `TransformersEngine` by providing three different optimization levels to convert more operations in FP8 or FP16.
+
+* The base optimization level (`O1`), passes communications of the weights (such as in DDP) in FP8, stores the weights of the model in FP16, and leaves the optimizer states in FP32. The main benefit of this optimization level is that we can reduce the communication bandwidth by essentially half. Additionally, more GPU memory is saved due to 1/2 of everything being cast in FP8, and the weights being cast to FP16. Notably, both the optimizer states remain in FP32.
+
+* The second optimization level (`O2`) improves upon this by also reducing the precision of the optimizer states. One is in FP8 while the other is in FP16. Generally it's been shown that this will only provide a net-gain of no degraded end accuracy, increased training speed, and reduced memory as now every state is either in FP16 or FP8. 
+
+* Finally, MS-AMP has a third optimization level (`O3`) which helps during DDP scenarios such as DeepSpeed. The weights of the model in memory are fully cast to FP8, and the master weights are now stored in FP16. This fully reduces memory by the highest factor as now not only is almost everything in FP8, only two states are left in FP16. Currently, only DeepSpeed versions up through 0.9.2 are supported, so this capability is not included in the Accelerate integration
+
+## Combining the two
+
+More experiments need to be performed but it's been noted that combining both MS-AMP and TransformersEngine can lead to the highest throughput by relying on NVIDIA's optimized FP8 operators and utilizing how MS-AMP reduces the memory overhead.

docs/source/concept_guides/performance.md

@@ -13,7 +13,7 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Comparing performance between different device setups
+# Comparing performance across distributed setups
 
 Evaluating and comparing the performance from different setups can be quite tricky if you don't know what to look for.
 For example, you cannot run the same script with the same batch size across TPU, multi-GPU, and single-GPU with Accelerate 
@@ -43,13 +43,13 @@ Why is this important? Under the hood this will set **5** different seed setting
     random.seed(seed)
     np.random.seed(seed)
     torch.manual_seed(seed)
-    torch.cuda.manual_seed_all(seed)
+    torch.cuda.manual_seed_all(seed) # or torch.xpu.manual_seed_all, etc
     # ^^ safe to call this function even if cuda is not available
-    if is_tpu_available():
+    if is_torch_xla_available():
         xm.set_rng_state(seed)
 ```
 
-The random state, numpy's state, torch, torch's cuda state, and if TPUs are available torch_xla's cuda state.
+The random state, numpy's state, torch, torch's device state, and if TPUs are available torch_xla's cuda state.
 
 ## Observed Batch Sizes 
 
@@ -74,7 +74,7 @@ In this example, there are two GPUs for "Multi-GPU" and a TPU pod with 8 workers
 
 ## Learning Rates 
 
-As noted in multiple sources[[1](https://aws.amazon.com/blogs/machine-learning/scalable-multi-node-deep-learning-training-using-gpus-in-the-aws-cloud/)][[2](https://docs.nvidia.com/clara/tlt-mi_archive/clara-train-sdk-v2.0/nvmidl/appendix/training_with_multiple_gpus.html)], the learning rate should be scaled *linearly* based on the number of devices present. The below 
+As noted in multiple sources[[1](https://aws.amazon.com/blogs/machine-learning/scalable-multi-node-deep-learning-training-using-gpus-in-the-aws-cloud/)][[2](https://docs.nvidia.com/clara/clara-train-sdk/pt/model.html#classification-models-multi-gpu-training)], the learning rate should be scaled *linearly* based on the number of devices present. The below 
 snippet shows doing so with Accelerate:
 
 <Tip>

docs/source/concept_guides/training_tpu.md

@@ -13,9 +13,9 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Training on TPUs with 🤗 Accelerate
+# Training on TPUs
 
-Training on TPUs can be slightly different from training on multi-gpu, even with 🤗 Accelerate. This guide aims to show you 
+Training on TPUs can be slightly different from training on multi-gpu, even with Accelerate. This guide aims to show you 
 where you should be careful and why, as well as the best practices in general.
 
 ## Training in a Notebook
@@ -36,7 +36,7 @@ Below is an example of a training function passed to the [`notebook_launcher`] i
 
 <Tip>
 
-    This code snippet is based off the one from the `simple_nlp_example` notebook found [here](https://github.com/huggingface/notebooks/blob/main/examples/accelerate/simple_nlp_example.ipynb) with slight 
+    This code snippet is based off the one from the `simple_nlp_example` notebook found [here](https://github.com/huggingface/notebooks/blob/main/examples/accelerate_examples/simple_nlp_example.ipynb) with slight 
     modifications for the sake of simplicity
 
 </Tip>
@@ -81,7 +81,7 @@ notebook_launcher(training_function)
 
 <Tip>
 
-    The `notebook_launcher` will default to 8 processes if 🤗 Accelerate has been configured for a TPU
+    The `notebook_launcher` will default to 8 processes if Accelerate has been configured for a TPU
 
 </Tip>
 
@@ -128,10 +128,10 @@ And finally calling the training function with:
 
 ## Mixed Precision and Global Variables 
 
-As mentioned in the [mixed precision tutorial](../usage_guides/mixed_precision), 🤗 Accelerate supports fp16 and bf16, both of which can be used on TPUs.
+As mentioned in the [mixed precision tutorial](../usage_guides/mixed_precision), Accelerate supports fp16 and bf16, both of which can be used on TPUs.
 That being said, ideally `bf16` should be utilized as it is extremely efficient to use.
 
-There are two "layers" when using `bf16` and 🤗 Accelerate on TPUs, at the base level and at the operation level. 
+There are two "layers" when using `bf16` and Accelerate on TPUs, at the base level and at the operation level. 
 
 At the base level, this is enabled when passing `mixed_precision="bf16"` to `Accelerator`, such as:
 ```python

docs/source/index.md

@@ -15,7 +15,7 @@ rendered properly in your Markdown viewer.
 
 # Accelerate
 
-🤗 Accelerate is a library that enables the same PyTorch code to be run across any distributed configuration by adding just four lines of code! In short, training and inference at scale made simple, efficient and adaptable.
+Accelerate is a library that enables the same PyTorch code to be run across any distributed configuration by adding just four lines of code! In short, training and inference at scale made simple, efficient and adaptable.
 
 ```diff
 + from accelerate import Accelerator
@@ -37,7 +37,7 @@ rendered properly in your Markdown viewer.
       scheduler.step()
 ```
 
-Built on `torch_xla` and `torch.distributed`, 🤗 Accelerate takes care of the heavy lifting, so you don't have to write any custom code to adapt to these platforms.
+Built on `torch_xla` and `torch.distributed`, Accelerate takes care of the heavy lifting, so you don't have to write any custom code to adapt to these platforms.
 Convert existing codebases to utilize [DeepSpeed](usage_guides/deepspeed), perform [fully sharded data parallelism](usage_guides/fsdp), and have automatic support for mixed-precision training! 
 
 <Tip> 
@@ -56,11 +56,11 @@ accelerate launch {my_script.py}
   <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="./basic_tutorials/overview"
       ><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">Tutorials</div>
-      <p class="text-gray-700">Learn the basics and become familiar with using 🤗 Accelerate. Start here if you are using 🤗 Accelerate for the first time!</p>
+      <p class="text-gray-700">Learn the basics and become familiar with using Accelerate. Start here if you are using Accelerate for the first time!</p>
     </a>
     <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="./usage_guides/explore"
       ><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 to help you achieve a specific goal. Take a look at these guides to learn how to use 🤗 Accelerate to solve real-world problems.</p>
+      <p class="text-gray-700">Practical guides to help you achieve a specific goal. Take a look at these guides to learn how to use Accelerate to solve real-world problems.</p>
     </a>
     <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="./concept_guides/gradient_synchronization"
       ><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>
@@ -68,7 +68,7 @@ accelerate launch {my_script.py}
    </a>
     <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="./package_reference/accelerator"
       ><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 🤗 Accelerate classes and methods work.</p>
+      <p class="text-gray-700">Technical descriptions of how Accelerate classes and methods work.</p>
     </a>
   </div>
 </div>

docs/source/package_reference/accelerator.md

@@ -15,197 +15,12 @@ rendered properly in your Markdown viewer.
 
 # Accelerator
 
-The [`Accelerator`] is the main class provided by 🤗 Accelerate. 
-It serves at the main entry point for the API. 
+The [`Accelerator`] is the main class for enabling distributed training on any type of training setup. Read the [Add Accelerator to your code](../basic_tutorials/migration) tutorial to learn more about how to add the [`Accelerator`] to your script.
 
-## Quick adaptation of your code
-
-To quickly adapt your script to work on any kind of setup with 🤗 Accelerate just:
-
-1. Initialize an [`Accelerator`] object (that we will call `accelerator` throughout this page) as early as possible in your script.
-2. Pass your dataloader(s), model(s), optimizer(s), and scheduler(s) to the [`~Accelerator.prepare`] method.
-3. Remove all the `.cuda()` or `.to(device)` from your code and let the `accelerator` handle the device placement for you. 
-
-<Tip>
-
-    Step three is optional, but considered a best practice.
-
-</Tip>
-
-4. Replace `loss.backward()` in your code with `accelerator.backward(loss)`
-5. Gather your predictions and labels before storing them or using them for metric computation using [`~Accelerator.gather`]
-
-<Tip warning={true}>
-
-    Step five is mandatory when using distributed evaluation
-    
-</Tip>
-
-In most cases this is all that is needed. The next section lists a few more advanced use cases and nice features
-you should search for and replace by the corresponding methods of your `accelerator`:
-
-## Advanced recommendations
-
-### Printing
-
-`print` statements should be replaced by [`~Accelerator.print`] to be printed once per process:
-
-```diff
-- print("My thing I want to print!")
-+ accelerator.print("My thing I want to print!")
-```
-
-### Executing processes
-
-#### Once on a single server
-
-For statements that should be executed once per server, use [`~Accelerator.is_local_main_process`]:
-
-```python
-if accelerator.is_local_main_process:
-    do_thing_once_per_server()
-```
-
-A function can be wrapped using the [`~Accelerator.on_local_main_process`] function to achieve the same 
-behavior on a function's execution:
-
-```python
-@accelerator.on_local_main_process
-def do_my_thing():
-    "Something done once per server"
-    do_thing_once_per_server()
-```
-
-#### Only ever once across all servers
-
-For statements that should only ever be executed once, use [`~Accelerator.is_main_process`]:
-
-```python
-if accelerator.is_main_process:
-    do_thing_once()
-```
-
-A function can be wrapped using the [`~Accelerator.on_main_process`] function to achieve the same 
-behavior on a function's execution:
-
-```python
-@accelerator.on_main_process
-def do_my_thing():
-    "Something done once per server"
-    do_thing_once()
-```
-
-#### On specific processes
-
-If a function should be ran on a specific overall or local process index, there are similar decorators 
-to achieve this:
-
-```python
-@accelerator.on_local_process(local_process_idx=0)
-def do_my_thing():
-    "Something done on process index 0 on each server"
-    do_thing_on_index_zero_on_each_server()
-```
-
-```python
-@accelerator.on_process(process_index=0)
-def do_my_thing():
-    "Something done on process index 0"
-    do_thing_on_index_zero()
-```
-
-### Synchronicity control
-
-Use [`~Accelerator.wait_for_everyone`] to make sure all processes join that point before continuing. (Useful before a model save for instance).
-
-### Saving and loading
-
-```python
-model = MyModel()
-model = accelerator.prepare(model)
-```
-
-Use [`~Accelerator.save_model`] instead of `torch.save` to save a model. It will remove all model wrappers added during the distributed process, get the state_dict of the model and save it. The state_dict will be in the same precision as the model being trained.
-
-```diff
-- torch.save(state_dict, "my_state.pkl")
-+ accelerator.save_model(model, save_directory)
-```
-
-[`~Accelerator.save_model`] can also save a model into sharded checkpoints or with safetensors format.
-Here is an example: 
-
-```python
-accelerator.save_model(model, save_directory, max_shard_size="1GB", safe_serialization=True)
-```
-
-#### 🤗 Transformers models
-
-If you are using models from the [🤗 Transformers](https://huggingface.co/docs/transformers/) library, you can use the `.save_pretrained()` method.
-
-```python
-from transformers import AutoModel
-
-model = AutoModel.from_pretrained("bert-base-cased")
-model = accelerator.prepare(model)
-
-# ...fine-tune with PyTorch...
-
-unwrapped_model = accelerator.unwrap_model(model)
-unwrapped_model.save_pretrained(
-    "path/to/my_model_directory",
-    is_main_process=accelerator.is_main_process,
-    save_function=accelerator.save,
-)
-```
-
-This will ensure your model stays compatible with other 🤗 Transformers functionality like the `.from_pretrained()` method.
-
-```python
-from transformers import AutoModel
-
-model = AutoModel.from_pretrained("path/to/my_model_directory")
-```
-
-### Operations
-
-Use [`~Accelerator.clip_grad_norm_`] instead of ``torch.nn.utils.clip_grad_norm_`` and [`~Accelerator.clip_grad_value_`] instead of ``torch.nn.utils.clip_grad_value``
-
-### Gradient Accumulation
-
-To perform gradient accumulation use [`~Accelerator.accumulate`] and specify a gradient_accumulation_steps. 
-This will also automatically ensure the gradients are synced or unsynced when on 
-multi-device training, check if the step should actually be performed, and auto-scale the loss:
-
-```diff
-- accelerator = Accelerator()
-+ accelerator = Accelerator(gradient_accumulation_steps=2)
-
-  for (input, label) in training_dataloader:
-+     with accelerator.accumulate(model):
-          predictions = model(input)
-          loss = loss_function(predictions, labels)
-          accelerator.backward(loss)
-          optimizer.step()
-          scheduler.step()
-          optimizer.zero_grad()
-```
-#### GradientAccumulationPlugin
-[[autodoc]] utils.GradientAccumulationPlugin
-
-
-Instead of passing `gradient_accumulation_steps` you can instantiate a GradientAccumulationPlugin and pass it to the [`Accelerator`]'s `__init__`
-as `gradient_accumulation_plugin`. You can only pass either one of `gradient_accumulation_plugin` or `gradient_accumulation_steps` passing both will raise an error.
-```diff
-from accelerate.utils import GradientAccumulationPlugin
-
-gradient_accumulation_plugin = GradientAccumulationPlugin(num_steps=2)
-- accelerator = Accelerator()
-+ accelerator = Accelerator(gradient_accumulation_plugin=gradient_accumulation_plugin)
-```
-
-In addition to the number of steps, this also lets you configure whether or not you adjust your learning rate scheduler to account for the change in steps due to accumulation.
-
-## Overall API documentation:
+## Accelerator[[api]]
 
 [[autodoc]] Accelerator
+
+## Utilities
+
+[[autodoc]] accelerate.utils.gather_object

docs/source/package_reference/big_modeling.md

@@ -15,32 +15,88 @@ rendered properly in your Markdown viewer.
 
 # Working with large models
 
-## Dispatching and Offloading Models
+## Dispatch and offload
+
+### init_empty_weights
 
 [[autodoc]] big_modeling.init_empty_weights
+
+### cpu_offload
+
 [[autodoc]] big_modeling.cpu_offload
+
+### cpu_offload_with_hook
+
+[[autodoc]] big_modeling.cpu_offload_with_hook
+
+### disk_offload
+
 [[autodoc]] big_modeling.disk_offload
+
+### dispatch_model
+
 [[autodoc]] big_modeling.dispatch_model
+
+### load_checkpoint_and_dispatch
+
 [[autodoc]] big_modeling.load_checkpoint_and_dispatch
+
+### load_checkpoint_in_model
+
 [[autodoc]] big_modeling.load_checkpoint_in_model
+
+### infer_auto_device_map
+
 [[autodoc]] utils.infer_auto_device_map
 
-## Model Hooks
+## Hooks
 
-### Hook Classes
+### ModelHook
 
 [[autodoc]] hooks.ModelHook
+
+### AlignDevicesHook
+
 [[autodoc]] hooks.AlignDevicesHook
+
+### SequentialHook
+
 [[autodoc]] hooks.SequentialHook
 
-### Adding Hooks
+## Adding Hooks
+
+### add_hook_to_module
 
 [[autodoc]] hooks.add_hook_to_module
+
+### attach_execution_device_hook
+
 [[autodoc]] hooks.attach_execution_device_hook
+
+### attach_align_device_hook
+
 [[autodoc]] hooks.attach_align_device_hook
+
+### attach_align_device_hook_on_blocks
+
 [[autodoc]] hooks.attach_align_device_hook_on_blocks
 
-### Removing Hooks
+## Removing Hooks
+
+### remove_hook_from_module
 
 [[autodoc]] hooks.remove_hook_from_module
-[[autodoc]] hooks.remove_hook_from_submodules
+
+### remove_hook_from_submodules
+
+[[autodoc]] hooks.remove_hook_from_submodules
+
+## Utilities
+
+### has_offloaded_params
+
+[[autodoc]] utils.has_offloaded_params
+
+### align_module_device
+
+[[autodoc]] utils.align_module_device

docs/source/package_reference/cli.md

@@ -145,10 +145,11 @@ values. They can also be passed in manually.
 
 The following arguments are useful for fine-tuning how available hardware should be used
 
-* `--mixed_precision {no,fp16,bf16}` (`str`) -- Whether or not to use mixed precision training. Choose between FP16 and BF16 (bfloat16) training. BF16 training is only supported on Nvidia Ampere GPUs and PyTorch 1.10 or later.
+* `--mixed_precision {no,fp16,bf16,fp8}` (`str`) -- Whether or not to use mixed precision training. Choose between FP16 and BF16 (bfloat16) training. BF16 training is only supported on Nvidia Ampere GPUs and PyTorch 1.10 or later.
 * `--num_processes NUM_PROCESSES` (`int`) -- The total number of processes to be launched in parallel.
 * `--num_machines NUM_MACHINES` (`int`) -- The total number of machines used in this training.
 * `--num_cpu_threads_per_process NUM_CPU_THREADS_PER_PROCESS` (`int`) -- The number of CPU threads per process. Can be tuned for optimal performance.
+* `--enable_cpu_affinity` (`bool`) -- Whether or not CPU affinity and balancing should be enabled. Currently only supported on NVIDIA hardware.
 
 **Training Paradigm Arguments**:
 
@@ -165,19 +166,26 @@ The following arguments are only useful when `multi_gpu` is passed or multi-gpu
 
 * `--gpu_ids` (`str`) -- What GPUs (by id) should be used for training on this machine as a comma-seperated list
 * `--same_network` (`bool`) -- Whether all machines used for multinode training exist on the same local network.
-* `--machine_rank MACHINE_RANK` (`int`) -- The rank of the machine on which this script is launched.
-* `--main_process_ip MAIN_PROCESS_IP` (`str`) -- The IP address of the machine of rank 0.
-* `--main_process_port MAIN_PROCESS_PORT` (`int`) -- The port to use to communicate with the machine of rank 0.
-* `--rdzv_backend` (`str`) -- The rendezvous method to use, such as "static" or "c10d"
+* `--machine_rank` (`int`) -- The rank of the machine on which this script is launched.
+* `--main_process_ip` (`str`) -- The IP address of the machine of rank 0.
+* `--main_process_port` (`int`) -- The port to use to communicate with the machine of rank 0.
+* `-t`, `--tee` (`str`) -- Tee std streams into a log file and also to console.
+* `--log_dir` (`str`) -- Base directory to use for log files when using torchrun/torch.distributed.run as launcher. Use with --tee to redirect std streams info log files.
+* `--role` (`str`) -- User-defined role for the workers.
+* `--rdzv_backend` (`str`) -- The rendezvous method to use, such as 'static' (the default) or 'c10d'
 * `--rdzv_conf` (`str`) -- Additional rendezvous configuration (<key1>=<value1>,<key2>=<value2>,...).
 * `--max_restarts` (`int`) -- Maximum number of worker group restarts before failing.
-* `--monitor_interval` (`float`) -- Interval, in seconds, to monitor the state of workers.
+* `--monitor_interval` (`int`) -- Interval, in seconds, to monitor the state of workers.
 
 **TPU Arguments**:
 
 The following arguments are only useful when `tpu` is passed or TPU training is configured through `accelerate config`: 
 
-* `--main_training_function MAIN_TRAINING_FUNCTION` (`str`) -- The name of the main function to be executed in your script.
+* `--tpu_cluster` (`bool`) -- Whether to use a GCP TPU pod for training.
+* `--tpu_use_sudo` (`bool`) -- Whether to use `sudo` when running the TPU training script in each pod.
+* `--vm` (`str`) -- List of single Compute VM instance names. If not provided we assume usage of instance groups. For TPU pods.
+* `--env` (`str`) -- List of environment variables to set on the Compute VM instances. For TPU pods.
+* `--main_training_function` (`str`) -- The name of the main function to be executed in your script (only for TPU training).
 * `--downcast_bf16` (`bool`) -- Whether when using bf16 precision on TPUs if both float and double tensors are cast to bfloat16 or if double tensors remain as float32.
 
 **DeepSpeed Arguments**:
@@ -188,6 +196,7 @@ The following arguments are only useful when `use_deepspeed` is passed or `deeps
 * `--zero_stage` (`int`) -- DeepSpeed's ZeRO optimization stage.
 * `--offload_optimizer_device` (`str`) -- Decides where (none|cpu|nvme) to offload optimizer states.
 * `--offload_param_device` (`str`) -- Decides where (none|cpu|nvme) to offload parameters.
+* `--offload_optimizer_nvme_path` (`str`) -- Decides Nvme Path to offload optimizer states.
 * `--gradient_accumulation_steps` (`int`) -- No of gradient_accumulation_steps used in your training script.
 * `--gradient_clipping` (`float`) -- Gradient clipping value used in your training script.
 * `--zero3_init_flag` (`str`) -- Decides Whether (true|false) to enable `deepspeed.zero.Init` for constructing massive models. Only applicable with DeepSpeed ZeRO Stage-3.
@@ -196,10 +205,11 @@ The following arguments are only useful when `use_deepspeed` is passed or `deeps
 * `--deepspeed_exclusion_filter` (`str`) -- DeepSpeed exclusion filter string when using mutli-node setup.
 * `--deepspeed_inclusion_filter` (`str`) -- DeepSpeed inclusion filter string when using mutli-node setup.
 * `--deepspeed_multinode_launcher` (`str`) -- DeepSpeed multi-node launcher to use.
+* `--deepspeed_moe_layer_cls_names` (`str`) -- comma-separated list of transformer MoE layer class names (case-sensitive) to wrap, e.g, `MixtralSparseMoeBlock` `Qwen2MoeSparseMoeBlock`, `JetMoEAttention,JetMoEBlock`
 
 **Fully Sharded Data Parallelism Arguments**:
 
-The following arguments are only useful when `use_fdsp` is passed or Fully Sharded Data Parallelism is configured through `accelerate config`:
+The following arguments are only useful when `use_fsdp` is passed or Fully Sharded Data Parallelism is configured through `accelerate config`:
 
 * `--fsdp_offload_params` (`str`) -- Decides Whether (true|false) to offload parameters and gradients to CPU.
 * `--fsdp_min_num_params` (`int`) -- FSDP's minimum number of parameters for Default Auto Wrapping.
@@ -208,6 +218,11 @@ The following arguments are only useful when `use_fdsp` is passed or Fully Shard
 * `--fsdp_transformer_layer_cls_to_wrap` (`str`) -- Transformer layer class name (case-sensitive) to wrap, e.g, `BertLayer`, `GPTJBlock`, `T5Block` ...
 * `--fsdp_backward_prefetch_policy` (`str`) -- FSDP's backward prefetch policy.
 * `--fsdp_state_dict_type` (`str`) -- FSDP's state dict type.
+* `--fsdp_forward_prefetch` (`str`) -- FSDP forward prefetch.
+* `--fsdp_use_orig_params` (`str`) -- If True, allows non-uniform `requires_grad` mixed in a FSDP unit.
+* `--fsdp_cpu_ram_efficient_loading` (`str`) -- If true, only the first process loads the pretrained model checkoint while all other processes have empty weights. When using this, `--fsdp_sync_module_states` needs to True.
+* `--fsdp_sync_module_states` (`str`) -- If true, each individually wrapped FSDP unit will broadcast module parameters from rank 0.
+* `--fsdp_activation_checkpointing` (`bool`) -- Decides Whether intermediate activations are freed during the forward pass, and a checkpoint is left as a placeholder
 
 **Megatron-LM Arguments**:
 
@@ -218,9 +233,21 @@ The following arguments are only useful when `use_megatron_lm` is passed or Mega
 * `--megatron_lm_num_micro_batches` (``) -- Megatron-LM's number of micro batches when PP degree > 1.
 * `--megatron_lm_sequence_parallelism` (``) -- Decides Whether (true|false) to enable Sequence Parallelism when TP degree > 1.
 * `--megatron_lm_recompute_activations` (``) -- Decides Whether (true|false) to enable Selective Activation Recomputation.
-* `--megatron_lm_use_distributed_optimizer` (``) -- Decides Whether (true|false) to use distributed optimizer which shards optimizer state and gradients across Data Pralellel (DP) ranks.
+* `--megatron_lm_use_distributed_optimizer` (``) -- Decides Whether (true|false) to use distributed optimizer which shards optimizer state and gradients across Data Parallel (DP) ranks.
 * `--megatron_lm_gradient_clipping` (``) -- Megatron-LM's gradient clipping value based on global L2 Norm (0 to disable).
 
+**FP8 Arguments**:
+
+* `--fp8_backend` (`str`) -- Choose a backend to train with FP8 (`te` or `msamp`)
+* `--fp8_use_autocast_during_eval` (`bool`) -- Whether to use FP8 autocast during eval mode (useful only when `--fp8_backend=te` is passed). Generally better metrics are found when this is not passed.
+* `--fp8_margin` (`int`) -- The margin to use for the gradient scaling (useful only when `--fp8_backend=te` is passed).
+* `--fp8_interval` (`int`) -- The interval to use for how often the scaling factor is recomputed (useful only when `--fp8_backend=te` is passed).
+* `--fp8_format` (`str`) -- The format to use for the FP8 recipe (useful only when `--fp8_backend=te` is passed).
+* `--fp8_amax_history_len` (`int`) -- The length of the history to use for the scaling factor computation (useful only when `--fp8_backend=te` is passed).
+* `--fp8_amax_compute_algo` (`str`) -- The algorithm to use for the scaling factor computation. (useful only when `--fp8_backend=te` is passed).
+* `--fp8_override_linear_precision` (`Tuple[bool, bool, bool]`) -- Whether or not to execute `fprop`, `dgrad`, and `wgrad` GEMMS in higher precision.
+* `--fp8_opt_level` (`str`) -- What level of 8-bit collective communication should be used with MS-AMP (useful only when `--fp8_backend=msamp` is passed)
+
 **AWS SageMaker Arguments**:
 
 The following arguments are only useful when training in SageMaker

docs/source/package_reference/deepspeed.md

@@ -13,16 +13,32 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Utilities for DeepSpeed
+# DeepSpeed utilities
+
+## DeepSpeedPlugin
+
+## get_active_deepspeed_plugin
+
+[[autodoc]] utils.get_active_deepspeed_plugin
 
 [[autodoc]] utils.DeepSpeedPlugin
 
-[[autodoc]] utils.DummyOptim
+[[autodoc]] utils.deepspeed.DummyScheduler
 
-[[autodoc]] utils.DummyScheduler
+## DeepSpeedEnginerWrapper
 
-[[autodoc]] utils.DeepSpeedEngineWrapper
+[[autodoc]] utils.deepspeed.DeepSpeedEngineWrapper
 
-[[autodoc]] utils.DeepSpeedOptimizerWrapper
+## DeepSpeedOptimizerWrapper
 
-[[autodoc]] utils.DeepSpeedSchedulerWrapper
+[[autodoc]] utils.deepspeed.DeepSpeedOptimizerWrapper
+
+## DeepSpeedSchedulerWrapper
+
+[[autodoc]] utils.deepspeed.DeepSpeedSchedulerWrapper
+
+## DummyOptim
+
+[[autodoc]] utils.deepspeed.DummyOptim
+
+## DummyScheduler

docs/source/package_reference/fp8.md

@@ -0,0 +1,38 @@
+<!--Copyright 2021 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.
+-->
+
+# FP8
+
+Below are functions and classes relative to the underlying FP8 implementation
+
+## FP8RecipeKwargs
+
+[[autodoc]] utils.FP8RecipeKwargs
+
+## convert_model
+
+[[autodoc]] utils.convert_model
+
+## has_transformer_engine_layers
+
+[[autodoc]] utils.has_transformer_engine_layers
+
+## contextual_fp8_autocast
+
+[[autodoc]] utils.contextual_fp8_autocast
+
+## apply_fp8_autowrap
+
+[[autodoc]] utils.apply_fp8_autowrap

docs/source/package_reference/fsdp.md

@@ -13,6 +13,20 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Utilities for Fully Sharded Data Parallelism
+# Fully Sharded Data Parallel utilities
 
-[[autodoc]] utils.FullyShardedDataParallelPlugin
+## enable_fsdp_ram_efficient_loading
+
+[[autodoc]] utils.enable_fsdp_ram_efficient_loading
+
+## disable_fsdp_ram_efficient_loading
+
+[[autodoc]] utils.disable_fsdp_ram_efficient_loading
+
+## merge_fsdp_weights
+
+[[autodoc]] utils.merge_fsdp_weights
+
+## FullyShardedDataParallelPlugin
+
+[[autodoc]] utils.FullyShardedDataParallelPlugin

docs/source/package_reference/inference.md

@@ -0,0 +1,22 @@
+<!--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.
+-->
+
+# Pipeline parallelism
+
+Accelerate supports pipeline parallelism for large-scale training with the PyTorch [torch.distributed.pipelining](https://pytorch.org/docs/stable/distributed.pipelining.html) API.
+
+## prepare_pippy
+
+[[autodoc]] inference.prepare_pippy

docs/source/package_reference/kwargs.md

@@ -13,7 +13,7 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Kwargs Handlers
+# Kwargs handlers
 
 The following objects can be passed to the main [`Accelerator`] to customize how some PyTorch objects
 related to distributed training or mixed precision are created.
@@ -30,6 +30,10 @@ related to distributed training or mixed precision are created.
 
 [[autodoc]] utils.FP8RecipeKwargs
 
+## ProfileKwargs
+
+[[autodoc]] utils.ProfileKwargs
+
 ## GradScalerKwargs
 
 [[autodoc]] GradScalerKwargs
@@ -37,3 +41,7 @@ related to distributed training or mixed precision are created.
 ## InitProcessGroupKwargs
 
 [[autodoc]] InitProcessGroupKwargs
+
+## KwargsHandler
+
+[[autodoc]] utils.KwargsHandler

docs/source/package_reference/launchers.md

@@ -17,6 +17,10 @@ rendered properly in your Markdown viewer.
 
 Functions for launching training on distributed processes.
 
+## notebook_launcher
 
 [[autodoc]] accelerate.notebook_launcher
+
+## debug_launcher
+
 [[autodoc]] accelerate.debug_launcher

docs/source/package_reference/logging.md

@@ -13,25 +13,9 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Logging with Accelerate
+# Logging
 
-Accelerate has its own logging utility to handle logging while in a distributed system.
-To utilize this replace cases of `logging` with `accelerate.logging`:
-```diff
-- import logging
-+ from accelerate.logging import get_logger
-- logger = logging.getLogger(__name__)
-+ logger = get_logger(__name__)
-```
-
-## Setting the log level
-
-The log level can be set with the `ACCELERATE_LOG_LEVEL` environment variable or by passing 
-`log_level` to `get_logger`:
-```python
-from accelerate.logging import get_logger
-
-logger = get_logger(__name__, log_level="INFO")
-```
+Refer to the [Troubleshooting guide](../usage_guides/troubleshooting#logging) or to the example below to learn 
+how to use Accelerate's logger. 
 
 [[autodoc]] logging.get_logger

docs/source/package_reference/megatron_lm.md

@@ -13,20 +13,36 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Utilities for Megatron-LM
+# Megatron-LM utilities
+
+## MegatronLMPlugin
 
 [[autodoc]] utils.MegatronLMPlugin
 
+## MegatronLMDummyScheduler
+
 [[autodoc]] utils.MegatronLMDummyScheduler
 
+## MegatronLMDummyDataLoader
+
 [[autodoc]] utils.MegatronLMDummyDataLoader
 
+## AbstractTrainStep
+
 [[autodoc]] utils.AbstractTrainStep
 
+## GPTTrainStep
+
 [[autodoc]] utils.GPTTrainStep
 
+## BertTrainStep
+
 [[autodoc]] utils.BertTrainStep
 
+## T5TrainStep
+
 [[autodoc]] utils.T5TrainStep
 
+## avg_losses_across_data_parallel_group
+
 [[autodoc]] utils.avg_losses_across_data_parallel_group

docs/source/package_reference/state.md

@@ -21,8 +21,14 @@ instances share the same state, which is initialized on the first instantiation.
 These classes are immutable and store information about certain configurations or 
 states.
 
+## PartialState
+
 [[autodoc]] state.PartialState
 
+## AcceleratorState
+
 [[autodoc]] state.AcceleratorState
 
+## GradientState
+
 [[autodoc]] state.GradientState

docs/source/package_reference/torch_wrappers.md

@@ -13,25 +13,36 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Wrapper classes for torch Dataloaders, Optimizers, and Schedulers
+# DataLoaders, Optimizers, and Schedulers
 
 The internal classes Accelerate uses to prepare objects for distributed training
 when calling [`~Accelerator.prepare`].
 
-## Datasets and DataLoaders
+## DataLoader utilities
 
 [[autodoc]] data_loader.prepare_data_loader
 [[autodoc]] data_loader.skip_first_batches
 
+## BatchSamplerShard
+
 [[autodoc]] data_loader.BatchSamplerShard
+
+## IterableDatasetShard
+
 [[autodoc]] data_loader.IterableDatasetShard
+
+## DataLoaderShard
+
 [[autodoc]] data_loader.DataLoaderShard
+
+## DataLoaderDispatcher
+
 [[autodoc]] data_loader.DataLoaderDispatcher
 
-## Optimizers 
+## AcceleratedOptimizer
 
 [[autodoc]] optimizer.AcceleratedOptimizer
 
-## Schedulers 
+## AcceleratedScheduler
 
 [[autodoc]] scheduler.AcceleratedScheduler

docs/source/package_reference/tracking.md

@@ -13,21 +13,38 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Experiment Tracking
+# Experiment Trackers
 
-## The Base Tracker Class
+## GeneralTracker
 
 [[autodoc]] tracking.GeneralTracker
 
-## Integrated Trackers
+## TensorBoardTracker
 
 [[autodoc]] tracking.TensorBoardTracker
     - __init__
+
+## WandBTracker
+
 [[autodoc]] tracking.WandBTracker
     - __init__
+
+## CometMLTracker
+
 [[autodoc]] tracking.CometMLTracker
     - __init__
+
+## AimTracker
+
 [[autodoc]] tracking.AimTracker
     - __init__
+
+## MLflowTracker
+
 [[autodoc]] tracking.MLflowTracker
     - __init__
+
+## ClearMLTracker
+
+[[autodoc]] tracking.ClearMLTracker
+    - __init__

docs/source/package_reference/utilities.md

@@ -13,7 +13,7 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Helpful Utilities
+# Utility functions and classes
 
 Below are a variety of utility functions that 🤗 Accelerate provides, broken down by use-case. 
 
@@ -40,6 +40,12 @@ The following are constants used when utilizing [`Accelerator.save_model`]
 
 These are basic dataclasses used throughout 🤗 Accelerate and they can be passed in as parameters.
 
+### Standalone
+
+These are standalone dataclasses used for checks, such as the type of distributed system being used
+
+[[autodoc]] utils.ComputeEnvironment
+
 [[autodoc]] utils.DistributedType
 
 [[autodoc]] utils.DynamoBackend
@@ -48,12 +54,30 @@ These are basic dataclasses used throughout 🤗 Accelerate and they can be pass
 
 [[autodoc]] utils.PrecisionType
 
-[[autodoc]] utils.ProjectConfiguration
+[[autodoc]] utils.RNGType
+
+[[autodoc]] utils.SageMakerDistributedType
+
+### Kwargs
+
+These are configurable arguments for specific interactions throughout the PyTorch ecosystem that Accelerate handles under the hood.
+
+[[autodoc]] utils.AutocastKwargs
+
+[[autodoc]] utils.DistributedDataParallelKwargs
+
+[[autodoc]] utils.FP8RecipeKwargs
+
+[[autodoc]] utils.GradScalerKwargs
+
+[[autodoc]] utils.InitProcessGroupKwargs
+
+[[autodoc]] utils.KwargsHandler
 
 ## Plugins
 
 These are plugins that can be passed to the [`Accelerator`] object. While they are defined elsewhere in the documentation, 
-for convience all of them are available to see here:
+for convenience all of them are available to see here:
 
 [[autodoc]] utils.DeepSpeedPlugin
 
@@ -65,6 +89,24 @@ for convience all of them are available to see here:
 
 [[autodoc]] utils.TorchDynamoPlugin
 
+## Configurations
+
+These are classes which can be configured and passed through to the appropriate integration
+
+[[autodoc]] utils.BnbQuantizationConfig
+
+[[autodoc]] utils.DataLoaderConfiguration
+
+[[autodoc]] utils.ProjectConfiguration
+
+## Environmental Variables
+
+These are environmental variables that can be enabled for different use cases
+
+* `ACCELERATE_DEBUG_MODE` (`str`): Whether to run accelerate in debug mode. More info available [here](../usage_guides/debug.md).
+
+
+
 
 ## Data Manipulation and Operations
 
@@ -72,16 +114,34 @@ These include data operations that mimic the same `torch` ops but can be used on
 
 [[autodoc]] utils.broadcast
 
+[[autodoc]] utils.broadcast_object_list
+
 [[autodoc]] utils.concatenate
 
+[[autodoc]] utils.convert_outputs_to_fp32
+
+[[autodoc]] utils.convert_to_fp32
+
 [[autodoc]] utils.gather
 
+[[autodoc]] utils.gather_object
+
+[[autodoc]] utils.get_grad_scaler
+
+[[autodoc]] utils.get_mixed_precision_context_manager
+
+[[autodoc]] utils.listify
+
 [[autodoc]] utils.pad_across_processes
 
+[[autodoc]] utils.recursively_apply
+
 [[autodoc]] utils.reduce
 
 [[autodoc]] utils.send_to_device
 
+[[autodoc]] utils.slice_tensors
+
 ## Environment Checks
 
 These functionalities check the state of the current working environment including information about the operating system itself, what it can support, and if particular dependencies are installed. 
@@ -96,7 +156,7 @@ These functionalities check the state of the current working environment includi
 
 [[autodoc]] utils.is_torch_version
 
-[[autodoc]] utils.is_tpu_available
+[[autodoc]] utils.is_torch_xla_available
 
 [[autodoc]] utils.is_xpu_available
 
@@ -110,9 +170,13 @@ These functionalities check the state of the current working environment includi
 
 When setting up 🤗 Accelerate for the first time, rather than running `accelerate config` [~utils.write_basic_config] can be used as an alternative for quick configuration.
 
-## Memory
+[[autodoc]] utils.set_numa_affinity
 
-[[autodoc]] utils.get_max_memory
+[[autodoc]] utils.environment.override_numa_affinity
+
+[[autodoc]] utils.purge_accelerate_environment
+
+## Memory
 
 [[autodoc]] utils.find_executable_batch_size
 
@@ -120,12 +184,30 @@ When setting up 🤗 Accelerate for the first time, rather than running `acceler
 
 These utilities relate to interacting with PyTorch models
 
+[[autodoc]] utils.calculate_maximum_sizes
+
+[[autodoc]] utils.compute_module_sizes
+
 [[autodoc]] utils.extract_model_from_parallel
 
+[[autodoc]] utils.get_balanced_memory
+
 [[autodoc]] utils.get_max_layer_size
 
+[[autodoc]] utils.infer_auto_device_map
+
+[[autodoc]] utils.load_checkpoint_in_model
+
+[[autodoc]] utils.load_offloaded_weights
+
+[[autodoc]] utils.load_state_dict
+
 [[autodoc]] utils.offload_state_dict
 
+[[autodoc]] utils.retie_parameters
+
+[[autodoc]] utils.set_module_tensor_to_device
+
 
 ## Parallel
 
@@ -135,6 +217,8 @@ These include general utilities that should be used when working in parallel.
 
 [[autodoc]] utils.save
 
+[[autodoc]] utils.load
+
 [[autodoc]] utils.wait_for_everyone
 
 
@@ -166,5 +250,3 @@ These include utilities that are useful to load checkpoints.
 These include utilities that are useful to quantize model.
 
 [[autodoc]] utils.load_and_quantize_model
-
-[[autodoc]] utils.BnbQuantizationConfig

docs/source/quicktour.md

@@ -9,19 +9,80 @@ Unless required by applicable law or agreed to in writing, software distributed
 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
+⚠️ 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.
 -->
 
-# Quick tour
+# Quicktour
 
-Let's have a look at the 🤗 Accelerate main features and traps to avoid.
+There are many ways to launch and run your code depending on your training environment ([torchrun](https://pytorch.org/docs/stable/elastic/run.html), [DeepSpeed](https://www.deepspeed.ai/), etc.) and available hardware. Accelerate offers a unified interface for launching and training on different distributed setups, allowing you to focus on your PyTorch training code instead of the intricacies of adapting your code to these different setups. This allows you to easily scale your PyTorch code for training and inference on distributed setups with hardware like GPUs and TPUs. Accelerate also provides Big Model Inference to make loading and running inference with really large models that usually don't fit in memory more accessible.
 
-## Main use
+This quicktour introduces the three main features of Accelerate:
 
-To use 🤗 Accelerate in your own script, you have to change four things:
+* a unified command line launching interface for distributed training scripts
+* a training library for adapting PyTorch training code to run on different distributed setups
+* Big Model Inference
 
-1. Import the [`Accelerator`] main class and instantiate one in an `accelerator` object:
+## Unified launch interface
+
+Accelerate automatically selects the appropriate configuration values for any given distributed training framework (DeepSpeed, FSDP, etc.) through a unified configuration file generated from the [`accelerate config`](package_reference/cli#accelerate-config) command. You could also pass the configuration values explicitly to the command line which is helpful in certain situations like if you're using SLURM.
+
+
+But in most cases, you should always run [`accelerate config`](package_reference/cli#accelerate-config) first to help Accelerate learn about your training setup.
+
+```bash
+accelerate config
+```
+
+The [`accelerate config`](package_reference/cli#accelerate-config) command creates and saves a default_config.yaml file in Accelerates cache folder. This file stores the configuration for your training environment, which helps Accelerate correctly launch your training script based on your machine.
+
+After you've configured your environment, you can test your setup with [`accelerate test`](package_reference/cli#accelerate-test), which launches a short script to test the distributed environment.
+
+```bash
+accelerate test
+```
+
+> [!TIP]
+> Add `--config_file` to the `accelerate test` or `accelerate launch` command to specify the location of the configuration file if it is saved in a non-default location like the cache.
+
+Once your environment is setup, launch your training script with [`accelerate launch`](package_reference/cli#accelerate-launch)!
+
+```bash
+accelerate launch path_to_script.py --args_for_the_script
+```
+
+To learn more, check out the [Launch distributed code](basic_tutorials/launch) tutorial for more information about launching your scripts.
+
+We also have a [configuration zoo](https://github.com/huggingface/accelerate/blob/main/examples/config_yaml_templates) which showcases a number of premade **minimal** example configurations for a variety of setups you can run.
+
+## Adapt training code
+
+The next main feature of Accelerate is the [`Accelerator`] class which adapts your PyTorch code to run on different distributed setups.
+
+You only need to add a few lines of code to your training script to enable it to run on multiple GPUs or TPUs.
+
+```diff
++ from accelerate import Accelerator
++ accelerator = Accelerator()
+
++ device = accelerator.device
++ model, optimizer, training_dataloader, scheduler = accelerator.prepare(
++     model, optimizer, training_dataloader, scheduler
++ )
+
+  for batch in training_dataloader:
+      optimizer.zero_grad()
+      inputs, targets = batch
+-     inputs = inputs.to(device)
+-     targets = targets.to(device)
+      outputs = model(inputs)
+      loss = loss_function(outputs, targets)
++     accelerator.backward(loss)
+      optimizer.step()
+      scheduler.step()
+```
+
+1. Import and instantiate the [`Accelerator`] class at the beginning of your training script. The [`Accelerator`] class initializes everything necessary for distributed training, and it automatically detects your training environment (a single machine with a GPU, a machine with several GPUs, several machines with multiple GPUs or a TPU, etc.) based on how the code was launched.
 
 ```python
 from accelerate import Accelerator
@@ -29,27 +90,19 @@ from accelerate import Accelerator
 accelerator = Accelerator()
 ```
 
-This should happen as early as possible in your training script as it will initialize everything necessary for
-distributed training. You don't need to indicate the kind of environment you are in (just one machine with a GPU, one
-machines with several GPUs, several machines with multiple GPUs or a TPU), the library will detect this automatically.
+2. Remove calls like `.cuda()` on your model and input data. The [`Accelerator`] class automatically places these objects on the appropriate device for you.
 
-2. Remove the call `.to(device)` or `.cuda()` for your model and input data. The `accelerator` object
-will handle this for you and place all those objects on the right device for you. If you know what you're doing, you
-can leave those `.to(device)` calls but you should use the device provided by the `accelerator` object:
-`accelerator.device`.
+> [!WARNING]
+> This step is *optional* but it is considered best practice to allow Accelerate to handle device placement. You could also deactivate automatic device placement by passing `device_placement=False` when initializing the [`Accelerator`]. If you want to explicitly place objects on a device with `.to(device)`, make sure you use `accelerator.device` instead. For example, if you create an optimizer before placing a model on `accelerator.device`, training fails on a TPU.
 
-To fully deactivate the automatic device placement, pass along `device_placement=False` when initializing your
-[`Accelerator`].
+> [!WARNING]
+> Accelerate does not use non-blocking transfers by default for its automatic device placement, which can result in potentially unwanted CUDA synchronizations.  You can enable non-blocking transfers by passing a [`~utils.dataclasses.DataLoaderConfiguration`] with `non_blocking=True` set as the `dataloader_config` when initializing the [`Accelerator`].  As usual, non-blocking transfers will only work if the dataloader also has `pin_memory=True` set.  Be wary that using non-blocking transfers from GPU to CPU may cause incorrect results if it results in CPU operations being performed on non-ready tensors.
 
-<Tip warning={true}>
+```py
+device = accelerator.device
+```
 
-    If you place your objects manually on the proper device, be careful to create your optimizer after putting your
-    model on `accelerator.device` or your training will fail on TPU.
-
-</Tip>
-
-3. Pass all objects relevant to training (optimizer, model, training dataloader, learning rate scheduler) to the
-[`~Accelerator.prepare`] method. This will make sure everything is ready for training.
+3. Pass all relevant PyTorch objects for training (optimizer, model, dataloader(s), learning rate scheduler) to the [`~Accelerator.prepare`] method as soon as they're created. This method wraps the model in a container optimized for your distributed setup, uses Accelerates version of the optimizer and scheduler, and creates a sharded version of your dataloader for distribution across GPUs or TPUs.
 
 ```python
 model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
@@ -57,73 +110,23 @@ model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
 )
 ```
 
-In particular, your training dataloader will be sharded across all GPUs/TPU cores available so that each one sees a
-different portion of the training dataset. Also, the random states of all processes will be synchronized at the
-beginning of each iteration through your dataloader, to make sure the data is shuffled the same way (if you decided to
-use `shuffle=True` or any kind of random sampler).
+4. Replace `loss.backward()` with [`~Accelerator.backward`] to use the correct `backward()` method for your training setup.
 
-<Tip>
+```py
+accelerator.backward(loss)
+```
 
-    The actual batch size for your training will be the number of devices used multiplied by the batch size you set in
-    your script: for instance training on 4 GPUs with a batch size of 16 set when creating the training dataloader will
-    train at an actual batch size of 64.
+Read [Accelerate’s internal mechanisms](concept_guides/internal_mechanism) guide to learn more details about how Accelerate adapts your code.
 
-</Tip>
+### Distributed evaluation
 
-Alternatively, you can use the option `split_batches=True` when creating and initializing your
-[`Accelerator`], in which case the batch size will always stay the same, whether you run your
-script on 1, 2, 4, or 64 GPUs.
-
-You should execute this instruction as soon as all objects for training are created, before starting your actual
-training loop.
-
-<Tip warning={true}>
-
-    You should only pass the learning rate scheduler to [`~Accelerator.prepare`] when the scheduler needs to be stepped
-    at each optimizer step.
-
-</Tip>
-
-<Tip warning={true}>
-
-    Your training dataloader may change length when going through this method: if you run on X GPUs, it will have its
-    length divided by X (since your actual batch size will be multiplied by X), unless you set
-    `split_batches=True`.
-
-</Tip>
-
-Any instruction using your training dataloader length (for instance if you want to log the number of total training
-steps) should go after the call to [`~Accelerator.prepare`].
-
-You can perfectly send your dataloader to [`~Accelerator.prepare`] on its own, but it's best to send the
-model and optimizer to [`~Accelerator.prepare`] together.
-
-You may or may not want to send your validation dataloader to [`~Accelerator.prepare`], depending on
-whether you want to run distributed evaluation or not (see below).
-
-4. Replace the line `loss.backward()` by `accelerator.backward(loss)`.
-
-And you're all set! With all these changes, your script will run on your local machine as well as on multiple GPUs or a
-TPU! You can either use your favorite tool to launch the distributed training, or you can use the 🤗 Accelerate
-launcher.
-
-
-## Distributed evaluation
-
-You can perform regular evaluation in your training script, if you leave your validation dataloader out of the
-[`~Accelerator.prepare`] method. In this case, you will need to put the input data on the
-`accelerator.device` manually.
-
-To perform distributed evaluation, send along your validation dataloader to the [`~Accelerator.prepare`]
-method:
+To perform distributed evaluation, pass your validation dataloader to the [`~Accelerator.prepare`] method:
 
 ```python
 validation_dataloader = accelerator.prepare(validation_dataloader)
 ```
 
-As for your training dataloader, it will mean that (should you run your script on multiple devices) each device will
-only see part of the evaluation data. This means you will need to group your predictions together. This is very easy to
-do with the [`~Accelerator.gather_for_metrics`] method.
+Each device in your distributed setup only receives a part of the evaluation data, which means you should group your predictions together with the [`~Accelerator.gather_for_metrics`] method. This method requires all tensors to be the same size on each process, so if your tensors have different sizes on each process (for instance when dynamically padding to the maximum length in a batch), you should use the [`~Accelerator.pad_across_processes`] method to pad you tensor to the largest size across processes. Note that the tensors needs to be 1D and that we concatenate the tensors along the first dimension. 
 
 ```python
 for inputs, targets in validation_dataloader:
@@ -134,387 +137,53 @@ for inputs, targets in validation_dataloader:
     metric.add_batch(all_predictions, all_targets)
 ```
 
-<Tip warning={true}>
+For more complex cases (e.g. 2D tensors, don't want to concatenate tensors, dict of 3D tensors), you can pass `use_gather_object=True` in `gather_for_metrics`. This will return the list of objects after gathering. Note that using it with GPU tensors is not well supported and inefficient.
 
-    Similar to the training dataloader, passing your validation dataloader through
-    [`~Accelerator.prepare`] may change it: if you run on X GPUs, it will have its length divided by X
-    (since your actual batch size will be multiplied by X), unless you set `split_batches=True`.
+> [!TIP]
+> Data at the end of a dataset may be duplicated so the batch can be equally divided among all workers. The [`~Accelerator.gather_for_metrics`] method automatically removes the duplicated data to calculate a more accurate metric.
 
-</Tip>
+## Big Model Inference
 
-Any instruction using your training dataloader length (for instance if you need the number of total training steps
-to create a learning rate scheduler) should go after the call to [`~Accelerator.prepare`]. 
+Accelerate's Big Model Inference has two main features, [`~accelerate.init_empty_weights`] and [`~accelerate.load_checkpoint_and_dispatch`], to load large models for inference that typically don't fit into memory.
 
-Some data at the end of the dataset may be duplicated so the batch can be divided equally among all workers. As a result, metrics
-should be calculated through the [`~Accelerator.gather_for_metrics`] method to automatically remove the duplicated data while gathering.
+> [!TIP]
+> Take a look at the [Handling big models for inference](concept_guides/big_model_inference) guide for a better understanding of how Big Model Inference works under the hood.
 
-<Tip>
+### Empty weights initialization
 
-    If for some reason you don't wish to have this automatically done, [`~Accelerator.gather`] can be used instead to gather 
-    the data across all processes and this can manually be done instead.
+The [`~accelerate.init_empty_weights`] context manager initializes models of any size by creating a *model skeleton* and moving and placing parameters each time they're created to PyTorch's [**meta**](https://pytorch.org/docs/main/meta.html) device. This way, not all weights are immediately loaded and only a small part of the model is loaded into memory at a time.
 
-</Tip>
+For example, loading an empty [Mixtral-8x7B](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1) model takes significantly less memory than fully loading the models and weights on the CPU.
 
+```py
+from accelerate import init_empty_weights
+from transformers import AutoConfig, AutoModelForCausalLM
 
-<Tip warning={true}>
-
-    The [`~Accelerator.gather`] and [`~Accelerator.gather_for_metrics`] methods require the tensors to be all the same size on each process. If
-    you have tensors of different sizes on each process (for instance when dynamically padding to the maximum length in
-    a batch), you should use the [`~Accelerator.pad_across_processes`] method to pad you tensor to the
-    biggest size across processes.
-
-</Tip>
-
-## Launching your distributed script
-
-You can use the regular commands to launch your distributed training (like `torch.distributed.run` for
-PyTorch), they are fully compatible with 🤗 Accelerate.
-
-🤗 Accelerate also provides a CLI tool that unifies all launchers, so you only have to remember one command. To use it,
-just run:
-
-```bash
-accelerate config
+config = AutoConfig.from_pretrained("mistralai/Mixtral-8x7B-Instruct-v0.1")
+with init_empty_weights():
+    model = AutoModelForCausalLM.from_config(config)
 ```
 
-on your machine and reply to the questions asked. This will save a *default_config.yaml* file in your cache folder for
-🤗 Accelerate. That cache folder is (with decreasing order of priority):
+### Load and dispatch weights
 
-- The content of your environment variable `HF_HOME` suffixed with *accelerate*.
-- If it does not exist, the content of your environment variable `XDG_CACHE_HOME` suffixed with
-  *huggingface/accelerate*.
-- If this does not exist either, the folder *~/.cache/huggingface/accelerate*
+The [`~accelerate.load_checkpoint_and_dispatch`] function loads full or sharded checkpoints into the empty model, and automatically distribute weights across all available devices.
 
-You can also specify with the flag `--config_file` the location of the file you want to save.
+The `device_map` parameter determines where to place each model layer, and specifiying `"auto"` places them on the GPU first, then the CPU, and finally the hard drive as memory-mapped tensors if there's still not enough memory. Use the `no_split_module_classes` parameter to indicate which modules shouldn't be split across devices (typically those with a residual connection).
 
-Once this is done, you can test everything is going well on your setup by running:
+```py
+from accelerate import load_checkpoint_and_dispatch
 
-```bash
-accelerate test
+model_checkpoint = "your-local-model-folder"
+model = load_checkpoint_and_dispatch(
+    model, checkpoint=model_checkpoint, device_map="auto", no_split_module_classes=['Block']
+)
 ```
 
-This will launch a short script that will test the distributed environment. If it runs fine, you are ready for the next
-step!
+## Next steps
 
-Note that if you specified a location for the config file in the previous step, you need to pass it here as well:
+Now that you've been introduced to the main Accelerate features, your next steps could include:
 
-```bash
-accelerate test --config_file path_to_config.yaml
-```
-
-Now that this is done, you can run your script with the following command:
-
-```bash
-accelerate launch path_to_script.py --args_for_the_script
-```
-
-If you stored the config file in a non-default location, you can indicate it to the launcher like this:
-
-```bash
-accelerate launch --config_file path_to_config.yaml path_to_script.py --args_for_the_script
-```
-
-You can also override any of the arguments determined by your config file. 
-To see the complete list of parameters that you can pass in, run `accelerate launch -h`. 
-
-Check out the [Launch tutorial](basic_tutorials/launch) for more information about launching your scripts. 
-
-
-## Launching training from a notebook
-
-In Accelerate 0.3.0, a new [`notebook_launcher`] has been introduced to help you launch your training
-function from a notebook. This launcher supports launching a training with TPUs on Colab or Kaggle, as well as training
-on several GPUs (if the machine on which you are running your notebook has them).
-
-Just define a function responsible for your whole training and/or evaluation in a cell of the notebook, then execute a
-cell with the following code:
-
-```python
-from accelerate import notebook_launcher
-
-notebook_launcher(training_function)
-```
-
-<Tip warning={true}>
-
-    Your [`Accelerator`] object should only be defined inside the training function. This is because the
-    initialization should be done inside the launcher only.
-
-</Tip>
-
-Check out the [Notebook Launcher tutorial](basic_tutorials/notebook) for more information about training on TPUs. 
-
-
-## Training on TPU
-
-If you want to launch your script on TPUs, there are a few caveats you should be aware of. Behind the scenes, the TPUs
-will create a graph of all the operations happening in your training step (forward pass, backward pass and optimizer
-step). This is why your first step of training will always be very long as building and compiling this graph for
-optimizations takes some time.
-
-The good news is that this compilation will be cached so the second step and all the following will be much faster. The
-bad news is that it only applies if all of your steps do exactly the same operations, which implies:
-
-- having all tensors of the same length in all your batches
-- having static code (i.e., not a for loop of length that could change from step to step)
-
-Having any of the things above change between two steps will trigger a new compilation which will, once again, take a
-lot of time. In practice, that means you must take special care to have all your tensors in your inputs of the same
-shape (so no dynamic padding for instance if you are in an NLP problem) and should not use layers with for loops that
-have different lengths depending on the inputs (such as an LSTM) or the training will be excruciatingly slow.
-
-To introduce special behavior in your script for TPUs you can check the `distributed_type` of your
-`accelerator`:
-
-```python docstyle-ignore
-from accelerate import DistributedType
-
-if accelerator.distributed_type == DistributedType.TPU:
-    # do something of static shape
-else:
-    # go crazy and be dynamic
-```
-
-The [NLP example](https://github.com/huggingface/accelerate/blob/main/examples/nlp_example.py) shows an example in a 
-situation with dynamic padding.
-
-One last thing to pay close attention to: if your model has tied weights (such as language models which tie the weights
-of the embedding matrix with the weights of the decoder), moving this model to the TPU (either yourself or after you
-passed your model to [`~Accelerator.prepare`]) will break the tying. You will need to retie the weights
-after. You can find an example of this in the [run_clm_no_trainer](https://github.com/huggingface/transformers/blob/master/examples/pytorch/language-modeling/run_clm.py) script in
-the Transformers repository.
-
-Check out the [TPU tutorial](concept_guides/training_tpu) for more information about training on TPUs. 
-
-
-## Other caveats
-
-We list here all smaller issues you could have in your script conversion and how to resolve them.
-
-### Execute a statement only on one processes
-
-Some of your instructions only need to run for one process on a given server: for instance a data download or a log
-statement. To do this, wrap the statement in a test like this:
-
-```python docstyle-ignore
-if accelerator.is_local_main_process:
-    # Is executed once per server
-```
-
-Another example is progress bars: to avoid having multiple progress bars in your output, you should only display one on
-the local main process:
-
-```python
-from tqdm.auto import tqdm
-
-progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
-```
-
-The *local* means per machine: if you are running your training on two servers with several GPUs, the instruction will
-be executed once on each of those servers. If you need to execute something only once for all processes (and not per
-machine) for instance, uploading the final model to the 🤗 model hub, wrap it in a test like this:
-
-```python docstyle-ignore
-if accelerator.is_main_process:
-    # Is executed once only
-```
-
-For printing statements you only want executed once per machine, you can just replace the `print` function by
-`accelerator.print`.
-
-
-### Defer execution
-
-When you run your usual script, instructions are executed in order. Using 🤗 Accelerate to deploy your script on several
-GPUs at the same time introduces a complication: while each process executes all instructions in order, some may be
-faster than others.
-
-You might need to wait for all processes to have reached a certain point before executing a given instruction. For
-instance, you shouldn't save a model before being sure every process is done with training. To do this, just write the
-following line in your code:
-
-```
-accelerator.wait_for_everyone()
-```
-
-This instruction will block all the processes that arrive first until all the other processes have reached that
-point (if you run your script on just one GPU or CPU, this won't do anything).
-
-
-### Saving/loading a model
-
-Saving the model you trained might need a bit of adjustment: first you should wait for all processes to reach that
-point in the script as shown above, and then, you should unwrap your model before saving it. This is because when going
-through the [`~Accelerator.prepare`] method, your model may have been placed inside a bigger model,
-which deals with the distributed training. This in turn means that saving your model state dictionary without taking
-any precaution will take that potential extra layer into account, and you will end up with weights you can't load back
-in your base model. The [`~Accelerator.save_model`] method will help you to achieve that. It will unwrap your model and save
-the model state dictionnary.
-
-Here is an example:
-```
-accelerator.wait_for_everyone()
-accelerator.save_model(model, save_directory)
-```
-The [`~Accelerator.save_model`] method can also save a model into sharded checkpoints or with safetensors format.
-Here is an example: 
-
-```python
-accelerator.wait_for_everyone()
-accelerator.save_model(model, save_directory, max_shard_size="1GB", safe_serialization=True)
-```
-
-If your script contains logic to load a checkpoint, we also recommend you load your weights in the unwrapped model
-(this is only useful if you use the load function after making your model go through
-[`~Accelerator.prepare`]). Here is an example:
-
-```python
-unwrapped_model = accelerator.unwrap_model(model)
-path_to_checkpoint = os.path.join(save_directory,"pytorch_model.bin")
-unwrapped_model.load_state_dict(torch.load(path_to_checkpoint))
-```
-
-Note that since all the model parameters are references to tensors, this will load your weights inside `model`.
-
-If you want to load a sharded checkpoint or a checkpoint with safetensors format into the model with a specific `device`, we recommend you to load it with [`~utils.load_checkpoint_in_model`] function. Here's an example:
-
-```python
-load_checkpoint_in_model(unwrapped_model, save_directory, device_map={"":device})
-```
-
-## Saving/loading entire states
-
-When training your model, you may want to save the current state of the model, optimizer, random generators, and potentially LR schedulers to be restored in the _same script_.
-You can use [`~Accelerator.save_state`] and [`~Accelerator.load_state`] respectively to do so.
-
-To further customize where and how states saved through [`~Accelerator.save_state`] the [`~utils.ProjectConfiguration`] class can be used. For example 
-if `automatic_checkpoint_naming` is enabled each saved checkpoint will be located then at `Accelerator.project_dir/checkpoints/checkpoint_{checkpoint_number}`.
-
-If you have registered any other stateful items to be stored through [`~Accelerator.register_for_checkpointing`] they will also be saved and/or loaded.
-
-<Tip>
-
-    Every object passed to [`~Accelerator.register_for_checkpointing`] must have a `load_state_dict` and `state_dict` function to be stored
-
-</Tip>
-
-
-### Gradient clipping
-
-If you are using gradient clipping in your script, you should replace the calls to
-`torch.nn.utils.clip_grad_norm_` or `torch.nn.utils.clip_grad_value_` with [`~Accelerator.clip_grad_norm_`]
-and [`~Accelerator.clip_grad_value_`] respectively.
-
-
-### Mixed Precision training
-
-If you are running your training in Mixed Precision with 🤗 Accelerate, you will get the best result with your loss being
-computed inside your model (like in Transformer models for instance). Every computation outside of the model will be
-executed in full precision (which is generally what you want for loss computation, especially if it involves a
-softmax). However you might want to put your loss computation inside the [`~Accelerator.autocast`] context manager:
-
-```
-with accelerator.autocast():
-    loss = complex_loss_function(outputs, target):
-```
-
-Another caveat with Mixed Precision training is that the gradient will skip a few updates at the beginning and
-sometimes during training: because of the dynamic loss scaling strategy, there are points during training where the
-gradients have overflown, and the loss scaling factor is reduced to avoid this happening again at the next step.
-
-This means that you may update your learning rate scheduler when there was no update, which is fine in general, but may
-have an impact when you have very little training data, or if the first learning rate values of your scheduler are very
-important. In this case, you can skip the learning rate scheduler updates when the optimizer step was not done like
-this:
-
-```
-if not accelerator.optimizer_step_was_skipped:
-    lr_scheduler.step()
-```
-
-### Gradient Accumulation 
-
-To perform gradient accumulation use [`~Accelerator.accumulate`] and specify a `gradient_accumulation_steps`. 
-This will also automatically ensure the gradients are synced or unsynced when on multi-device training, check if the step should
-actually be performed, and auto-scale the loss:
-
-```python
-accelerator = Accelerator(gradient_accumulation_steps=2)
-model, optimizer, training_dataloader = accelerator.prepare(model, optimizer, training_dataloader)
-
-for input, label in training_dataloader:
-    with accelerator.accumulate(model):
-        predictions = model(input)
-        loss = loss_function(predictions, label)
-        accelerator.backward(loss)
-        optimizer.step()
-        scheduler.step()
-        optimizer.zero_grad()
-```
-
-### DeepSpeed
-
-DeepSpeed support is experimental, so the underlying API will evolve in the near future and may have some slight
-breaking changes. In particular, 🤗 Accelerate does not support DeepSpeed config you have written yourself yet, this
-will be added in a next version.
-
-<Tip warning={true}>
-
-    The [`notebook_launcher`] does not support the DeepSpeed integration yet.
-
-</Tip>
-
-## Internal mechanism
-
-Internally, the library works by first analyzing the environment in which the script is launched to determine which
-kind of distributed setup is used, how many different processes there are and which one the current script is in. All
-that information is stored in the [`~AcceleratorState`].
-
-This class is initialized the first time you instantiate an [`~Accelerator`] as well as performing any
-specific initialization your distributed setup needs. Its state is then uniquely shared through all instances of
-[`~state.AcceleratorState`].
-
-Then, when calling [`~Accelerator.prepare`], the library:
-
-- wraps your model(s) in the container adapted for the distributed setup,
-- wraps your optimizer(s) in a [`~optimizer.AcceleratedOptimizer`],
-- creates a new version of your dataloader(s) in a [`~data_loader.DataLoaderShard`].
-
-While the model(s) and optimizer(s) are just put in simple wrappers, the dataloader(s) are re-created. This is mostly
-because PyTorch does not let the user change the `batch_sampler` of a dataloader once it's been created and the
-library handles the sharding of your data between processes by changing that `batch_sampler` to yield every other
-`num_processes` batches.
-
-The [`~data_loader.DataLoaderShard`] subclasses `DataLoader` to add the following functionality:
-
-- it synchronizes the appropriate random number generator of all processes at each new iteration, to ensure any
-  randomization (like shuffling) is done the exact same way across processes.
-- it puts the batches on the proper device before yielding them (unless you have opted out of
-  `device_placement=True`).
-
-The random number generator synchronization will by default synchronize:
-
-- the `generator` attribute of a given sampler (like the PyTorch `RandomSampler`) for PyTorch >= 1.6
-- the main random number generator in PyTorch <=1.5.1
-
-You can choose which random number generator(s) to synchronize with the `rng_types` argument of the main
-[`Accelerator`]. In PyTorch >= 1.6, it is recommended to rely on a local `generator` to avoid
-setting the same seed in the main random number generator in all processes.
-
-<Tip warning={true}>
-
-    Synchronization of the main torch (or CUDA or XLA) random number generator will affect any other potential random
-    artifacts you could have in your dataset (like random data augmentation) in the sense that all processes will get
-    the same random numbers from the torch random modules (so will apply the same random data augmentation if it's
-    controlled by torch).
-
-</Tip>
-
-<Tip>
-
-    The randomization part of your custom sampler, batch sampler or iterable dataset should be done using a local
-    `torch.Generator` object (in PyTorch >= 1.6), see the traditional `RandomSampler`, as an example.
-
-</Tip>
-
-For more details about the internals, see the [Internals page](package_reference/torch_wrappers).
+* Check out the [tutorials](basic_tutorials/overview) for a gentle walkthrough of Accelerate. This is especially useful if you're new to distributed training and the library.
+* Dive into the [guides](usage_guides/explore) to see how to use Accelerate for specific use-cases.
+* Deepen your conceptual understanding of how Accelerate works internally by reading the [concept guides](concept_guides/internal_mechanism).
+* Look up classes and commands in the [API reference](package_reference/accelerator) to see what parameters and options are available.

docs/source/usage_guides/big_modeling.md

@@ -13,15 +13,15 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Handling big models for inference
+# Big Model Inference
 
-One of the biggest advancements 🤗 Accelerate provides is the concept of [large model inference](../concept_guides/big_model_inference) wherein you can perform *inference* on models that cannot fully fit on your graphics card. 
+One of the biggest advancements Accelerate provides is [Big Model Inference](../concept_guides/big_model_inference), which allows you to perform inference with models that don't fully fit on your graphics card.
 
-This tutorial will be broken down into two parts showcasing how to use both 🤗 Accelerate and 🤗 Transformers (a higher API-level) to make use of this idea.
+This tutorial will show you how to use Big Model Inference in Accelerate and the Hugging Face ecosystem.
 
-## Using 🤗 Accelerate
+## Accelerate
 
-For these tutorials, we'll assume a typical workflow for loading your model in such that:
+A typical workflow for loading a PyTorch model is shown below. `ModelClass` is a model that exceeds the GPU memory of your device (mps or cuda or xpu).
 
 ```py
 import torch
@@ -31,9 +31,7 @@ state_dict = torch.load(checkpoint_file)
 my_model.load_state_dict(state_dict)
 ```
 
-Note that here we assume that `ModelClass` is a model that takes up more video-card memory than what can fit on your device (be it `mps` or `cuda`).
-
-The first step is to init an empty skeleton of the model which won't take up any RAM using the [`init_empty_weights`] context manager:
+With Big Model Inference, the first step is to init an empty skeleton of the model with the `init_empty_weights` context manager. This doesn't require any memory because `my_model` is "parameterless".
 
 ```py
 from accelerate import init_empty_weights
@@ -41,22 +39,14 @@ with init_empty_weights():
     my_model = ModelClass(...)
 ```
 
-With this `my_model` currently is "parameterless", hence leaving the smaller footprint than what one would normally get loading this onto the CPU directly. 
+Next, the weights are loaded into the model for inference.
 
-Next we need to load in the weights to our model so we can perform inference.
+The [`load_checkpoint_and_dispatch`] method loads a checkpoint inside your empty model and dispatches the weights for each layer across all available devices, starting with the fastest devices (GPU, MPS, XPU, NPU, MLU, MUSA) first before moving to the slower ones (CPU and hard drive).
 
-For this we will use [`load_checkpoint_and_dispatch`], which as the name implies will load a checkpoint inside your empty model and dispatch the weights for each layer across all the devices you have available (GPU/MPS and CPU RAM). 
+Setting `device_map="auto"` automatically fills all available space on the GPU(s) first, then the CPU, and finally, the hard drive (the absolute slowest option) if there is still not enough memory.
 
-To determine how this `dispatch` can be performed, generally specifying `device_map="auto"` will be good enough as 🤗 Accelerate
-will attempt to fill all the space in your GPU(s), then loading them to the CPU, and finally if there is not enough RAM it will be loaded to the disk (the absolute slowest option). 
-
-<Tip>
-
-For more details on desigining your own device map, see this section of the [concept guide](../concept_guide/big_model_inference#desigining-a-device-map)
-
-</Tip>
-
-See an example below:
+> [!TIP]
+> Refer to the [Designing a device map](../concept_guides/big_model_inference#designing-a-device-map) guide for more details on how to design your own device map.
 
 ```py
 from accelerate import load_checkpoint_and_dispatch
@@ -66,42 +56,29 @@ model = load_checkpoint_and_dispatch(
 )
 ```
 
-<Tip>
+If there are certain “chunks” of layers that shouldn’t be split, pass them to `no_split_module_classes` (see [here](../concept_guides/big_model_inference#loading-weights) for more details).
 
-    If there are certain "chunks" of layers that shouldn't be split, you can pass them in as `no_split_module_classes`. Read more about it [here](../concept_guides/big_model_inference#loading-weights)
+A models weights can also be sharded into multiple checkpoints to save memory, such as when the `state_dict` doesn't fit in memory (see [here](../concept_guides/big_model_inference#sharded-checkpoints) for more details).
 
-</Tip>
-
-<Tip>
-
-    Also to save on memory (such as if the `state_dict` will not fit in RAM), a model's weights can be divided and split into multiple checkpoint files. Read more about it [here](../concept_guides/big_model_inference#sharded-checkpoints)
-
-</Tip>
-
-Now that the model is dispatched fully, you can perform inference as normal with the model:
+Now that the model is fully dispatched, you can perform inference.
 
 ```py
 input = torch.randn(2,3)
-input = input.to("cuda")
+device_type = next(iter(model.parameters())).device.type
+input = input.to(device_type)
 output = model(input)
 ```
 
-What will happen now is each time the input gets passed through a layer, it will be sent from the CPU to the GPU (or disk to CPU to GPU), the output is calculated, and then the layer is pulled back off the GPU going back down the line. While this adds some overhead to the inference being performed, through this method it is possible to run **any size model** on your system, as long as the largest layer is capable of fitting on your GPU. 
+Each time an input is passed through a layer, it is sent from the CPU to the GPU (or disk to CPU to GPU), the output is calculated, and the layer is removed from the GPU going back down the line. While this adds some overhead to inference, it enables you to run any size model on your system, as long as the largest layer fits on your GPU.
 
-<Tip>
+Multiple GPUs, or "model parallelism", can be utilized but only one GPU will be active at any given moment. This forces the GPU to wait for the previous GPU to send it the output. You should launch your script normally with Python instead of other tools like torchrun and accelerate launch.
 
-    Multiple GPUs can be utilized, however this is considered "model parallism" and as a result only one GPU will be active at a given moment, waiting for the prior one to send it the output. You should launch your script normally with `python`
-    and not need `torchrun`, `accelerate launch`, etc.
+> [!TIP]
+> You may also be interested in *pipeline parallelism* which utilizes all available GPUs at once, instead of only having one GPU active at a time. This approach is less flexbile though. For more details, refer to the [Memory-efficient pipeline parallelism](./distributed_inference#memory-efficient-pipeline-parallelism-experimental) guide.
 
-</Tip>
+<Youtube id="MWCSGj9jEAo"/>
 
-For a visual representation of this, check out the animation below:
-
-<Youtube id="MWCSGj9jEAo" />
-
-### Complete Example
-
-Below is the full example showcasing what we performed above:
+Take a look at a full example of Big Model Inference below.
 
 ```py
 import torch
@@ -115,36 +92,33 @@ model = load_checkpoint_and_dispatch(
 )
 
 input = torch.randn(2,3)
-input = input.to("cuda")
+device_type = next(iter(model.parameters())).device.type
+input = input.to(device_type)
 output = model(input)
 ```
 
-## Using 🤗 Transformers, 🤗 Diffusers, and other 🤗 Open Source Libraries
+## Hugging Face ecosystem
 
-Libraries that support 🤗 Accelerate big model inference include all of the earlier logic in their `from_pretrained` constructors. 
+Other libraries in the Hugging Face ecosystem, like Transformers or Diffusers, supports Big Model Inference in their [`~transformers.PreTrainedModel.from_pretrained`] constructors.
 
-These operate by specifying a string representing the model to download from the [🤗 Hub](https://hf.co/models) and then denoting `device_map="auto"` along with a few extra parameters. 
+You just need to add `device_map="auto"` in [`~transformers.PreTrainedModel.from_pretrained`] to enable Big Model Inference.
 
-As a brief example, we will look at using `transformers` and loading in Big Science's T0pp model. 
+For example, load Big Sciences T0pp 11 billion parameter model with Big Model Inference.
 
 ```py
 from transformers import AutoModelForSeq2SeqLM
 
-model = AutoModelForSeq2SeqLM("bigscience/T0pp", device_map="auto")
+model = AutoModelForSeq2SeqLM.from_pretrained("bigscience/T0pp", device_map="auto")
 ```
 
-After loading the model in, the initial steps from before to prepare a model have all been done and the model is fully
-ready to make use of all the resources in your machine. Through these constructors, you can also save *more* memory by
-specifying the precision the model is loaded into as well, through the `torch_dtype` parameter, such as:
+After loading the model, the empty init and smart dispatch steps from before are executed and the model is fully ready to make use of all the resources in your machine. Through these constructors, you can also save more memory by specifying the `torch_dtype` parameter to load a model in a lower precision.
 
 ```py
 from transformers import AutoModelForSeq2SeqLM
 
-model = AutoModelForSeq2SeqLM("bigscience/T0pp", device_map="auto", torch_dtype=torch.float16)
+model = AutoModelForSeq2SeqLM.from_pretrained("bigscience/T0pp", device_map="auto", torch_dtype=torch.float16)
 ```
 
-To learn more about this, check out the 🤗 Transformers documentation available [here](https://huggingface.co/docs/transformers/main/en/main_classes/model#large-model-loading).
+## Next steps
 
-## Where to go from here
-
-For a much more detailed look at big model inference, be sure to check out the [Conceptual Guide on it](../concept_guides/big_model_inference)
+For a more detailed explanation of Big Model Inference, make sure to check out the [conceptual guide](../concept_guides/big_model_inference)!

docs/source/usage_guides/checkpoint.md

@@ -15,8 +15,8 @@ rendered properly in your Markdown viewer.
 
 # Checkpointing
 
-When training a PyTorch model with 🤗 Accelerate, you may often want to save and continue a state of training. Doing so requires
-saving and loading the model, optimizer, RNG generators, and the GradScaler. Inside 🤗 Accelerate are two convenience functions to achieve this quickly:
+When training a PyTorch model with Accelerate, you may often want to save and continue a state of training. Doing so requires
+saving and loading the model, optimizer, RNG generators, and the GradScaler. Inside Accelerate are two convenience functions to achieve this quickly:
 - Use [`~Accelerator.save_state`] for saving everything mentioned above to a folder location
 - Use [`~Accelerator.load_state`] for loading everything stored from an earlier `save_state`
 

docs/source/usage_guides/ddp_comm_hook.md

@@ -0,0 +1,325 @@
+<!--
+Copyright 2022 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.
+-->
+
+# DDP Communication Hooks
+
+Distributed Data Parallel (DDP) communication hooks provide a generic interface to control how gradients are communicated across workers by overriding the vanilla allreduce in `DistributedDataParallel`. A few built-in communication hooks are provided, and users can easily apply any of these hooks to optimize communication.
+
+
+- **FP16 Compression Hook**: Compresses gradients by casting them to half-precision floating-point format (`torch.float16`), reducing communication overhead.
+- **BF16 Compression Hook**: Similar to FP16, but uses the Brain Floating Point format (`torch.bfloat16`), which can be more efficient on certain hardware.
+- **PowerSGD Hook**: An advanced gradient compression algorithm that provides high compression rates and can accelerate bandwidth-bound distributed training.
+
+In this tutorial, you will see how to quickly set up DDP communication hooks and perform training with the utilities provided in Accelerate, which can be as simple as adding just one new line of code! This demonstrates how to use DDP communication hooks to optimize gradient communication in distributed training with the Accelerate library.
+
+## FP16 Compression Hook
+
+<hfoptions id="fp16">
+<hfoption id="PyTorch">
+
+```python
+import torch
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.distributed.algorithms.ddp_comm_hooks import default_hooks
+
+class MyModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.layer = torch.nn.Linear(10, 10)
+
+    def forward(self, x):
+        return self.layer(x)
+
+model = MyModel()
+model = DDP(model, device_ids=[torch.cuda.current_device()])
+model.register_comm_hook(state=None, hook=default_hooks.fp16_compress_hook)
+
+# Training loop
+for data, targets in data_loader:
+    outputs = model(data)
+    loss = criterion(outputs, targets)
+    loss.backward()
+    optimizer.step()
+    optimizer.zero_grad()
+```
+
+</hfoption>
+<hfoption id="Accelerate">
+
+```python
+from accelerate import Accelerator, DDPCommunicationHookType, DistributedDataParallelKwargs
+import torch
+
+class MyModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.layer = torch.nn.Linear(10, 10)
+
+    def forward(self, x):
+        return self.layer(x)
+
+# DDP Communication Hook setup
+ddp_kwargs = DistributedDataParallelKwargs(comm_hook=DDPCommunicationHookType.FP16)
+accelerator = Accelerator(kwargs_handlers=[ddp_kwargs])
+
+model = MyModel()
+optimizer = torch.optim.Adam(model.parameters())
+data_loader = DataLoader(dataset, batch_size=16)
+
+model, optimizer, data_loader = accelerator.prepare(model, optimizer, data_loader)
+
+# Training loop
+for data, targets in data_loader:
+    outputs = model(data)
+    loss = criterion(outputs, targets)
+    accelerator.backward(loss)
+    optimizer.step()
+    optimizer.zero_grad()
+```
+
+</hfoption>
+</hfoptions>
+
+### BF16 Compression Hook
+
+<Tip warning={true}>
+
+BF16 Compression Hook API is experimental, and it requires NCCL version later than 2.9.6.
+
+</Tip>
+
+<hfoptions id="bf16">
+<hfoption id="PyTorch">
+
+```python
+import torch
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.distributed.algorithms.ddp_comm_hooks import default_hooks
+
+class MyModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.layer = torch.nn.Linear(10, 10)
+
+    def forward(self, x):
+        return self.layer(x)
+
+model = MyModel()
+model = DDP(model, device_ids=[torch.cuda.current_device()])
+model.register_comm_hook(state=None, hook=default_hooks.bf16_compress_hook)
+
+# Training loop
+for data, targets in data_loader:
+    outputs = model(data)
+    loss = criterion(outputs, targets)
+    loss.backward()
+    optimizer.step()
+    optimizer.zero_grad()
+```
+
+</hfoption>
+<hfoption id="Accelerate">
+
+```python
+from accelerate import Accelerator, DDPCommunicationHookType, DistributedDataParallelKwargs
+import torch
+
+class MyModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.layer = torch.nn.Linear(10, 10)
+
+    def forward(self, x):
+        return self.layer(x)
+
+# DDP Communication Hook setup
+ddp_kwargs = DistributedDataParallelKwargs(comm_hook=DDPCommunicationHookType.BF16)
+accelerator = Accelerator(kwargs_handlers=[ddp_kwargs])
+
+model = MyModel()
+optimizer = torch.optim.Adam(model.parameters())
+data_loader = DataLoader(dataset, batch_size=16)
+
+model, optimizer, data_loader = accelerator.prepare(model, optimizer, data_loader)
+
+# Training loop
+for data, targets in data_loader:
+    outputs = model(data)
+    loss = criterion(outputs, targets)
+    accelerator.backward(loss)
+    optimizer.step()
+    optimizer.zero_grad()
+```
+
+</hfoption>
+</hfoptions>
+
+### PowerSGD Hook
+
+<Tip warning={true}>
+
+PowerSGD typically requires extra memory of the same size as the model’s gradients to enable error feedback, which can compensate for biased compressed communication and improve accuracy.
+
+</Tip>
+
+<hfoptions id="powerSGD">
+<hfoption id="PyTorch">
+
+```python
+import torch
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.distributed.algorithms.ddp_comm_hooks import powerSGD_hook
+
+class MyModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.layer = torch.nn.Linear(10, 10)
+
+    def forward(self, x):
+        return self.layer(x)
+
+model = MyModel()
+model = DDP(model, device_ids=[torch.cuda.current_device()])
+state = powerSGD_hook.PowerSGDState(process_group=None)
+model.register_comm_hook(state=state, hook=powerSGD_hook.powerSGD_hook)
+
+# Training loop
+for data, targets in data_loader:
+    outputs = model(data)
+    loss = criterion(outputs, targets)
+    loss.backward()
+    optimizer.step()
+    optimizer.zero_grad()
+```
+
+</hfoption>
+<hfoption id="Accelerate">
+
+```python
+from accelerate import Accelerator, DDPCommunicationHookType, DistributedDataParallelKwargs
+import torch
+
+class MyModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.layer = torch.nn.Linear(10, 10)
+
+    def forward(self, x):
+        return self.layer(x)
+
+# DDP Communication Hook setup
+ddp_kwargs = DistributedDataParallelKwargs(comm_hook=DDPCommunicationHookType.POWER_SGD)
+accelerator = Accelerator(kwargs_handlers=[ddp_kwargs])
+
+model = MyModel()
+optimizer = torch.optim.Adam(model.parameters())
+data_loader = DataLoader(dataset, batch_size=16)
+
+model, optimizer, data_loader = accelerator.prepare(model, optimizer, data_loader)
+
+# Training loop
+for data, targets in data_loader:
+    outputs = model(data)
+    loss = criterion(outputs, targets)
+    accelerator.backward(loss)
+    optimizer.step()
+    optimizer.zero_grad()
+```
+
+</hfoption>
+</hfoptions>
+
+## DDP Communication Hooks utilities
+
+There are two additional utilities for supporting optional functionalities with the communication hooks.
+
+### comm_wrapper
+
+`comm_wrapper` is an option to wrap a communication hook with additional functionality. For example, it can be used to combine FP16 compression with other communication strategies. Currently supported wrappers are `no`, `fp16`, and `bf16`.
+
+```python
+from accelerate import Accelerator, DDPCommunicationHookType, DistributedDataParallelKwargs
+import torch
+
+class MyModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.layer = torch.nn.Linear(10, 10)
+
+    def forward(self, x):
+        return self.layer(x)
+
+# DDP Communication Hook setup
+ddp_kwargs = DistributedDataParallelKwargs(
+    comm_hook=DDPCommunicationHookType.POWER_SGD,
+    comm_wrapper=DDPCommunicationHookType.FP16
+)
+accelerator = Accelerator(kwargs_handlers=[ddp_kwargs])
+
+model = MyModel()
+optimizer = torch.optim.Adam(model.parameters())
+data_loader = DataLoader(dataset, batch_size=16)
+
+model, optimizer, data_loader = accelerator.prepare(model, optimizer, data_loader)
+
+# Training loop
+for data, targets in data_loader:
+    outputs = model(data)
+    loss = criterion(outputs, targets)
+    accelerator.backward(loss)
+    optimizer.step()
+    optimizer.zero_grad()
+```
+
+### comm_state_option
+
+`comm_state_option` allows you to pass additional state information required by certain communication hooks. This is particularly useful for stateful hooks like `PowerSGD`, which require maintaining hyperparameters and internal states across training steps. Below is an example showcasing the use of `comm_state_option` with the `PowerSGD` hook.
+
+```python
+from accelerate import Accelerator, DDPCommunicationHookType, DistributedDataParallelKwargs
+import torch
+
+class MyModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.layer = torch.nn.Linear(10, 10)
+
+    def forward(self, x):
+        return self.layer(x)
+
+# DDP Communication Hook setup
+ddp_kwargs = DistributedDataParallelKwargs(
+    comm_hook=DDPCommunicationHookType.POWER_SGD,
+    comm_state_option={"matrix_approximation_rank": 2}
+)
+accelerator = Accelerator(kwargs_handlers=[ddp_kwargs])
+
+model = MyModel()
+optimizer = torch.optim.Adam(model.parameters())
+data_loader = DataLoader(dataset, batch_size=16)
+
+model, optimizer, data_loader = accelerator.prepare(model, optimizer, data_loader)
+
+# Training loop
+for data, targets in data_loader:
+    outputs = model(data)
+    loss = criterion(outputs, targets)
+    accelerator.backward(loss)
+    optimizer.step()
+    optimizer.zero_grad()
+```
+
+For more advanced usage and additional hooks, refer to the [PyTorch DDP Communication Hooks documentation](https://pytorch.org/docs/stable/ddp_comm_hooks.html).

docs/source/usage_guides/debug.md

@@ -1,93 +0,0 @@
-<!--Copyright 2022 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.
--->
-
-# Debugging Distributed Operations
-
-When running scripts in a distributed fashion, often functions such as [`Accelerator.gather`] and [`Accelerator.reduce`] (and others) are neccessary to grab tensors across devices and perform certain operations on them. However, if the tensors which are being grabbed are not the proper shapes then this will result in your code hanging forever. The only sign that exists of this truly happening is hitting a timeout exception from `torch.distributed`, but this can get quite costly as usually the timeout is 10 minutes.
-
-Accelerate now has a `debug` mode which adds a neglible amount of time to each operation, but allows it to verify that the inputs you are bringing in can *actually* perform the operation you want **without** hitting this timeout problem!
-
-## Visualizing the problem
-
-To have a tangible example of this issue, let's take the following setup (on 2 GPUs):
-
-```python
-from accelerate import PartialState
-
-state = PartialState()
-if state.process_index == 0:
-    tensor = torch.tensor([[0.0, 1, 2, 3, 4]]).to(state.device)
-else:
-    tensor = torch.tensor([[[0.0, 1, 2, 3, 4], [5, 6, 7, 8, 9]]]).to(state.device)
-
-broadcast_tensor = broadcast(tensor)
-print(broadcast_tensor)
-```
-
-We've created a single tensor on each device, with two radically different shapes. With this setup if we want to perform an operation such as [`utils.broadcast`], we would forever hit a timeout because `torch.distributed` requires that these operations have the **exact same shape** across all processes for it to work.
-
-If you run this yourself, you will find that `broadcast_tensor` can be printed on the main process, but its results won't quite be right, and then it will just hang never printing it on any of the other processes:
-
-```
->>> tensor([[0, 1, 2, 3, 4]], device='cuda:0')
-```
-
-## The solution
-
-By enabling Accelerate's operational debug mode, Accelerate will properly find and catch errors such as this and provide a very clear traceback immediatly: 
-
-```
-Traceback (most recent call last):
-  File "/home/zach_mueller_huggingface_co/test.py", line 18, in <module>
-    main()
-  File "/home/zach_mueller_huggingface_co/test.py", line 15, in main
-        main()broadcast_tensor = broadcast(tensor)
-  File "/home/zach_mueller_huggingface_co/accelerate/src/accelerate/utils/operations.py", line 303, in wrapper
-    broadcast_tensor = broadcast(tensor)
-accelerate.utils.operations.DistributedOperationException: Cannot apply desired operation due to shape mismatches. All shapes across devices must be valid.
-
-Operation: `accelerate.utils.operations.broadcast`
-Input shapes:
-  - Process 0: [1, 5]
-  - Process 1: [1, 2, 5]
-```
-
-This explains that the shapes across our devices were *not* the same, and that we should ensure that they match properly to be compatible. Typically this means that there is either an extra dimension, or certain dimensions are incompatible with the operation.
-
-To enable this please do one of the following:
-
-Enable it through the questionarre during `accelerate config` (recommended)
-
-From the CLI: 
-
-```
-accelerate launch --debug {my_script.py} --arg1 --arg2
-```
-
-As an environmental variable (which avoids the need for `accelerate launch`):
-
-```
-ACCELERATE_DEBUG_MODE="1" accelerate launch {my_script.py} --arg1 --arg2
-```
-
-Manually changing the `config.yaml` file:
-
-```diff
- compute_environment: LOCAL_MACHINE
-+debug: true
-```
-
-
-

docs/source/usage_guides/deepspeed.md

@@ -9,13 +9,13 @@ Unless required by applicable law or agreed to in writing, software distributed
 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
+⚠️ 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.
 -->
 
-# DeepSpeed 
+# DeepSpeed
 
-[DeepSpeed](https://github.com/microsoft/DeepSpeed) implements everything described in the [ZeRO paper](https://arxiv.org/abs/1910.02054). Currently, it provides full support for:
+[DeepSpeed](https://github.com/microsoft/DeepSpeed) implements everything described in the [ZeRO paper](https://arxiv.org/abs/1910.02054). Some of the salient optimizations are:
 
 1. Optimizer state partitioning (ZeRO stage 1)
 2. Gradient partitioning (ZeRO stage 2)
@@ -23,6 +23,7 @@ rendered properly in your Markdown viewer.
 4. Custom mixed precision training handling
 5. A range of fast CUDA-extension-based optimizers
 6. ZeRO-Offload to CPU and Disk/NVMe
+7. Hierarchical partitioning of model parameters (ZeRO++)
 
 ZeRO-Offload has its own dedicated paper: [ZeRO-Offload: Democratizing Billion-Scale Model Training](https://arxiv.org/abs/2101.06840). And NVMe-support is described in the paper [ZeRO-Infinity: Breaking the GPU
 Memory Wall for Extreme Scale Deep Learning](https://arxiv.org/abs/2104.07857).
@@ -32,19 +33,19 @@ DeepSpeed ZeRO-2 is primarily used only for training, as its features are of no
 DeepSpeed ZeRO-3 can be used for inference as well since it allows huge models to be loaded on multiple GPUs, which
 won't be possible on a single GPU.
 
-🤗 Accelerate integrates [DeepSpeed](https://github.com/microsoft/DeepSpeed) via 2 options:
+Accelerate integrates [DeepSpeed](https://github.com/microsoft/DeepSpeed) via 2 options:
 
 1. Integration of the DeepSpeed features via `deepspeed config file` specification in `accelerate config` . You just supply your custom config file or use our template. Most of
-   this document is focused on this feature. This supports all the core features of DeepSpeed and gives user a lot of flexibility. 
+   this document is focused on this feature. This supports all the core features of DeepSpeed and gives user a lot of flexibility.
    User may have to change a few lines of code depending on the config.
-2. Integration via `deepspeed_plugin`.This supports subset of the DeepSpeed features and uses default options for the rest of the configurations. 
+2. Integration via `deepspeed_plugin`.This supports subset of the DeepSpeed features and uses default options for the rest of the configurations.
    User need not change any code and is good for those who are fine with most of the default settings of DeepSpeed.
 
 ## What is integrated?
 
 Training:
 
-1. DeepSpeed ZeRO training supports the full ZeRO stages 1, 2 and 3 as well as CPU/Disk offload of optimizer states, gradients and parameters. 
+1. Accelerate integrates all features of DeepSpeed ZeRO. This includes all the ZeRO stages 1, 2 and 3 as well as ZeRO-Offload, ZeRO-Infinity (which can offload to disk/NVMe) and ZeRO++.
 Below is a short description of Data Parallelism using ZeRO - Zero Redundancy Optimizer along with diagram from this [blog post](https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/)
 ![ZeRO Data Parallelism](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-zero.png)
 
@@ -60,6 +61,8 @@ Below is a short description of Data Parallelism using ZeRO - Zero Redundancy Op
 
  e. **Param Offload**: Offloads the model parameters to CPU/Disk building on top of ZERO Stage 3
 
+ f. **Hierarchical Partitioning**: Enables efficient multi-node training with data-parallel training across nodes and ZeRO-3 sharding within a node, built on top of ZeRO Stage 3.
+
 <u>Note</u>: With respect to Disk Offload, the disk should be an NVME for decent speed but it technically works on any Disk
 
 Inference:
@@ -74,8 +77,8 @@ Inference:
 **Pre-Requisites**: Install DeepSpeed version >=0.6.5. Please refer to the [DeepSpeed Installation details](https://github.com/microsoft/DeepSpeed#installation)
 for more information.
 
-We will first look at easy to use integration via `accelerate config`. 
-Followed by more flexible and feature rich `deepspeed config file` integration. 
+We will first look at easy to use integration via `accelerate config`.
+Followed by more flexible and feature rich `deepspeed config file` integration.
 
 ### Accelerate DeepSpeed Plugin
 On your machine(s) just run:
@@ -154,10 +157,18 @@ Currently, `Accelerate` supports following config through the CLI:
 `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_optimizer_nvme_path`: Decides Nvme Path to offload optimizer states. If unspecified, will default to 'none'.
 `offload_param_device`: [none] Disable parameter offloading, [cpu] offload parameters to CPU, [nvme] offload parameters to NVMe SSD. Only applicable with ZeRO Stage-3.
+`offload_param_nvme_path`: Decides Nvme Path to offload parameters. If unspecified, will default to 'none'.
 `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. 
+`mixed_precision`: `no` for FP32 training, `fp16` for FP16 mixed-precision training and `bf16` for BF16 mixed-precision training.
+`deepspeed_moe_layer_cls_names`: Comma-separated list of transformer Mixture-of-Experts (MoE) layer class names (case-sensitive) to wrap ,e.g, `MixtralSparseMoeBlock`, `Qwen2MoeSparseMoeBlock`, `JetMoEAttention,JetMoEBlock` ...
+`deepspeed_hostfile`: DeepSpeed hostfile for configuring multi-node compute resources.
+`deepspeed_exclusion_filter`: DeepSpeed exclusion filter string when using mutli-node setup.
+`deepspeed_inclusion_filter`: DeepSpeed inclusion filter string when using mutli-node setup.
+`deepspeed_multinode_launcher`: DeepSpeed multi-node launcher to use. If unspecified, will default to `pdsh`.
+`deepspeed_config_file`: path to the DeepSpeed config file in `json` format. See the next section for more details on this.
 ```
 To be able to tweak more options, you will need to use a DeepSpeed config file.
 
@@ -168,8 +179,8 @@ On your machine(s) just run:
 accelerate config
 ```
 
-and answer the questions asked. It will ask whether you want to use a config file for deepspeed to which you answer yes 
-and provide the path to the deepspeed config file. 
+and answer the questions asked. It will ask whether you want to use a config file for deepspeed to which you answer yes
+and provide the path to the deepspeed config file.
 This will generate a config file that will be used automatically to properly set the
 default options when doing
 
@@ -349,17 +360,38 @@ accelerate launch examples/by_feature/deepspeed_with_config_support.py \
 --report_to "wandb"\
 ```
 
+**ZeRO++ Config Example**
+You can use the features of ZeRO++ by using the appropriate config parameters. Note that ZeRO++ is an extension for ZeRO Stage 3. Here is how the config file can be modified, from [DeepSpeed's ZeRO++ tutorial](https://www.deepspeed.ai/tutorials/zeropp/):
+
+```json
+{
+    "zero_optimization": {
+        "stage": 3,
+        "reduce_bucket_size": "auto",
+
+        "zero_quantized_weights": true,
+        "zero_hpz_partition_size": 8,
+        "zero_quantized_gradients": true,
+
+        "contiguous_gradients": true,
+        "overlap_comm": true
+    }
+}
+```
+
+For hierarchical partitioning, the partition size `zero_hpz_partition_size` should ideally be set to the number of GPUs per node. (For example, the above config file assumes 8 GPUs per node)
+
 **Important code changes when using DeepSpeed Config File**
 
-1. DeepSpeed Optimizers and Schedulers. For more information on these, 
+1. DeepSpeed Optimizers and Schedulers. For more information on these,
 see the [DeepSpeed Optimizers](https://deepspeed.readthedocs.io/en/latest/optimizers.html) and [DeepSpeed Schedulers](https://deepspeed.readthedocs.io/en/latest/schedulers.html) documentation.
 We will look at the changes needed in the code when using these.
-   
+
    a. DS Optim + DS Scheduler: The case when both `optimizer` and `scheduler` keys are present in the DeepSpeed config file.
    In this situation, those will be used and the user has to use `accelerate.utils.DummyOptim` and `accelerate.utils.DummyScheduler` to replace the PyTorch/Custom optimizers and schedulers in their code.
    Below is the snippet from `examples/by_feature/deepspeed_with_config_support.py` showing this:
    ```python
-    # Creates Dummy Optimizer if `optimizer` was spcified in the config file else creates Adam Optimizer
+    # Creates Dummy Optimizer if `optimizer` was specified in the config file else creates Adam Optimizer
     optimizer_cls = (
         torch.optim.AdamW
         if accelerator.state.deepspeed_plugin is None
@@ -368,7 +400,7 @@ We will look at the changes needed in the code when using these.
     )
     optimizer = optimizer_cls(optimizer_grouped_parameters, lr=args.learning_rate)
 
-    # Creates Dummy Scheduler if `scheduler` was spcified in the config file else creates `args.lr_scheduler_type` Scheduler
+    # Creates Dummy Scheduler if `scheduler` was specified in the config file else creates `args.lr_scheduler_type` Scheduler
     if (
         accelerator.state.deepspeed_plugin is None
         or "scheduler" not in accelerator.state.deepspeed_plugin.deepspeed_config
@@ -388,16 +420,25 @@ We will look at the changes needed in the code when using these.
    In this situation, no code changes are needed from the user and this is the case when using integration via DeepSpeed Plugin.
    In the above example we can see that the code remains unchanged if the `optimizer` and `scheduler` keys are absent in the DeepSpeed config file.
 
-   c. Custom Optim + DS Scheduler: The case when only `scheduler` key is present in the DeepSpeed config file. 
-   In this situation, the user has to use `accelerate.utils.DummyScheduler` to replace the PyTorch/Custom scheduler in their code. 
+   c. Custom Optim + DS Scheduler: The case when only `scheduler` key is present in the DeepSpeed config file.
+   In this situation, the user has to use `accelerate.utils.DummyScheduler` to replace the PyTorch/Custom scheduler in their code.
 
-   d. DS Optim + Custom Scheduler: The case when only `optimizer` key is present in the DeepSpeed config file. 
+   d. DS Optim + Custom Scheduler: The case when only `optimizer` key is present in the DeepSpeed config file.
    This will result in an error because you can only use DS Scheduler when using DS Optim.
 
-2. Notice the `auto` values in the above example DeepSpeed config files. These are automatically handled by `prepare` method 
-based on model, dataloaders, dummy optimizer and dummy schedulers provided to `prepare` method. 
+2. Notice the `auto` values in the above example DeepSpeed config files. These are automatically handled by `prepare` method
+based on model, dataloaders, dummy optimizer and dummy schedulers provided to `prepare` method.
 Only the `auto` fields specified in above examples are handled by `prepare` method and the rest have to be explicitly specified by the user.
 
+The `auto` values are calculated as:
+
+- `reduce_bucket_size`: `hidden_size * hidden_size`
+- `stage3_prefetch_bucket_size`: `int(0.9 * hidden_size * hidden_size)`
+- `stage3_param_persistence_threshold`: `10 * hidden_size`
+
+For the `auto` feature to work for these 3 config entries - Accelerate will use `model.config.hidden_size` or `max(model.config.hidden_sizes)` as `hidden_size`. If neither of these is available, the launching will fail and you will have to set these 3 config entries manually. Remember the first 2 config entries are the communication buffers - the larger they are the more efficient the comms will be, and the larger they are the more GPU memory they will consume, so it's a tunable performance trade-off.
+
+
 **Things to note when using DeepSpeed Config File**
 
 Below is a sample script using `deepspeed_config_file` in different scenarios.
@@ -482,11 +523,11 @@ use_cpu: false
 3. Output of `accelerate launch test.py`:
 
 ```bash
-ValueError: When using `deepspeed_config_file`, the following accelerate config variables will be ignored: 
-['gradient_accumulation_steps', 'gradient_clipping', 'zero_stage', 'offload_optimizer_device', 'offload_param_device', 
+ValueError: When using `deepspeed_config_file`, the following accelerate config variables will be ignored:
+['gradient_accumulation_steps', 'gradient_clipping', 'zero_stage', 'offload_optimizer_device', 'offload_param_device',
 'zero3_save_16bit_model', 'mixed_precision'].
 Please specify them appropriately in the DeepSpeed config file.
-If you are using an accelerate config file, remove others config variables mentioned in the above specified list.
+If you are using an accelerate config file, remove other config variables mentioned in the above specified list.
 The easiest method is to create a new config following the questionnaire via `accelerate config`.
 It will only ask for the necessary config variables when using `deepspeed_config_file`.
 ```
@@ -499,15 +540,15 @@ It will only ask for the necessary config variables when using `deepspeed_config
 $ accelerate config
 -------------------------------------------------------------------------------------------------------------------------------
 In which compute environment are you running?
-This machine                                                                                                                   
+This machine
 -------------------------------------------------------------------------------------------------------------------------------
-Which type of machine are you using?                                                                                           
-multi-GPU                                                                                                                      
-How many different machines will you use (use more than 1 for multi-node training)? [1]:                                       
-Do you wish to optimize your script with torch dynamo?[yes/NO]:                                                                
-Do you want to use DeepSpeed? [yes/NO]: yes                                                                                    
-Do you want to specify a json file to a DeepSpeed config? [yes/NO]: yes                                                        
-Please enter the path to the json DeepSpeed config file: ds_config.json                                                        
+Which type of machine are you using?
+multi-GPU
+How many different machines will you use (use more than 1 for multi-node training)? [1]:
+Do you wish to optimize your script with torch dynamo?[yes/NO]:
+Do you want to use DeepSpeed? [yes/NO]: yes
+Do you want to specify a json file to a DeepSpeed config? [yes/NO]: yes
+Please enter the path to the json DeepSpeed config file: ds_config.json
 Do you want to enable `deepspeed.zero.Init` when using ZeRO Stage-3 for constructing massive models? [yes/NO]: yes
 How many GPU(s) should be used for distributed training? [1]:4
 accelerate configuration saved at ds_config_sample.yaml
@@ -585,10 +626,10 @@ Mixed precision type: fp16
 ds_config: {'bf16': {'enabled': False}, 'zero_optimization': {'stage': 3, 'stage3_gather_16bit_weights_on_model_save': True, 'offload_optimizer': {'device': 'nvme'}, 'offload_param': {'device': 'cpu'}}, 'gradient_clipping': 1.0, 'train_batch_size': 'auto', 'train_micro_batch_size_per_gpu': 'auto', 'gradient_accumulation_steps': 5, 'steps_per_print': inf, 'fp16': {'enabled': True, 'auto_cast': True}}
 ```
 
-**Note**: 
-1. Remaining `"auto"` values are handled in `accelerator.prepare()` call as explained in point 2 of 
+**Note**:
+1. Remaining `"auto"` values are handled in `accelerator.prepare()` call as explained in point 2 of
 `Important code changes when using DeepSpeed Config File`.
-2. Only when `gradient_accumulation_steps` is `auto`, the value passed while creating `Accelerator` object via `Accelerator(gradient_accumulation_steps=k)` will be used. When using DeepSpeed Plugin, the value from it will be used and it will overwrite the value passed while creating Accelerator object. 
+2. Only when `gradient_accumulation_steps` is `auto`, the value passed while creating `Accelerator` object via `Accelerator(gradient_accumulation_steps=k)` will be used. When using DeepSpeed Plugin, the value from it will be used and it will overwrite the value passed while creating Accelerator object.
 
 ## Saving and loading
 
@@ -599,7 +640,7 @@ ZeRO Stage-3 has 2 options:
 
    a. Saving the entire 16bit model weights to directly load later on using `model.load_state_dict(torch.load(pytorch_model.bin))`.
    For this, either set `zero_optimization.stage3_gather_16bit_weights_on_model_save` to True in DeepSpeed Config file or set
-   `zero3_save_16bit_model` to True in DeepSpeed Plugin. 
+   `zero3_save_16bit_model` to True in DeepSpeed Plugin.
    **Note that this option requires consolidation of the weights on one GPU it can be slow and memory demanding, so only use this feature when needed.**
    Below is the snippet from `examples/by_feature/deepspeed_with_config_support.py` showing this:
    ```python
@@ -623,15 +664,15 @@ ZeRO Stage-3 has 2 options:
    Below is the snippet from `examples/by_feature/deepspeed_with_config_support.py` showing this:
    ```python
    success = model.save_checkpoint(PATH, ckpt_id, checkpoint_state_dict)
-   status_msg = "checkpointing: PATH={}, ckpt_id={}".format(PATH, ckpt_id)
+   status_msg = f"checkpointing: PATH={PATH}, ckpt_id={ckpt_id}"
    if success:
        logging.info(f"Success {status_msg}")
    else:
        logging.warning(f"Failure {status_msg}")
-   ``` 
+   ```
    This will create ZeRO model and optimizer partitions along with `zero_to_fp32.py` script in checkpoint directory.
-   You can use this script to do offline consolidation.  
-   It requires no configuration files or GPUs. Here is an example of its usage:  
+   You can use this script to do offline consolidation.
+   It requires no configuration files or GPUs. Here is an example of its usage:
    ```bash
    $ cd /path/to/checkpoint_dir
    $ ./zero_to_fp32.py . pytorch_model.bin
@@ -655,7 +696,7 @@ ZeRO Stage-3 has 2 options:
    Note that all these functions require ~2x memory (general RAM) of the size of the final checkpoint.
 
 ## ZeRO Inference
-DeepSpeed ZeRO Inference supports ZeRO stage 3 with ZeRO-Infinity. 
+DeepSpeed ZeRO Inference supports ZeRO stage 3 with ZeRO-Infinity.
 It uses the same ZeRO protocol as training, but it doesn't use an optimizer and a lr scheduler and only stage 3 is relevant.
 With accelerate integration, you just need to prepare the model and dataloader as shown below:
 
@@ -663,11 +704,11 @@ With accelerate integration, you just need to prepare the model and dataloader a
 model, eval_dataloader = accelerator.prepare(model, eval_dataloader)
 ```
 
-## Few caveats to be aware of 
+## Few caveats to be aware of
 
 1. Current integration doesn’t support Pipeline Parallelism of DeepSpeed.
-2. Current integration doesn’t support `mpu`, limiting the tensor parallelism which is supported in Megatron-LM. 
-3. Current integration doesn’t support multiple models. 
+2. Current integration doesn’t support `mpu`, limiting the tensor parallelism which is supported in Megatron-LM.
+3. Current integration doesn’t support multiple models.
 
 ## DeepSpeed Resources
 
@@ -683,7 +724,15 @@ Papers:
 - [ZeRO: Memory Optimizations Toward Training Trillion Parameter Models](https://arxiv.org/abs/1910.02054)
 - [ZeRO-Offload: Democratizing Billion-Scale Model Training](https://arxiv.org/abs/2101.06840)
 - [ZeRO-Infinity: Breaking the GPU Memory Wall for Extreme Scale Deep Learning](https://arxiv.org/abs/2104.07857)
+- [ZeRO++: Extremely Efficient Collective Communication for Giant Model Training](https://arxiv.org/abs/2306.10209)
 
-Finally, please, remember that 🤗 `Accelerate` only integrates DeepSpeed, therefore if you
+
+Finally, please, remember that `Accelerate` only integrates DeepSpeed, therefore if you
 have any problems or questions with regards to DeepSpeed usage, please, file an issue with [DeepSpeed GitHub](https://github.com/microsoft/DeepSpeed/issues).
 
+
+<Tip>
+
+    For those interested in the similarities and differences between FSDP and DeepSpeed, please check out the [concept guide here](../concept_guides/fsdp_and_deepspeed)!
+    
+</Tip>

docs/source/usage_guides/deepspeed_multiple_model.md

@@ -0,0 +1,246 @@
+<!--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 contains specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# Using multiple models with DeepSpeed
+
+<Tip warning={true}>
+
+    This guide assumes that you have read and understood the [DeepSpeed usage guide](./deepspeed.md).
+
+</Tip>
+
+Running multiple models with Accelerate and DeepSpeed is useful for:
+
+* Knowledge distillation
+* Post-training techniques like RLHF (see the [TRL](https://github.com/huggingface/trl) library for more examples)
+* Training multiple models at once
+
+Currently, Accelerate has a **very experimental API** to help you use multiple models.
+
+This tutorial will focus on two common use cases:
+
+1. Knowledge distillation, where a smaller student model is trained to mimic a larger, better-performing teacher.  If the student model fits on a single GPU, we can use ZeRO-2 for training and ZeRO-3 to shard the teacher for inference. This is significantly faster than using ZeRO-3 for both models.
+2. Training multiple *disjoint* models at once.
+
+## Knowledge distillation
+
+Knowledge distillation is a good example of using multiple models, but only training one of them.
+
+Normally, you would use a single [`utils.DeepSpeedPlugin`] for both models. However, in this case, there are two separate configurations. Accelerate allows you to create and use multiple plugins **if and only if** they are in a `dict` so that you can reference and enable the proper plugin when needed.
+
+```python
+from accelerate.utils import DeepSpeedPlugin
+
+zero2_plugin = DeepSpeedPlugin(hf_ds_config="zero2_config.json")
+zero3_plugin = DeepSpeedPlugin(hf_ds_config="zero3_config.json")
+
+deepspeed_plugins = {"student": zero2_plugin, "teacher": zero3_plugin}
+```
+
+The `zero2_config.json` should be configured for full training (so specify `scheduler` and `optimizer` if you are not utilizing your own), while `zero3_config.json` should only be configured for the inference model, as shown in the example below.
+
+```json
+{
+    "bf16": {
+        "enabled": "auto"
+    },
+    "zero_optimization": {
+        "stage": 3,
+        "overlap_comm": true,
+        "reduce_bucket_size": "auto",
+        "stage3_prefetch_bucket_size": "auto",
+        "stage3_param_persistence_threshold": "auto",
+        "stage3_max_live_parameters": "auto",
+        "stage3_max_reuse_distance": "auto",
+    },
+    "train_micro_batch_size_per_gpu": 1
+}
+```
+
+An example `zero2_config.json` configuration is shown below.
+
+```json
+{
+    "bf16": {
+        "enabled": "auto"
+    },
+    "optimizer": {
+        "type": "AdamW",
+        "params": {
+            "lr": "auto",
+            "weight_decay": "auto",
+            "torch_adam": true,
+            "adam_w_mode": true
+        }
+    },
+    "scheduler": {
+        "type": "WarmupLR",
+        "params": {
+            "warmup_min_lr": "auto",
+            "warmup_max_lr": "auto",
+            "warmup_num_steps": "auto"
+        }
+    },
+    "zero_optimization": {
+        "stage": 2,
+        "offload_optimizer": {
+            "device": "cpu",
+            "pin_memory": true
+        },
+    },
+    "gradient_accumulation_steps": 1,
+    "gradient_clipping": "auto",
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+}
+```
+
+<Tip>
+
+    DeepSpeed will raise an error if `train_micro_batch_size_per_gpu` isn't specified, even if this particular model isn't being trained.
+
+</Tip>
+
+From here, create a single [`Accelerator`] and pass in both configurations.
+
+```python
+from accelerate import Accelerator
+
+accelerator = Accelerator(deepspeed_plugins=deepspeed_plugins)
+```
+
+Now let's see how to use them.
+
+### Student model
+
+By default, Accelerate sets the first item in the `dict` as the default or enabled plugin (`"student"` plugin). Verify this by using the [`utils.deepspeed.get_active_deepspeed_plugin`] function to see which plugin is enabled.
+
+```python
+active_plugin = get_active_deepspeed_plugin(accelerator.state)
+assert active_plugin is deepspeed_plugins["student"]
+```
+
+[`AcceleratorState`] also keeps the active DeepSpeed plugin saved in `state.deepspeed_plugin`.
+```python
+assert active_plugin is accelerator.deepspeed_plugin
+```
+
+Since `student` is the currently active plugin, let's go ahead and prepare the model, optimizer, and scheduler.
+
+```python
+student_model, optimizer, scheduler = ...
+student_model, optimizer, scheduler, train_dataloader = accelerator.prepare(student_model, optimizer, scheduler, train_dataloader)
+```
+
+Now it's time to deal with the teacher model.
+
+### Teacher model
+
+First, you need to specify in [`Accelerator`] that the `zero3_config.json` configuration should be used.
+
+```python
+accelerator.state.select_deepspeed_plugin("teacher")
+```
+
+This disables the `"student"` plugin and enables the `"teacher"` plugin instead. The
+DeepSpeed stateful config inside of Transformers is updated, and it changes which plugin configuration gets called when using
+`deepspeed.initialize()`. This allows you to use the automatic `deepspeed.zero.Init`  context manager integration Transformers provides.
+
+```python
+teacher_model = AutoModel.from_pretrained(...)
+teacher_model = accelerator.prepare(teacher_model)
+```
+
+Otherwise, you should manually initialize the model with `deepspeed.zero.Init`.
+```python
+with deepspeed.zero.Init(accelerator.deepspeed_plugin.config):
+    model = MyModel(...)
+```
+
+### Training
+
+From here, your training loop can be whatever you like, as long as `teacher_model` is never being trained on.
+
+```python
+teacher_model.eval()
+student_model.train()
+for batch in train_dataloader:
+    with torch.no_grad():
+        output_teacher = teacher_model(**batch)
+    output_student = student_model(**batch)
+    # Combine the losses or modify it in some way
+    loss = output_teacher.loss + output_student.loss
+    accelerator.backward(loss)
+    optimizer.step()
+    scheduler.step()
+    optimizer.zero_grad()
+```
+
+## Train multiple disjoint models
+
+Training multiple models is a more complicated scenario.
+In its current state, we assume each model is **completely disjointed** from the other during training.
+
+This scenario still requires two [`utils.DeepSpeedPlugin`]'s to be made. However, you also need a second [`Accelerator`], since different `deepspeed` engines are being called at different times. A single [`Accelerator`] can only carry one instance at a time.
+
+Since the [`state.AcceleratorState`] is a stateful object though, it is already aware of both [`utils.DeepSpeedPlugin`]'s available. You can just instantiate a second [`Accelerator`] with no extra arguments.
+
+```python
+first_accelerator = Accelerator(deepspeed_plugins=deepspeed_plugins)
+second_accelerator = Accelerator()
+```
+
+You can call either `first_accelerator.state.select_deepspeed_plugin()` to enable or disable
+a particular plugin, and then call [`prepare`].
+
+```python
+# can be `accelerator_0`, `accelerator_1`, or by calling `AcceleratorState().select_deepspeed_plugin(...)`
+first_accelerator.state.select_deepspeed_plugin("first_model")
+first_model = AutoModel.from_pretrained(...)
+# For this example, `get_training_items` is a nonexistent function that gets the setup we need for training
+first_optimizer, first_scheduler, train_dl, eval_dl = get_training_items(model1)
+first_model, first_optimizer, first_scheduler, train_dl, eval_dl = accelerator.prepare(
+    first_model, first_optimizer, first_scheduler, train_dl, eval_dl
+)
+
+second_accelerator.state.select_deepspeed_plugin("second_model")
+second_model = AutoModel.from_pretrained(...)
+# For this example, `get_training_items` is a nonexistent function that gets the setup we need for training
+second_optimizer, second_scheduler, _, _ = get_training_items(model2)
+second_model, second_optimizer, second_scheduler = accelerator.prepare(
+    second_model, second_optimizer, second_scheduler
+)
+```
+
+And now you can train:
+
+```python
+for batch in dl:
+    outputs1 = first_model(**batch)
+    first_accelerator.backward(outputs1.loss)
+    first_optimizer.step()
+    first_scheduler.step()
+    first_optimizer.zero_grad()
+    
+    outputs2 = model2(**batch)
+    second_accelerator.backward(outputs2.loss)
+    second_optimizer.step()
+    second_scheduler.step()
+    second_optimizer.zero_grad()
+```
+
+## Resources
+
+To see more examples, please check out the [related tests](https://github.com/huggingface/accelerate/blob/main/src/accelerate/test_utils/scripts/external_deps/test_ds_multiple_model.py) currently in [Accelerate].

docs/source/usage_guides/distributed_inference.md

@@ -13,14 +13,20 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Distributed Inference with 🤗 Accelerate
+# Distributed inference
 
-Distributed inference is a common use case, especially with natural language processing (NLP) models. Users often want to
-send a number of different prompts, each to a different GPU, and then get the results back. This also has other cases
-outside of just NLP, however for this tutorial we will focus on just this idea of each GPU receiving a different prompt,
-and then returning the results.
+Distributed inference can fall into three brackets:
 
-## The Problem
+1. Loading an entire model onto each GPU and sending chunks of a batch through each GPU's model copy at a time
+2. Loading parts of a model onto each GPU and processing a single input at one time
+3. Loading parts of a model onto each GPU and using what is called scheduled Pipeline Parallelism to combine the two prior techniques. 
+
+We're going to go through the first and the last bracket, showcasing how to do each as they are more realistic scenarios.
+
+
+## Sending chunks of a batch automatically to each loaded model
+
+This is the most memory-intensive solution, as it requires each GPU to keep a full copy of the model in memory at a given time. 
 
 Normally when doing this, users send the model to a specific device to load it from the CPU, and then move each prompt to a different device. 
 
@@ -50,14 +56,13 @@ def run_inference(rank, world_size):
 ```
 One will notice how we have to check the rank to know what prompt to send, which can be a bit tedious.
 
-A user might then also think that with 🤗 Accelerate, using the `Accelerator` to prepare a dataloader for such a task might also be 
-a simple way to manage this. (To learn more, check out the relvent section in the [Quick Tour](../quicktour#distributed-evaluation))
+A user might then also think that with Accelerate, using the `Accelerator` to prepare a dataloader for such a task might also be 
+a simple way to manage this. (To learn more, check out the relevant section in the [Quick Tour](../quicktour#distributed-evaluation))
 
 Can it manage it? Yes. Does it add unneeded extra code however: also yes.
 
-## The Solution
 
-With 🤗 Accelerate, we can simplify this process by using the [`Accelerator.split_between_processes`] context manager (which also exists in `PartialState` and `AcceleratorState`). 
+With Accelerate, we can simplify this process by using the [`Accelerator.split_between_processes`] context manager (which also exists in `PartialState` and `AcceleratorState`). 
 This function will automatically split whatever data you pass to it (be it a prompt, a set of tensors, a dictionary of the prior data, etc.) across all the processes (with a potential
 to be padded) for you to use right away.
 
@@ -77,7 +82,7 @@ with distributed_state.split_between_processes(["a dog", "a cat"]) as prompt:
     result.save(f"result_{distributed_state.process_index}.png")
 ```
 
-And then to launch the code, we can use the 🤗 Accelerate:
+And then to launch the code, we can use the Accelerate:
 
 If you have generated a config file to be used using `accelerate config`:
 
@@ -134,3 +139,97 @@ with distributed_state.split_between_processes(["a dog", "a cat", "a chicken"],
 
 On the first GPU, the prompts will be `["a dog", "a cat"]`, and on the second GPU it will be `["a chicken", "a chicken"]`.
 Make sure to drop the final sample, as it will be a duplicate of the previous one.
+
+You can find more complex examples [here](https://github.com/huggingface/accelerate/tree/main/examples/inference/distributed) such as how to use it with LLMs.
+
+## Memory-efficient pipeline parallelism (experimental)
+
+This next part will discuss using *pipeline parallelism*. This is an **experimental** API that utilizes [torch.distributed.pipelining](https://pytorch.org/docs/stable/distributed.pipelining.html#) as a native solution. 
+
+The general idea with pipeline parallelism is: say you have 4 GPUs and a model big enough it can be *split* on four GPUs using `device_map="auto"`. With this method you can send in 4 inputs at a time (for example here, any amount works) and each model chunk will work on an input, then receive the next input once the prior chunk finished, making it *much* more efficient **and faster** than the method described earlier. Here's a visual taken from the PyTorch repository:
+
+![Pipeline parallelism example](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/pipeline_parallel.png)
+
+To illustrate how you can use this with Accelerate, we have created an [example zoo](https://github.com/huggingface/accelerate/tree/main/examples/inference) showcasing a number of different models and situations. In this tutorial, we'll show this method for GPT2 across two GPUs.
+
+Before you proceed, please make sure you have the latest PyTorch version installed by running the following:
+
+```bash
+pip install torch
+```
+
+Start by creating the model on the CPU:
+
+```{python}
+from transformers import GPT2ForSequenceClassification, GPT2Config
+
+config = GPT2Config()
+model = GPT2ForSequenceClassification(config)
+model.eval()
+```
+
+Next you'll need to create some example inputs to use. These help `torch.distributed.pipelining` trace the model.
+
+<Tip warning={true}>
+    However you make this example will determine the relative batch size that will be used/passed
+    through the model at a given time, so make sure to remember how many items there are!
+</Tip>
+
+```{python}
+input = torch.randint(
+    low=0,
+    high=config.vocab_size,
+    size=(2, 1024),  # bs x seq_len
+    device="cpu",
+    dtype=torch.int64,
+    requires_grad=False,
+)
+```
+Next we need to actually perform the tracing and get the model ready. To do so, use the [`inference.prepare_pippy`] function and it will fully wrap the model for pipeline parallelism automatically:
+
+```{python}
+from accelerate.inference import prepare_pippy
+example_inputs = {"input_ids": input}
+model = prepare_pippy(model, example_args=(input,))
+```
+
+<Tip>
+
+    There are a variety of parameters you can pass through to `prepare_pippy`:
+    
+    * `split_points` lets you determine what layers to split the model at. By default we use wherever `device_map="auto" declares, such as `fc` or `conv1`.
+
+    * `num_chunks` determines how the batch will be split and sent to the model itself (so `num_chunks=1` with four split points/four GPUs will have a naive MP where a single input gets passed between the four layer split points)
+
+</Tip>
+
+From here, all that's left is to actually perform the distributed inference!
+
+<Tip warning={true}>
+
+When passing inputs, we highly recommend to pass them in as a tuple of arguments. Using `kwargs` is supported, however, this approach is experimental.
+</Tip>
+
+```{python}
+args = some_more_arguments
+with torch.no_grad():
+    output = model(*args)
+```
+
+When finished all the data will be on the last process only:
+
+```{python}
+from accelerate import PartialState
+if PartialState().is_last_process:
+    print(output)
+```
+
+<Tip>
+
+    If you pass in `gather_output=True` to [`inference.prepare_pippy`], the output will be sent
+    across to all the GPUs afterwards without needing the `is_last_process` check. This is 
+    `False` by default as it incurs a communication call.
+    
+</Tip>
+
+And that's it! To explore more, please check out the inference examples in the [Accelerate repo](https://github.com/huggingface/accelerate/tree/main/examples/inference/pippy) and our [documentation](../package_reference/inference) as we work to improving this integration. 

docs/source/usage_guides/explore.md

@@ -13,14 +13,14 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Learning how to incorporate 🤗 Accelerate features quickly!
+# Start Here!
 
 Please use the interactive tool below to help you get started with learning about a particular 
-feature of 🤗 Accelerate and how to utilize it! It will provide you with a code diff, an explaination
+feature of Accelerate and how to utilize it! It will provide you with a code diff, an explanation
 towards what is going on, as well as provide you with some useful links to explore more within
 the documentation!
 
-Most code examples start from the following python code before integrating 🤗 Accelerate in some way:
+Most code examples start from the following python code before integrating Accelerate in some way:
 
 ```python
 for batch in dataloader:

docs/source/usage_guides/fsdp.md

@@ -36,27 +36,34 @@ default options when doing
 accelerate launch my_script.py --args_to_my_script
 ```
 
-For instance, here is how you would run the NLP example (from the root of the repo) with FSDP enabled:
+For instance, here is how you would run `examples/nlp_example.py` (from the root of the repo) with FSDP enabled:
 
 ```bash
 compute_environment: LOCAL_MACHINE
-deepspeed_config: {}
+debug: false
 distributed_type: FSDP
 downcast_bf16: 'no'
 fsdp_config:
   fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
   fsdp_backward_prefetch_policy: BACKWARD_PRE
+  fsdp_forward_prefetch: false
+  fsdp_cpu_ram_efficient_loading: true
   fsdp_offload_params: false
-  fsdp_sharding_strategy: 1
-  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sync_module_states: true
   fsdp_transformer_layer_cls_to_wrap: BertLayer
+  fsdp_use_orig_params: true
 machine_rank: 0
-main_process_ip: null
-main_process_port: null
 main_training_function: main
-mixed_precision: 'no'
+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
 ```
 
@@ -66,40 +73,30 @@ accelerate launch examples/nlp_example.py
 
 Currently, `Accelerate` supports the following config through the CLI:
 
-```bash
-`Sharding Strategy`: [1] FULL_SHARD (shards optimizer states, gradients and parameters), [2] SHARD_GRAD_OP (shards optimizer states and gradients), [3] NO_SHARD (DDP), [4] HYBRID_SHARD (shards optimizer states, gradients and parameters within each node while each node has full copy), [5] HYBRID_SHARD_ZERO2 (shards optimizer states and gradients within each node while each node has full copy)
+`fsdp_sharding_strategy`: [1] FULL_SHARD (shards optimizer states, gradients and parameters), [2] SHARD_GRAD_OP (shards optimizer states and gradients), [3] NO_SHARD (DDP), [4] HYBRID_SHARD (shards optimizer states, gradients and parameters within each node while each node has full copy), [5] HYBRID_SHARD_ZERO2 (shards optimizer states and gradients within each node while each node has full copy). For more information, please refer the official [PyTorch docs](https://pytorch.org/docs/stable/fsdp.html#torch.distributed.fsdp.ShardingStrategy).
 
-`Offload Params`: Decides Whether to offload parameters and gradients to CPU
+`fsdp_offload_params` : Decides Whether to offload parameters and gradients to CPU
 
-`Auto Wrap Policy`: [1] TRANSFORMER_BASED_WRAP, [2] SIZE_BASED_WRAP, [3] NO_WRAP
+`fsdp_auto_wrap_policy`: [1] TRANSFORMER_BASED_WRAP, [2] SIZE_BASED_WRAP, [3] NO_WRAP
 
-`Transformer Layer Class to Wrap`: When using `TRANSFORMER_BASED_WRAP`, user specifies comma-separated string of transformer layer class names (case-sensitive) to wrap ,e.g, 
-`BertLayer`, `GPTJBlock`, `T5Block`, `BertLayer,BertEmbeddings,BertSelfOutput`...
-This is important because submodules that share weights (e.g., embedding layer) should not end up in different FSDP wrapped units.
-Using this policy, wrapping happens for each block containing Multi-Head Attention followed by couple of MLP layers. 
-Remaining layers including the shared embeddings are conveniently wrapped in same outermost FSDP unit.
-Therefore, use this for transformer based models.
-You can use the `model._no_split_modules` for 🤗 Transformer models by answering `yes` to 
-`Do you want to use the model's `_no_split_modules` to wrap. Only applicable for 🤗 Transformers`. 
-It will try to use `model._no_split_modules` when available.  
+`fsdp_transformer_layer_cls_to_wrap`: Only applicable for Transformers. When using `fsdp_auto_wrap_policy=TRANSFORMER_BASED_WRAP`, a user may provide a comma-separated string of transformer layer class names (case-sensitive) to wrap, e.g., `BertLayer`, `GPTJBlock`, `T5Block`, `BertLayer,BertEmbeddings,BertSelfOutput`. This is important because submodules that share weights (e.g., embedding layers) should not end up in different FSDP wrapped units. Using this policy, wrapping happens for each block containing Multi-Head Attention followed by a couple of MLP layers. Remaining layers including the shared embeddings are conveniently wrapped in same outermost FSDP unit. Therefore, use this for transformer-based models. You can use the `model._no_split_modules` for Transformer models by answering `yes` to `Do you want to use the model's `_no_split_modules` to wrap. It will try to use `model._no_split_modules` when possible.
 
-`Min Num Params`: minimum number of parameters when using `SIZE_BASED_WRAP`
+`fsdp_min_num_params`: minimum number of parameters when using `fsdp_auto_wrap_policy=SIZE_BASED_WRAP`.
 
-`Backward Prefetch`: [1] BACKWARD_PRE, [2] BACKWARD_POST, [3] NO_PREFETCH
+`fsdp_backward_prefetch_policy`: [1] BACKWARD_PRE, [2] BACKWARD_POST, [3] NO_PREFETCH
 
-`State Dict Type`: [1] FULL_STATE_DICT, [2] LOCAL_STATE_DICT, [3] SHARDED_STATE_DICT 
+`fsdp_forward_prefetch`: if True, then FSDP explicitly prefetches the next upcoming all-gather while executing in the forward pass. Should only be used for static-graph models since the prefetching follows the first iteration’s execution order. i.e., if the sub-modules' order changes dynamically during the model's execution do not enable this feature.
 
-`Forward Prefetch`: if True, then FSDP explicitly prefetches the next upcoming
-all-gather while executing in the forward pass. only use with Static graphs.
+`fsdp_state_dict_type`: [1] FULL_STATE_DICT, [2] LOCAL_STATE_DICT, [3] SHARDED_STATE_DICT
 
-`Use Orig Params`: If True, allows non-uniform `requires_grad` during init, which means support for interspersed frozen and trainable paramteres. 
-Useful in cases such as parameter-efficient fine-tuning. 
-Please refer this [blog](https://dev-discuss.pytorch.org/t/rethinking-pytorch-fully-sharded-data-parallel-fsdp-from-first-principles/1019)
+`fsdp_use_orig_params`: If True, allows non-uniform `requires_grad` during init, which means support for interspersed frozen and trainable parameters. This setting is useful in cases such as parameter-efficient fine-tuning as discussed in [this post](https://dev-discuss.pytorch.org/t/rethinking-pytorch-fully-sharded-data-parallel-fsdp-from-first-principles/1019). This option also allows one to have multiple optimizer param groups. This should be `True` when creating an optimizer before preparing/wrapping the model with FSDP.
 
-`Sync Module States`: If True, each individually wrapped FSDP unit will broadcast module parameters from rank 0
-```
+`fsdp_cpu_ram_efficient_loading`: Only applicable for Transformers models. If True, only the first process loads the pretrained model checkpoint while all other processes have empty weights. This should be set to False if you experience errors when loading the pretrained Transformers model via `from_pretrained` method. When this setting is True `fsdp_sync_module_states` also must to be True, otherwise all the processes except the main process would have random weights leading to unexpected behaviour during training. For this to work, make sure the distributed process group is initialized before calling Transformers `from_pretrained` method. When using Trainer API, the distributed process group is initialized when you create an instance of `TrainingArguments` class.
 
-For additional and more nuanced control, you can specify other FSDP parameters via `FullyShardedDataParallelPlugin`. 
+`fsdp_sync_module_states`: If True, each individually wrapped FSDP unit will broadcast module parameters from rank 0.
+
+
+For additional and more nuanced control, you can specify other FSDP parameters via `FullyShardedDataParallelPlugin`.
 When creating `FullyShardedDataParallelPlugin` object, pass it the parameters that weren't part of the accelerate config or if you want to override them.
 The FSDP parameters will be picked based on the accelerate config file or launch command arguments and other parameters that you will pass directly through the `FullyShardedDataParallelPlugin` object will set/override that.
 
@@ -126,9 +123,9 @@ Below is the code snippet to save using `save_state` utility of accelerate.
 accelerator.save_state("ckpt")
 ```
 
-Inspect the ckeckpoint folder to see model and optimizer as shards per process:
+Inspect the checkpoint folder to see model and optimizer as shards per process:
 ```
-ls ckpt 
+ls ckpt
 # optimizer_0  pytorch_model_0  random_states_0.pkl  random_states_1.pkl  scheduler.bin
 
 cd ckpt
@@ -146,7 +143,7 @@ To load them back for resuming the training, use the `load_state` utility of acc
 accelerator.load_state("ckpt")
 ```
 
-When using transformers `save_pretrained`, pass `state_dict=accelerator.get_state_dict(model)` to save the model state dict. 
+When using transformers `save_pretrained`, pass `state_dict=accelerator.get_state_dict(model)` to save the model state dict.
   Below is an example:
 
 ```diff
@@ -160,67 +157,44 @@ When using transformers `save_pretrained`, pass `state_dict=accelerator.get_stat
 
 ### State Dict
 
-`accelerator.get_state_dict` will call the underlying `model.state_dict` implementation.  With a model wrapped by FSDP, the default behavior of `state_dict` is to gather all of the state in the rank 0 device.  This can cause CUDA out of memory errors if the parameters don't fit on a single GPU.
-
-To avoid this, PyTorch provides a context manager that adjusts the behavior of `state_dict`.  To offload some of the state dict onto CPU, you can use the following code:
-
-```
-from torch.distributed.fsdp import FullyShardedDataParallel as FSDP, StateDictType, FullStateDictConfig
-
-full_state_dict_config = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
-with FSDP.state_dict_type(unwrapped_model, StateDictType.FULL_STATE_DICT, full_state_dict_config):
-    state = accelerator.get_state_dict(unwrapped_model)
-```
+`accelerator.get_state_dict` will call the underlying `model.state_dict` implementation using `FullStateDictConfig(offload_to_cpu=True, rank0_only=True)` context manager to get the state dict only for rank 0 and it will be offloaded to CPU.
 
 You can then pass `state` into the `save_pretrained` method.  There are several modes for `StateDictType` and `FullStateDictConfig` that you can use to control the behavior of `state_dict`.  For more information, see the [PyTorch documentation](https://pytorch.org/docs/stable/fsdp.html).
 
-## A few caveats to be aware of
+If you choose to use `StateDictType.SHARDED_STATE_DICT`, the weights of the model during `Accelerator.save_state` will be split into `n` files for each sub-split on the model. To merge them back into
+a single dictionary to load back into the model later after training you can use the `merge_weights` utility:
 
-- PyTorch FSDP auto wraps sub-modules, flattens the parameters and shards the parameters in place.
-  Due to this, any optimizer created before model wrapping gets broken and occupies more memory.
-  Hence, it is highly recommended and efficient to prepare the model before creating the optimizer.
-  `Accelerate` will automatically wrap the model and create an optimizer for you in case of single model with a warning message.
-  > FSDP Warning: When using FSDP, it is efficient and recommended to call prepare for the model before creating the optimizer
+```py
+from accelerate.utils import merge_fsdp_weights
 
-However, below is the recommended way to prepare model and optimizer while using FSDP:
+# Our weights are saved usually in a `pytorch_model_fsdp_{model_number}` folder
+merge_fsdp_weights("pytorch_model_fsdp_0", "output_path", safe_serialization=True)
+```
+The final output will then either be saved to `model.safetensors` or `pytorch_model.bin` (if `safe_serialization=False` is passed). 
 
-```diff
-  model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", return_dict=True)
-+ model = accelerator.prepare(model)
-
-  optimizer = torch.optim.AdamW(params=model.parameters(), lr=lr)
-
-- model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
--        model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
--    )
-
-+ optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
-+         optimizer, train_dataloader, eval_dataloader, lr_scheduler
-+    )
+This can also be called using the CLI:
+```bash
+accelerate merge-weights pytorch_model_fsdp_0/ output_path
 ```
 
-- In case of a single model, if you have created the optimizer with multiple parameter groups and called prepare with them together,
-  then the parameter groups will be lost and the following warning is displayed:
-  > FSDP Warning: When using FSDP, several parameter groups will be conflated into
-  > a single one due to nested module wrapping and parameter flattening.
-  
-  This is because parameter groups created before wrapping will have no meaning post wrapping due to parameter flattening of nested FSDP modules into 1D arrays (which can consume many layers).
-  For instance, below are the named parameters of an FSDP model on GPU 0 (When using 2 GPUs. Around 55M (110M/2) params in 1D arrays as this will have the 1st shard of the parameters). 
-  Here, if one has applied no weight decay for [bias, LayerNorm.weight] the named parameters of an unwrapped BERT model, 
-  it can't be applied to the below FSDP wrapped model as there are no named parameters with either of those strings and 
-  the parameters of those layers are concatenated with parameters of various other layers.
-  ```
-  {
-    '_fsdp_wrapped_module.flat_param': torch.Size([494209]), 
-    '_fsdp_wrapped_module._fpw_module.bert.embeddings.word_embeddings._fsdp_wrapped_module.flat_param': torch.Size([11720448]), 
-    '_fsdp_wrapped_module._fpw_module.bert.encoder._fsdp_wrapped_module.flat_param': torch.Size([42527232])
-  }
-  ```
 
+## Mapping between FSDP sharding strategies and DeepSpeed ZeRO Stages
+* `FULL_SHARD` maps to the DeepSpeed `ZeRO Stage-3`. Shards optimizer states, gradients and parameters.
+* `SHARD_GRAD_OP` maps to the DeepSpeed `ZeRO Stage-2`. Shards optimizer states and gradients.
+* `NO_SHARD` maps to `ZeRO Stage-0`. No sharding wherein each GPU has full copy of model, optimizer states and gradients.
+* `HYBRID_SHARD` maps to `ZeRO++ Stage-3` wherein `zero_hpz_partition_size=<num_gpus_per_node>`. Here, this will shard optimizer states, gradients and parameters within each node while each node has full copy.
 
-- In case of multiple models, it is necessary to prepare the models before creating optimizers or else it will throw an error. 
-Then pass the optimizers to the prepare call in the same order as corresponding models else `accelerator.save_state()` and `accelerator.load_state()` will result in wrong/unexpected behaviour.
-- This feature is incompatible with `--predict_with_generate` in the `run_translation.py` script of 🤗 `Transformers` library.
+## A few caveats to be aware of
+
+- In case of multiple models, pass the optimizers to the prepare call in the same order as corresponding models else `accelerator.save_state()` and `accelerator.load_state()` will result in wrong/unexpected behaviour.
+- This feature is incompatible with `--predict_with_generate` in the `run_translation.py` script of `Transformers` library.
 
 For more control, users can leverage the `FullyShardedDataParallelPlugin`. After creating an instance of this class, users can pass it to the Accelerator class instantiation.
 For more information on these options, please refer to the PyTorch [FullyShardedDataParallel](https://github.com/pytorch/pytorch/blob/0df2e863fbd5993a7b9e652910792bd21a516ff3/torch/distributed/fsdp/fully_sharded_data_parallel.py#L236) code.
+
+
+<Tip>
+
+    For those interested in the similarities and differences between FSDP and DeepSpeed, please check out the [concept guide here](../concept_guides/fsdp_and_deepspeed)!
+    
+</Tip>

docs/source/usage_guides/gradient_accumulation.md

@@ -13,7 +13,7 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Performing gradient accumulation with 🤗 Accelerate
+# Performing gradient accumulation with Accelerate
 
 Gradient accumulation is a technique where you can train on bigger batch sizes than 
 your machine would normally be able to fit into memory. This is done by accumulating gradients over
@@ -22,7 +22,7 @@ several batches, and only stepping the optimizer after a certain number of batch
 While technically standard gradient accumulation code would work fine in a distributed setup, it is not the most efficient
 method for doing so and you may experience considerable slowdowns!
 
-In this tutorial you will see how to quickly setup gradient accumulation and perform it with the utilities provided in 🤗 Accelerate,
+In this tutorial you will see how to quickly setup gradient accumulation and perform it with the utilities provided in Accelerate,
 which can total to adding just one new line of code!
 
 This example will use a very simplistic PyTorch training loop that performs gradient accumulation every two batches:
@@ -47,9 +47,9 @@ for index, batch in enumerate(training_dataloader):
         optimizer.zero_grad()
 ```
 
-## Converting it to 🤗 Accelerate
+## Converting it to Accelerate
 
-First the code shown earlier will be converted to utilize 🤗 Accelerate without the special gradient accumulation helper:
+First the code shown earlier will be converted to utilize Accelerate without the special gradient accumulation helper:
 
 ```diff
 + from accelerate import Accelerator
@@ -79,9 +79,9 @@ First the code shown earlier will be converted to utilize 🤗 Accelerate withou
 
 </Tip>
 
-## Letting 🤗 Accelerate handle gradient accumulation
+## Letting Accelerate handle gradient accumulation
 
-All that is left now is to let 🤗 Accelerate handle the gradient accumulation for us. To do so you should pass in a `gradient_accumulation_steps` parameter to [`Accelerator`], dictating the number 
+All that is left now is to let Accelerate handle the gradient accumulation for us. To do so you should pass in a `gradient_accumulation_steps` parameter to [`Accelerator`], dictating the number 
 of steps to perform before each call to `step()` and how to automatically adjust the loss during the call to [`~Accelerator.backward`]:
 
 ```diff
@@ -118,13 +118,29 @@ You can remove all the special checks for the step number and the loss adjustmen
 As you can see the [`Accelerator`] is able to keep track of the batch number you are on and it will automatically know whether to step through the prepared optimizer and how to adjust the loss. 
 
 <Tip>
+
 Typically with gradient accumulation, you would need to adjust the number of steps to reflect the change in total batches you are 
-training on. 🤗 Accelerate automagically does this for you by default. Behind the scenes we instantiate a GradientAccumulationPlugin configured to do this.
+training on. Accelerate automagically does this for you by default. Behind the scenes we instantiate a [`GradientAccumulationPlugin`] configured to do this.
+
+</Tip>
+
+<Tip warning={true}>
+
+The [`state.GradientState`] is sync'd with the active dataloader being iterated upon. As such it assumes naively that when we have reached the end of the dataloader everything will sync and a step will be performed. To disable this, set `sync_with_dataloader` to be `False` in the [`GradientAccumulationPlugin`]:
+
+```{python}
+from accelerate import Accelerator
+from accelerate.utils import GradientAccumulationPlugin
+
+plugin = GradientAccumulationPlugin(sync_with_dataloader=False)
+accelerator = Accelerator(..., gradient_accumulation_plugin=plugin)
+```
+
 </Tip>
 
 ## The finished code
 
-Below is the finished implementation for performing gradient accumulation with 🤗 Accelerate
+Below is the finished implementation for performing gradient accumulation with Accelerate
 
 ```python
 from accelerate import Accelerator
@@ -155,7 +171,7 @@ To learn more about what magic this wraps around, read the [Gradient Synchroniza
 
 ## Self-contained example
 
-Here is a self-contained example that you can run to see gradient accumulation in action with 🤗 Accelerate:
+Here is a self-contained example that you can run to see gradient accumulation in action with Accelerate:
 
 ```python
 import torch
@@ -171,38 +187,46 @@ set_seed(0)
 x = torch.tensor([1., 2., 3., 4., 5., 6., 7., 8.])
 y = torch.tensor([2., 4., 6., 8., 10., 12., 14., 16.])
 gradient_accumulation_steps = 4
-batch_size = len(x) // gradient_accumulation_steps
+per_device_batch_size = len(x) // gradient_accumulation_steps
 
 # define dataset and dataloader
 dataset = TensorDataset(x, y)
-dataloader = DataLoader(dataset, batch_size=batch_size)
+dataloader = DataLoader(dataset, batch_size=per_device_batch_size)
 
 # define model, optimizer and loss function
-model = torch.zeros((1, 1), requires_grad=True)
+class SimpleLinearModel(torch.nn.Module):
+    def __init__(self):
+        super(SimpleLinearModel, self).__init__()
+        self.weight = torch.nn.Parameter(torch.zeros((1, 1)))
+
+    def forward(self, inputs):
+        return inputs @ self.weight
+
+model = SimpleLinearModel()
 model_clone = copy.deepcopy(model)
 criterion = torch.nn.MSELoss()
-model_optimizer = torch.optim.SGD([model], lr=0.02)
+model_optimizer = torch.optim.SGD(model.parameters(), lr=0.02)
 accelerator = Accelerator(gradient_accumulation_steps=gradient_accumulation_steps)
 model, model_optimizer, dataloader = accelerator.prepare(model, model_optimizer, dataloader)
-model_clone_optimizer = torch.optim.SGD([model_clone], lr=0.02)
-print(f"initial model weight is {model.mean().item():.5f}")
-print(f"initial model weight is {model_clone.mean().item():.5f}")
+model_clone_optimizer = torch.optim.SGD(model_clone.parameters(), lr=0.02)
+print(f"initial model weight is {model.weight.mean().item():.5f}")
+print(f"initial model weight is {model_clone.weight.mean().item():.5f}")
 for i, (inputs, labels) in enumerate(dataloader):
     with accelerator.accumulate(model):
         inputs = inputs.view(-1, 1)
         print(i, inputs.flatten())
         labels = labels.view(-1, 1)
-        outputs = inputs @ model
+        outputs = model(inputs)
         loss = criterion(outputs, labels)
         accelerator.backward(loss)
         model_optimizer.step()
         model_optimizer.zero_grad()
-loss = criterion(x.view(-1, 1) @ model_clone, y.view(-1, 1))
+loss = criterion(x.view(-1, 1) @ model_clone.weight, y.view(-1, 1))
 model_clone_optimizer.zero_grad()
 loss.backward()
 model_clone_optimizer.step()
-print(f"w/ accumulation, the final model weight is {model.mean().item():.5f}")
-print(f"w/o accumulation, the final model weight is {model_clone.mean().item():.5f}")
+print(f"w/ accumulation, the final model weight is {model.weight.mean().item():.5f}")
+print(f"w/o accumulation, the final model weight is {model_clone.weight.mean().item():.5f}")
 ```
 ```
 initial model weight is 0.00000
@@ -214,3 +238,233 @@ initial model weight is 0.00000
 w/ accumulation, the final model weight is 2.04000
 w/o accumulation, the final model weight is 2.04000
 ```
+
+## Gradient accumulation on training samples of variable size
+
+As was pointed out in this [blog-post](https://huggingface.co/blog/gradient_accumulation), which points out a common error that occurs when performing gradient accumulation on training samples of variable size:
+
+>  [...] for gradient accumulation across token-level tasks like causal LM training, the correct loss should be computed by the **total loss across all batches in a gradient accumulation step** divided by the **total number of all non padding tokens in those batches**. This is not the same as the average of the per-batch loss values. 
+
+In other words, some adjustements must be made on losses that operate on a token-level basis.
+
+### Skeleton code
+
+```python
+from accelerate import Accelerator
+import math
+import contextlib
+
+gradient_accumulation_steps = 2
+accelerator = Accelerator(gradient_accumulation_steps=gradient_accumulation_steps)
+model, optimizer, training_dataloader, scheduler = accelerator.prepare(
+    model, optimizer, training_dataloader, scheduler
+)
+
+training_iterator = iter(training_dataloader)
+num_samples_in_epoch = len(training_dataloader)
+remainder = num_samples_in_epoch % gradient_accumulation_steps
+remainder = remainder if remainder != 0 else gradient_accumulation_steps
+total_updates = math.ceil(num_samples_in_epoch / gradient_accumulation_steps)
+        
+
+total_batched_samples = 0
+for update_step in range(total_updates):
+        # In order to correctly the total number of non-padded tokens on which we'll compute the cross-entropy loss
+        # we need to pre-load the full local batch - i.e the next per_device_batch_size * accumulation_steps samples
+        batch_samples = []
+        num_batches_in_step = gradient_accumulation_steps if update_step != (total_updates - 1) else remainder
+        for _ in range(num_batches_in_step):
+            batch_samples += [next(training_iterator)]
+            
+        # get local num items in batch 
+        num_items_in_batch = sum([(batch["labels"].ne(-100)).sum() for batch in batch_samples])
+        # to compute it correctly in a multi-device DDP training, we need to gather the total number of items in the full batch.
+        num_items_in_batch = accelerator.gather(num_items_in_batch).sum().item()
+            
+        for i, batch in enumerate(batch_samples):
+            # if we perform gradient accumulation in a multi-devices set-up, we want to avoid unecessary communications when accumulating
+            # cf: https://muellerzr.github.io/blog/gradient_accumulation.html
+            if (i < len(batch_samples) - 1 and accelerator.num_processes > 1):
+                ctx = model.no_sync
+            else:
+                ctx = contextlib.nullcontext
+            
+            total_batched_samples += 1
+
+            with ctx():
+                inputs, targets = batch
+                outputs = model(inputs)
+                loss = loss_function(outputs, targets) # the loss function shoud sum over samples rather than averaging
+                
+                # We multiply by num_processes because the DDP calculates the average gradient across all devices whereas dividing by num_items_in_batch already takes into account all devices
+                # Same reason for gradient_accumulation_steps, but this times it's Accelerate that calculate the average gradient across the accumulated steps
+                loss = (loss * gradient_accumulation_steps * accelerator.num_processes) / num_items_in_batch
+                
+                accelerator.backward(loss)
+
+        # Sync gradients and perform optimization steps once every gradient_accumulation_steps
+        optimizer.step()
+        scheduler.step()
+        optimizer.zero_grad()
+```
+
+### Self-contained causal LM example
+
+```py
+import torch
+import copy
+from accelerate import Accelerator
+from accelerate.utils import set_seed
+from accelerate.logging import  get_logger
+from torch.utils.data import Dataset, DataLoader
+import math
+import contexlib
+
+# seed
+set_seed(0)
+logger = get_logger(__name__)
+
+class MyDataset(Dataset):
+    def __init__(self, num_samples):
+        super().__init__()
+        self.len = num_samples
+
+    def __getitem__(self, index):
+        input_ids = torch.arange(1, index+2, dtype=torch.float32)
+        labels = torch.remainder(input_ids, 2)
+        return {"input_ids": input_ids, "labels": labels}
+
+    def __len__(self):
+        return self.len
+    
+def collate_fn(features):
+    input_ids = torch.nn.utils.rnn.pad_sequence([f["input_ids"] for f in features], batch_first=True, padding_value=-100)
+    labels = torch.nn.utils.rnn.pad_sequence([f["labels"] for f in features], batch_first=True, padding_value=-100)
+    return {"input_ids": input_ids[..., None], "labels": labels[..., None]}
+
+# define toy inputs and labels
+gradient_accumulation_steps = 2
+per_device_batch_size = 4
+
+# define accelerator
+accelerator = Accelerator(gradient_accumulation_steps=gradient_accumulation_steps)
+
+# define dataset and dataloader
+# for this toy example, we'll compute gradient descent over one single global batch
+dataset = MyDataset(per_device_batch_size*gradient_accumulation_steps*accelerator.num_processes)
+dataloader = DataLoader(dataset, batch_size=per_device_batch_size, collate_fn=collate_fn)
+
+# define model, model_optimizer and loss function
+model = torch.nn.Linear(1, 2, bias=False)
+model_clone = copy.deepcopy(model)
+criterion = torch.nn.CrossEntropyLoss(reduction="sum") # must sum over samples rather than averaging
+model_optimizer = torch.optim.SGD(model.parameters(), lr=0.08)
+
+
+logger.warning(f"initial model weight is {model.weight.detach().cpu().squeeze()}")
+logger.warning(f"initial model clone weight is {model_clone.weight.detach().cpu().squeeze()}")
+
+# prepare artifacts - accelerator handles device placement and dataloader splitting
+model, model_optimizer = accelerator.prepare(model, model_optimizer)
+dataloader = accelerator.prepare_data_loader(dataloader, device_placement=True)
+training_iterator = iter(dataloader)
+
+num_samples_in_epoch = len(dataloader)
+remainder = num_samples_in_epoch % gradient_accumulation_steps
+remainder = remainder if remainder != 0 else gradient_accumulation_steps
+total_gradient_updates = math.ceil(num_samples_in_epoch / gradient_accumulation_steps)
+
+total_batched_samples = 0
+for update_step in range(total_gradient_updates):
+        # In order to correctly the total number of non-padded tokens on which we'll compute the cross-entropy loss
+        # we need to pre-load the full local batch - i.e the next per_device_batch_size * accumulation_steps samples
+        batch_samples = []
+        num_batches_in_step = gradient_accumulation_steps if update_step != (total_gradient_updates - 1) else remainder
+        for _ in range(num_batches_in_step):
+            batch_samples += [next(training_iterator)]
+            
+        # get local num items in batch 
+        local_num_items_in_batch = sum([(batch["labels"].ne(-100)).sum() for batch in batch_samples])
+        logger.warning(f"Step {update_step} - Device {accelerator.process_index} - num items in the local batch {local_num_items_in_batch}", main_process_only=False)
+
+        # to compute it correctly in a multi-device DDP training, we need to gather the total number of items in the full batch.
+        num_items_in_batch = accelerator.gather(local_num_items_in_batch).sum().item()
+        logger.warning(f"Total num items {num_items_in_batch}")
+
+        for i, batch in enumerate(batch_samples):
+            inputs, labels = batch["input_ids"], batch["labels"]
+            total_batched_samples += 1
+            # if we perform gradient accumulation in a multi-devices set-up, we want to avoid unecessary communications when accumulating
+            # cf: https://muellerzr.github.io/blog/gradient_accumulation.html
+            if (i < len(batch_samples) - 1 and accelerator.num_processes > 1):
+                ctx = model.no_sync
+            else:
+                ctx = contextlib.nullcontext
+            with ctx():
+
+                outputs = model(inputs)
+                loss = criterion(outputs.view(-1, 2), labels.view(-1).to(torch.int64))
+                
+                # We multiply by num_processes because the DDP calculates the average gradient across all devices whereas dividing by num_items_in_batch already takes into account all devices
+                # Same reason for gradient_accumulation_steps, but this times it's Accelerate that calculate the average gradient across the accumulated steps 
+                loss = (loss * gradient_accumulation_steps * accelerator.num_processes) / num_items_in_batch
+                accelerator.backward(loss)
+        model_optimizer.step()
+        model_optimizer.zero_grad()
+                
+
+logger.warning(f"Device {accelerator.process_index} - w/ accumulation, the final model weight is {accelerator.unwrap_model(model).weight.detach().cpu().squeeze()}", main_process_only=False)
+
+# We know do the same operation but on a single device and without gradient accumulation
+
+if accelerator.is_main_process:
+    # prepare one single entire batch
+    dataloader = DataLoader(dataset, batch_size=len(dataset), collate_fn=collate_fn)
+    full_batch_without_accum = next(iter(dataloader))
+    total_inputs, total_labels = full_batch_without_accum["input_ids"], full_batch_without_accum["labels"]
+    model_clone_optimizer = torch.optim.SGD(model_clone.parameters(), lr=0.08)
+    
+    # train the cloned model
+    loss = torch.nn.CrossEntropyLoss(reduction="mean")(model_clone(total_inputs).view(-1, 2), total_labels.view(-1).to(torch.int64))
+    model_clone_optimizer.zero_grad()
+    loss.backward()
+    model_clone_optimizer.step()
+    
+    # We should have the same final weights.
+    logger.warning(f"w/o accumulation, the final model weight is {model_clone.weight.detach().cpu().squeeze()}")
+
+```
+
+Results on a single device - gradient accumulation steps set to 1 and batch_size set to 8:
+```
+initial model weight is tensor([-0.0075,  0.5364])
+initial model clone weight is tensor([-0.0075,  0.5364])
+Step 0 - Device 0 - num items in the local batch 36
+Total num items 36
+Device 0 - w/ accumulation, the final model weight is tensor([0.0953, 0.4337])
+w/o accumulation, the final model weight is tensor([0.0953, 0.4337])
+```
+
+Results on a two devices set-up - gradient accumulation steps set to 2 and batch_size set to 4.
+```
+initial model weight is tensor([-0.0075,  0.5364])
+initial model clone weight is tensor([-0.0075,  0.5364])
+Step 0 - Device 0 - num items in the local batch 52
+Step 0 - Device 1 - num items in the local batch 84
+Total num items 136
+Device 1 - w/ accumulation, the final model weight is tensor([0.2117, 0.3172])
+Device 0 - w/ accumulation, the final model weight is tensor([0.2117, 0.3172])
+w/o accumulation, the final model weight is tensor([0.2117, 0.3172])
+```
+
+### To go further:
+
+Please find a complete example script on a real world training run in the examples folder at the path [`accelerate/examples/by_feature/gradient_accumulation_for_autoregressive_models.py`](https://github.com/huggingface/accelerate/blob/main/examples/by_feature/gradient_accumulation_for_autoregressive_models.py).
+
+Running it on several training configurations with constant global batch size equal to 32 gives the following graph:
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/hf-audio/gradient_accumulation_example/resolve/main/training_losses.png">
+</div>
+
+Note that the training losses are exactly the same up to training step 20. The small deviation after this training step occurs at the very end of the first epoch, because, by [default](https://huggingface.co/docs/accelerate/en/package_reference/torch_wrappers#accelerate.data_loader.prepare_data_loader.even_batches), the dataloader duplicates the samples at the beginning of the dataset when the total batch size doesn't exactly divide the dataset.

docs/source/usage_guides/ipex.md

@@ -40,7 +40,7 @@ Check more approaches for [IPEX installation](https://intel.github.io/intel-exte
 
 ## How It Works For Training optimization in CPU
 
-🤗 Accelerate has integrated [IPEX](https://github.com/intel/intel-extension-for-pytorch), all you need to do is enabling it through the config.
+Accelerate has integrated [IPEX](https://github.com/intel/intel-extension-for-pytorch), all you need to do is enabling it through the config.
 
 **Scenario 1**: Acceleration of No distributed CPU training
 
@@ -115,8 +115,11 @@ What is the IP address of the machine that will host the main process? 36.112.23
 What is the port you will use to communicate with the main process? 29500
 Are all the machines on the same local network? Answer `no` if nodes are on the cloud and/or on different network hosts [YES/no]: yes
 Do you want to use Intel PyTorch Extension (IPEX) to speed up training on CPU? [yes/NO]:yes
+Do you want accelerate to launch mpirun? [yes/NO]: yes
+Please enter the path to the hostfile to use with mpirun [~/hostfile]: ~/hostfile
+Enter the number of oneCCL worker threads [1]: 1
 Do you wish to optimize your script with torch dynamo?[yes/NO]:NO
-How many CPU(s) should be used for distributed training? [1]:16
+How many processes should be used for distributed training? [1]:16
 -----------------------------------------------------------------------------------------------------------------------------------------------------------
 Do you wish to use FP16 or BF16 (mixed precision)?
 bf16
@@ -135,6 +138,9 @@ main_process_ip: 36.112.23.24
 main_process_port: 29500
 main_training_function: main
 mixed_precision: bf16
+mpirun_config:
+  mpirun_ccl: '1'
+  mpirun_hostfile: /home/user/hostfile
 num_machines: 4
 num_processes: 16
 rdzv_backend: static
@@ -148,6 +154,7 @@ use_cpu: true
 Set following env and using intel MPI to launch the training
 
 In node0, you need to create a configuration file which contains the IP addresses of each node (for example hostfile) and pass that configuration file path as an argument.
+If you selected to have Accelerate launch `mpirun`, ensure that the location of your hostfile matches the path in the config.
 ```bash
 $ cat hostfile
 xxx.xxx.xxx.xxx #node0 ip
@@ -155,7 +162,18 @@ xxx.xxx.xxx.xxx #node1 ip
 xxx.xxx.xxx.xxx #node2 ip
 xxx.xxx.xxx.xxx #node3 ip
 ```
-Now, run the following command in node0 and **16DDP** will be enabled in node0,node1,node2,node3 with BF16 mixed precision:
+When Accelerate is launching `mpirun`, source the oneCCL bindings setvars.sh to get your Intel MPI environment, and then
+run your script using `accelerate launch`. Note that the python script and environment needs to exist on all of the
+machines being used for multi-CPU training.
+```bash
+oneccl_bindings_for_pytorch_path=$(python -c "from oneccl_bindings_for_pytorch import cwd; print(cwd)")
+source $oneccl_bindings_for_pytorch_path/env/setvars.sh
+
+accelerate launch examples/nlp_example.py
+```
+Otherwise, if you selected not to have Accelerate launch `mpirun`, run the following command in node0 and **16DDP** will
+be enabled in node0,node1,node2,node3 with BF16 mixed precision. When using this method, the python script, python
+environment, and accelerate config file need to be present on all of the machines used for multi-CPU training.
 ```bash
 oneccl_bindings_for_pytorch_path=$(python -c "from oneccl_bindings_for_pytorch import cwd; print(cwd)")
 source $oneccl_bindings_for_pytorch_path/env/setvars.sh

docs/source/usage_guides/local_sgd.md

@@ -13,12 +13,12 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Using Local SGD with 🤗 Accelerate
+# Using Local SGD with Accelerate
 
 Local SGD is a technique for distributed training where gradients are not synchronized every step. Thus, each process updates its own version of the model weights and after a given number of steps these weights are synchronized by averaging across all processes. This improves communication efficiency and can lead to substantial training speed up especially when a computer lacks a faster interconnect such as NVLink.
 Unlike gradient accumulation (where improving communication efficiency requires increasing the effective batch size), Local SGD does not require changing a batch size or a learning rate / schedule. However, if necessary, Local SGD can be combined with gradient accumulation as well.
 
-In this tutorial you will see how to quickly setup  Local SGD 🤗 Accelerate. Compared to a standard Accelerate setup, this requires only two extra lines of code.
+In this tutorial you will see how to quickly setup  Local SGD Accelerate. Compared to a standard Accelerate setup, this requires only two extra lines of code.
 
 This example will use a very simplistic PyTorch training loop that performs gradient accumulation every two batches:
 
@@ -42,9 +42,9 @@ for index, batch in enumerate(training_dataloader):
         optimizer.zero_grad()
 ```
 
-## Converting it to 🤗 Accelerate
+## Converting it to Accelerate
 
-First the code shown earlier will be converted to use 🤗 Accelerate  with neither a LocalSGD or a gradient accumulation helper:
+First the code shown earlier will be converted to use Accelerate  with neither a LocalSGD or a gradient accumulation helper:
 
 ```diff
 + from accelerate import Accelerator
@@ -67,9 +67,9 @@ First the code shown earlier will be converted to use 🤗 Accelerate  with neit
           scheduler.step()
 ```
 
-## Letting 🤗 Accelerate handle model synchronization 
+## Letting Accelerate handle model synchronization 
 
-All that is left now is to let 🤗 Accelerate handle model parameter synchronization **and** the gradient accumulation for us. For simplicity let us assume we need to synchronize every 8 steps. This is
+All that is left now is to let Accelerate handle model parameter synchronization **and** the gradient accumulation for us. For simplicity let us assume we need to synchronize every 8 steps. This is
 achieved by adding one `with LocalSGD` statement and one call `local_sgd.step()` after every optimizer step:
 
 ```diff
@@ -88,7 +88,7 @@ achieved by adding one `with LocalSGD` statement and one call `local_sgd.step()`
 +           local_sgd.step()
 ```
 
-Under the hood, the Local SGD code **disables** automatic gradient synchornization (but accumulation still works as expected!). Instead it averages model parameters every `local_sgd_steps` steps (as well as in the end of the training loop).
+Under the hood, the Local SGD code **disables** automatic gradient synchronization (but accumulation still works as expected!). Instead it averages model parameters every `local_sgd_steps` steps (as well as at the end of the training loop).
 
 ## Limitations
 

docs/source/usage_guides/low_precision_training.md

@@ -0,0 +1,145 @@
+<!--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.
+-->
+
+# Low Precision Training Methods
+
+Accelerate provides integrations to train on lower precision methods using specified supported hardware through the `TransformersEngine` and `MS-AMP` packages. This documentation will help guide you through what hardware is supported, how to configure your [`Accelerator`] to leverage the low precision methods, and what you can expect when training. 
+
+## What training on FP8 means
+
+To explore more of the nitty-gritty in training in FP8 with PyTorch and Accelerate, check out the [concept_guide](../concept_guides/low_precision_training) on why this can be difficult. But essentially rather than training in BF16, some (or all) aspects of training a model can be performed using 8 bits instead of 16. The challenge is doing so without degrading final performance. 
+
+This is only enabled on specific NVIDIA hardware, namely:
+
+* Anything after the 3000 series consumer graphics cards (such as the 4090)
+* Hopper-based GPU architectures (such as the `H100` and `H200`)
+
+What this will result in is some gain in the memory used (as we've cut the needed memory in half for some parts of training) and an increase in throughput *should* be seen as well for larger models that can replace certain layers with FP8-enabled ones.
+
+## Configuring the Accelerator
+
+Currently two different backends for FP8 are supported (`TransformersEngine` and `MS-AMP`), each with different capabilities and configurations. 
+
+To use either, the same core API is used. Just pass `mixed_precision="fp8"` to either the [`Accelerator`], during `accelerate config` when prompted about mixed precision, or as part of your `config.yaml` file in the `mixed_precision` key:
+
+```{python}
+from accelerate import Accelerator
+accelerator = Accelerator(mixed_precision="fp8")
+```
+
+By default, if `MS-AMP` is available in your environment, Accelerate will automatically utilize it as a backend. To specify it yourself (and customize other parts of the FP8 mixed precision setup), you can utilize the [`utils.FP8RecipeKwargs`] or clarify it in your config `yaml`/during `accelerate launch`:
+
+```{python}
+from accelerate import Accelerator
+from accelerate.utils import FP8RecipeKwargs
+kwargs = [FP8RecipeKwargs(backend="msamp")]
+# Or to specify the backend as `TransformersEngine` even if MS-AMP is installed
+# kwargs = [FP8RecipeKwargs(backend="te")]
+accelerator = Accelerator(mixed_precision="fp8", kwarg_handlers=kwargs)
+```
+
+```{yaml}
+mixed_precision: fp8
+fp8_config:
+  amax_compute_algorithm: max
+  amax_history_length: 1024
+  backend: TE
+  fp8_format: HYBRID
+  interval: 1
+  margin: 0
+  override_linear_precision: false
+  use_autocast_during_eval: false
+```
+
+## Configuring MS-AMP
+
+Of the two, `MS-AMP` is traditionally the easier one to configure as there is only a single argument: the optimization level. 
+
+Currently two levels of optimization are supported in the Accelerate integration, `"O1"` and `"O2"` (using the letter 'o', not zero). 
+
+* `"O1"` will cast the weight gradients and `all_reduce` communications to happen in 8-bit, while the rest are done in 16 bit. This reduces the general GPU memory usage and speeds up communication bandwidths.
+* `"O2"` will also cast first-order optimizer states into 8 bit, while the second order states are in FP16. (Currently just the `Adam` optimizer is supported). This tries its best to minimize final accuracy degradation and will save the highest potential memory.
+
+To specify an optimization level, pass it to the `FP8KwargsHandler` by setting the `optimization_level` argument:
+
+```{python}
+from accelerate import Accelerator
+from accelerate.utils import FP8RecipeKwargs
+kwargs = [FP8RecipeKwargs(backend="msamp", optimization_level="O2")]
+accelerator = Accelerator(mixed_precision="fp8", kwarg_handlers=kwargs)
+```
+
+Or during `accelerate launch` via `--fp8_backend=msamp --fp8_opt_level=O2`
+
+Similarly this can be set in your `config.yaml`:
+
+```{yaml}
+mixed_precision: fp8
+fp8_config:
+    backend: MSAMP
+    opt_level: O2
+```
+
+## Configuring TransformersEngine
+
+TransformersEngine has much more available for customizing how and what FP8 calculations are performed. A full list of supported arguments and what they mean are available in [NVIDIA's documentation](https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/api/common.html), however they are restated as part of [`FP8KwargsHandler`]'s docstring for your convenience. 
+
+Accelerate tries to set sensible defaults, but exploring and tweaking the various parameters yourself can lead to better performance potentially.
+
+To use it, specify `backend="te"` and modify any of the arguments you want as part of your kwarg handler:
+
+```{python}
+from accelerate import Accelerator
+from accelerate.utils import FP8RecipeKwargs
+kwargs = [FP8RecipeKwargs(backend="te", ...)]
+accelerator = Accelerator(mixed_precision="fp8", kwarg_handlers=kwargs)
+```
+
+Or during `accelerate launch` via `--fp8_backend=te ...`. Use `accelerate launch --fp8_backend=te -h` to see relevent arguments.
+
+Similarly this can be set in your `config.yaml`:
+
+```{yaml}
+mixed_precision: fp8
+fp8_config:
+    amax_compute_algorithm: max
+    amax_history_length: 1024
+    backend: TE
+    fp8_format: HYBRID
+    interval: 1
+    margin: 0
+    override_linear_precision: false
+    use_autocast_during_eval: false
+```
+
+## Example Zoo
+
+We have examples showcasing training with FP8 both with accelerate and its underlying implementation available in the accelerate repo.
+Currently we support scripts showcasing:
+
+* Single GPU
+* Distributed Data Parallelism (Multi-GPU)
+* Fully Sharded Data Parallelism
+* DeepSpeed ZeRO 1 through 3
+
+Find out more [here](https://github.com/huggingface/accelerate/tree/main/benchmarks/fp8)
+
+## Further Reading
+
+To learn more about training in FP8 please check out the following resources:
+
+* [Our concept guide](../concept_guides/low_precision_training) detailing into more about both TransformersEngine and MS-AMP
+* [The `transformers-engine` documentation](https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/api/common.html)
+* [The `MS-AMP` documentation](https://azure.github.io/MS-AMP/docs/)

docs/source/usage_guides/megatron_lm.md

@@ -9,7 +9,7 @@ Unless required by applicable law or agreed to in writing, software distributed
 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
+⚠️ 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.
 -->
 
@@ -32,7 +32,7 @@ independently and in parallel by each shard followed by syncing across all GPUs
 In a simple transformer layer, this leads to 2 `all-reduces` in the forward path and 2 in the backward path.
 For more details, please refer research paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using
 Model Parallelism](https://arxiv.org/pdf/1909.08053.pdf) and 
-this section of 🤗 blogpost [The Technology Behind BLOOM Training](https://huggingface.co/blog/bloom-megatron-deepspeed#tensor-parallelism).
+this section of blogpost [The Technology Behind BLOOM Training](https://huggingface.co/blog/bloom-megatron-deepspeed#tensor-parallelism).
 
 
 b. **Pipeline Parallelism (PP)**: Reduces memory footprint and enables large scale training via inter-node parallelization. 
@@ -41,7 +41,7 @@ Layers are distributed uniformly across PP stages. For example, if a model has `
 pipeline parallelism, each GPU will have `6` layers (24/4). For more details on schedules to reduce the idle time of PP,
 please refer to the research paper [Efficient Large-Scale Language Model Training on GPU Clusters
 Using Megatron-LM](https://arxiv.org/pdf/2104.04473.pdf) and 
-this section of 🤗 blogpost [The Technology Behind BLOOM Training](https://huggingface.co/blog/bloom-megatron-deepspeed#pipeline-parallelism).
+this section of blogpost [The Technology Behind BLOOM Training](https://huggingface.co/blog/bloom-megatron-deepspeed#pipeline-parallelism).
 
 c. **Sequence Parallelism (SP)**: Reduces memory footprint without any additional communication. Only applicable when using TP.
 It reduces activation memory required as it prevents the same copies to be on the tensor parallel ranks 
@@ -57,7 +57,7 @@ d. **Data Parallelism (DP)** via Distributed Optimizer: Reduces the memory footp
 For example, when using Adam optimizer with mixed-precision training, each parameter accounts for 12 bytes of memory.
 This gets distributed equally across the GPUs, i.e., each parameter would account for 3 bytes (12/4) if we have 4 GPUs.
 For more details, please refer the research paper [ZeRO: Memory Optimizations Toward Training Trillion
-Parameter Models](https://arxiv.org/pdf/1910.02054.pdf) and following section of 🤗 blog 
+Parameter Models](https://arxiv.org/pdf/1910.02054.pdf) and following section of blog 
 [The Technology Behind BLOOM Training](https://huggingface.co/blog/bloom-megatron-deepspeed#zero-data-parallelism).
 
 e. **Selective Activation Recomputation**: Reduces the memory footprint of activations significantly via smart activation checkpointing.
@@ -72,9 +72,9 @@ PyTorch JIT compiled Fused GeLU and Fused Bias+Dropout+Residual addition.
 g. **Support for Indexed datasets**: Efficient binary format of datasets for large scale training. Support for the `mmap`, `cached` index file and the `lazy` loader format.
 
 h. **Checkpoint reshaping and interoperability**: Utility for reshaping Megatron-LM checkpoints of variable 
-tensor and pipeline parallel sizes to the beloved 🤗 Transformers sharded checkpoints as it has great support with plethora of tools
-such as 🤗 Accelerate Big Model Inference, Megatron-DeepSpeed Inference etc. 
-Support is also available for converting 🤗 Transformers sharded checkpoints to Megatron-LM checkpoint of variable tensor and pipeline parallel sizes
+tensor and pipeline parallel sizes to the beloved Transformers sharded checkpoints as it has great support with plethora of tools
+such as Accelerate Big Model Inference, Megatron-DeepSpeed Inference etc. 
+Support is also available for converting Transformers sharded checkpoints to Megatron-LM checkpoint of variable tensor and pipeline parallel sizes
 for large scale training.  
 
 
@@ -107,13 +107,16 @@ cd ..
 4. Installing Megatron-LM
 
 ```
-pip install git+https://github.com/huggingface/Megatron-LM.git
+git clone https://github.com/NVIDIA/Megatron-LM.git
+cd Megatron-LM
+git checkout core_r0.5.0
+pip install --no-use-pep517 -e .
 ```
 
 ## Accelerate Megatron-LM Plugin
 
 Important features are directly supported via the `accelerate config` command. 
-An example of thr corresponding questions for using Megatron-LM features is shown below:
+An example of the corresponding questions for using Megatron-LM features is shown below:
 
 ```bash
 :~$ accelerate config --config_file "megatron_gpt_config.yaml"
@@ -128,7 +131,7 @@ Do you want to enable Sequence Parallelism? [YES/no]:
 What is the Pipeline Parallelism degree/size? [1]:2
 What is the number of micro-batches? [1]:2
 Do you want to enable selective activation recomputation? [YES/no]: 
-Do you want to use distributed optimizer which shards optimizer state and gradients across data pralellel ranks? [YES/no]: 
+Do you want to use distributed optimizer which shards optimizer state and gradients across data parallel ranks? [YES/no]: 
 What is the gradient clipping value based on global L2 Norm (0 to disable)? [1.0]: 
 How many GPU(s) should be used for distributed training? [1]:4
 Do you wish to use FP16 or BF16 (mixed precision)? [NO/fp16/bf16]: bf16
@@ -355,8 +358,8 @@ def main():
 
 2. For using the Megatron-LM datasets, a few more changes are required. Dataloaders for these datasets
 are available only on rank 0 of each tensor parallel group. As such, there are rank where dataloader won't be
-avaiable and this requires tweaks to the training loop. Being able to do all this shows how
-felixble and extensible 🤗 Accelerate is. The changes required are as follows.
+available and this requires tweaks to the training loop. Being able to do all this shows how
+flexible and extensible Accelerate is. The changes required are as follows.
 
 a. For Megatron-LM indexed datasets, we need to use `MegatronLMDummyDataLoader` 
 and pass the required dataset args to it such as `data_path`, `seq_length` etc. 
@@ -388,7 +391,7 @@ c. Changes to training and evaluation loops as dataloader is only available on t
 So, we need to iterate only if the dataloader isn't `None` else provide empty dict
 As such, we loop using `while` loop and break when `completed_steps` is equal to `args.max_train_steps`
 This is similar to the Megatron-LM setup wherein user has to provide `max_train_steps` when using Megaton-LM indexed datasets.
-This displays how flexible and extensible 🤗 Accelerate is.
+This displays how flexible and extensible Accelerate is.
 
 ```python
 while completed_steps < args.max_train_steps:
@@ -411,10 +414,10 @@ while completed_steps < args.max_train_steps:
     
 ## Utility for Checkpoint reshaping and interoperability
 
-1. The scripts for these are present in 🤗 Transformers library under respective models. 
+1. The scripts for these are present in Transformers library under respective models. 
 Currently, it is available for GPT model [checkpoint_reshaping_and_interoperability.py](https://github.com/huggingface/transformers/blob/main/src/transformers/models/megatron_gpt2/checkpoint_reshaping_and_interoperability.py)
 
-2. Below is an example of conversion of checkpoint from Megatron-LM to universal 🤗 Transformers sharded checkpoint.
+2. Below is an example of conversion of checkpoint from Megatron-LM to universal Transformers sharded checkpoint.
 ```bash
 python checkpoint_reshaping_and_interoperability.py \
 --convert_checkpoint_from_megatron_to_transformers \
@@ -542,12 +545,12 @@ megatron_lm_plugin = MegatronLMPlugin(other_megatron_args=other_megatron_args)
 This covers Decoder only, Encode only and Encoder-Decoder model classes.
 
 2. Only loss is returned from model forward pass as 
-there is quite complex interplay of pipeline, tensor and data parallelsim behind the scenes.
+there is quite complex interplay of pipeline, tensor and data parallelism behind the scenes.
 The `model(**batch_data)` call return loss(es) averaged across the data parallel ranks.
 This is fine for most cases wherein pre-training jobs are run using Megatron-LM features and
 you can easily compute the `perplexity` using the loss. 
 For GPT model, returning logits in addition to loss(es) is supported. 
-These logits aren't gathered across data prallel ranks. Use `accelerator.utils.gather_across_data_parallel_groups`
+These logits aren't gathered across data parallel ranks. Use `accelerator.utils.gather_across_data_parallel_groups`
 to gather logits across data parallel ranks. These logits along with labels can be used for computing various 
 performance metrics. 
 
@@ -566,18 +569,18 @@ setting is synonymous with gradient accumulation.
 
 7. When using Megatron-LM, use `accelerator.save_state` and `accelerator.load_state` for saving and loading checkpoints.
 
-8. Below are the mapping from Megatron-LM model architectures to the the equivalent 🤗 transformers model architectures.
-Only these 🤗 transformers model architectures are supported.
+8. Below are the mapping from Megatron-LM model architectures to the the equivalent transformers model architectures.
+Only these transformers model architectures are supported.
 
 a. Megatron-LM [BertModel](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/bert_model.py) : 
-🤗 transformers models with `megatron-bert` in config's model type, e.g., 
+transformers models with `megatron-bert` in config's model type, e.g., 
 [MegatronBERT](https://huggingface.co/docs/transformers/model_doc/megatron-bert)
     
 b. Megatron-LM [GPTModel](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py) : 
-🤗 transformers models with `gpt2` in config's model type, e.g., 
+transformers models with `gpt2` in config's model type, e.g., 
 [OpenAI GPT2](https://huggingface.co/docs/transformers/model_doc/gpt2)
    
 c. Megatron-LM [T5Model](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/t5_model.py) : 
-🤗 transformers models with `t5` in  config's model type, e.g., 
+transformers models with `t5` in  config's model type, e.g., 
 [T5](https://huggingface.co/docs/transformers/model_doc/t5) and 
-[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)
+[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)

docs/source/usage_guides/memory.md

@@ -1,58 +0,0 @@
-<!--Copyright 2022 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.
--->
-
-# Memory Utilities
-
-One of the most frustrating errors when it comes to running training scripts is hitting "CUDA Out-of-Memory", 
-as the entire script needs to be restarted, progress is lost, and typically a developer would want to simply
-start their script and let it run.
-
-`Accelerate` provides a utility heavily based on [toma](https://github.com/BlackHC/toma) to give this capability.
-
-## find_executable_batch_size
-
-This algorithm operates with exponential decay, decreasing the batch size in half after each failed run on some 
-training script. To use it, restructure your training function to include an inner function that includes this wrapper, 
-and build your dataloaders inside it. At a minimum, this could look like 4 new lines of code. 
-> Note: The inner function *must* take in the batch size as the first parameter, but we do not pass one to it when called. The wrapper handles this for us
-
-It should also be noted that anything which will consume CUDA memory and passed to the `accelerator` **must** be declared inside the inner function,
-such as models and optimizers.
-
-```diff
-def training_function(args):
-    accelerator = Accelerator()
-
-+   @find_executable_batch_size(starting_batch_size=args.batch_size)
-+   def inner_training_loop(batch_size):
-+       nonlocal accelerator # Ensure they can be used in our context
-+       accelerator.free_memory() # Free all lingering references
-        model = get_model()
-        model.to(accelerator.device)
-        optimizer = get_optimizer()
-        train_dataloader, eval_dataloader = get_dataloaders(accelerator, batch_size)
-        lr_scheduler = get_scheduler(
-            optimizer, 
-            num_training_steps=len(train_dataloader)*num_epochs
-        )
-        model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
-            model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
-        )
-        train(model, optimizer, train_dataloader, lr_scheduler)
-        validate(model, eval_dataloader)
-+   inner_training_loop()
-```
-
-To find out more, check the documentation [here](../package_reference/utilities#accelerate.find_executable_batch_size).

docs/source/usage_guides/model_size_estimator.md

@@ -13,12 +13,12 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Understanding how big of a model can fit on your machine
+# Model memory estimator
 
 One very difficult aspect when exploring potential models to use on your machine is knowing just how big of a model will *fit* into memory with your current graphics card (such as loading the model onto CUDA).
 
-To help alleviate this, 🤗 Accelerate has a CLI interface through `accelerate estimate-memory`. This tutorial will 
-help walk you through using it, what to expect, and at the end link to the interactive demo hosted on the 🤗 Hub which will 
+To help alleviate this, Accelerate has a CLI interface through `accelerate estimate-memory`. This tutorial will 
+help walk you through using it, what to expect, and at the end link to the interactive demo hosted on the Hub which will 
 even let you post those results directly on the model repo!
 
 Currently we support searching for models that can be used in `timm` and `transformers`.
@@ -32,6 +32,27 @@ Currently we support searching for models that can be used in `timm` and `transf
 
 </Tip>
 
+## Gradio Demos
+
+Below are a few gradio demos related to what was described above. The first is the official Hugging Face memory estimation space, utilizing Accelerate directly:
+
+<div class="block dark:hidden">
+	<iframe 
+        src="https://hf-accelerate-model-memory-usage.hf.space?__theme=light"
+        width="850"
+        height="1600"
+    ></iframe>
+</div>
+<div class="hidden dark:block">
+    <iframe 
+        src="https://hf-accelerate-model-memory-usage.hf.space?__theme=dark"
+        width="850"
+        height="1600"
+    ></iframe>
+</div>
+
+A community member has taken the idea and expanded it further, allowing you to filter models directly and see if you can run a particular LLM given GPU constraints and LoRA configurations. To play with it, see [here](https://huggingface.co/spaces/Vokturz/can-it-run-llm) for more details.
+
 ## The Command
 
 When using `accelerate estimate-memory`, you need to pass in the name of the model you want to use, potentially the framework
@@ -114,8 +135,3 @@ This calculation is accurate within a few % of the actual value, so it is a very
 
 When performing inference you can expect to add up to an additional 20% as found by [EleutherAI](https://blog.eleuther.ai/transformer-math/). We'll be conducting research into finding a more accurate estimate to these values, and will update 
 this calculator once done.
-
-## Live Gradio Demo
-
-Lastly, we invite you to try the [live Gradio demo](https://huggingface.co/spaces/hf-accelerate/model-memory-usage) of this utility,
-which includes an option to post a discussion thread on a models repository with this data. Doing so will help provide access to these numbers in the community faster and help users know what you've learned!

docs/source/usage_guides/mps.md

@@ -44,11 +44,8 @@ accelerate launch /examples/cv_example.py --data_dir images
 
 ## A few caveats to be aware of
 
-1. We strongly recommend to install PyTorch >= 1.13 (nightly version at the time of writing) on your MacOS machine. 
-It has major fixes related to model correctness and performance improvements for transformer based models.
-Please refer to https://github.com/pytorch/pytorch/issues/82707 for more details.
-2. Distributed setups `gloo` and `nccl` are not working with `mps` device. 
+1. Distributed setups `gloo` and `nccl` are not working with `mps` device. 
 This means that currently only single GPU of `mps` device type can be used.
 
-Finally, please, remember that, 🤗 `Accelerate` only integrates MPS backend, therefore if you
+Finally, please, remember that, `Accelerate` only integrates MPS backend, therefore if you
 have any problems or questions with regards to MPS backend usage, please, file an issue with [PyTorch GitHub](https://github.com/pytorch/pytorch/issues).

docs/source/usage_guides/profiler.md

@@ -0,0 +1,337 @@
+<!--
+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 contains specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# Profiler
+
+Profiler is a tool that allows the collection of performance metrics during training and inference. Profiler’s context manager API can be used to better understand what model operators are the most expensive, examine their input shapes and stack traces, study device kernel activity, and visualize the execution trace. It provides insights into the performance of your model, allowing you to optimize and improve it.
+
+This guide explains how to use PyTorch Profiler to measure the time and memory consumption of the model’s operators and how to integrate this with Accelerate. We will cover various use cases and provide examples for each.
+
+## Using profiler to analyze execution time
+
+Profiler allows one to check which operators were called during the execution of a code range wrapped with a profiler context manager.
+
+Let’s see how we can use profiler to analyze the execution time:
+
+<hfoptions id="cpu execution time">
+<hfoption id="PyTorch">
+
+```python
+import torch
+import torchvision.models as models
+from torch.profiler import profile, record_function, ProfilerActivity
+
+model = models.resnet18()
+inputs = torch.randn(5, 3, 224, 224)
+
+with profile(activities=[ProfilerActivity.CPU], record_shapes=True) as prof:
+    model(inputs)
+
+print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=10))
+```
+
+</hfoption>
+<hfoption id="Accelerate">
+
+```python
+from accelerate import Accelerator, ProfileKwargs
+import torch
+import torchvision.models as models
+
+model = models.resnet18()
+inputs = torch.randn(5, 3, 224, 224)
+
+profile_kwargs = ProfileKwargs(
+    activities=["cpu"],
+    record_shapes=True
+)
+
+accelerator = Accelerator(cpu=True, kwargs_handlers=[profile_kwargs])
+model = accelerator.prepare(model)
+
+with accelerator.profile() as prof:
+    with torch.no_grad():
+        model(inputs)
+
+print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=10))
+```
+
+</hfoption>
+</hfoptions>
+
+The resulting table output (omitting some columns):
+
+```
+---------------------------------  ------------  ------------  ------------  ------------  
+                             Name      Self CPU     CPU total  CPU time avg    # of Calls  
+---------------------------------  ------------  ------------  ------------  ------------  
+                     aten::conv2d     171.000us      52.260ms       2.613ms            20  
+                aten::convolution     227.000us      52.089ms       2.604ms            20  
+               aten::_convolution     270.000us      51.862ms       2.593ms            20  
+         aten::mkldnn_convolution      51.273ms      51.592ms       2.580ms            20  
+                 aten::batch_norm     118.000us       7.059ms     352.950us            20  
+     aten::_batch_norm_impl_index     315.000us       6.941ms     347.050us            20  
+          aten::native_batch_norm       6.305ms       6.599ms     329.950us            20  
+                 aten::max_pool2d      40.000us       4.008ms       4.008ms             1  
+    aten::max_pool2d_with_indices       3.968ms       3.968ms       3.968ms             1  
+                       aten::add_     780.000us     780.000us      27.857us            28  
+---------------------------------  ------------  ------------  ------------  ------------  
+Self CPU time total: 67.016ms
+```
+
+To get a finer granularity of results and include operator input shapes, pass `group_by_input_shape=True` (note: this requires running the profiler with `record_shapes=True`):
+
+```python
+print(prof.key_averages(group_by_input_shape=True).table(sort_by="cpu_time_total", row_limit=10))
+```
+
+## Using profiler to analyze memory consumption
+
+Profiler can also show the amount of memory (used by the model’s tensors) that was allocated (or released) during the execution of the model’s operators. To enable memory profiling functionality pass `profile_memory=True`.
+
+<hfoptions id="memory consumption">
+<hfoption id="PyTorch">
+
+```python
+model = models.resnet18()
+inputs = torch.randn(5, 3, 224, 224)
+
+with profile(activities=[ProfilerActivity.CPU],
+        profile_memory=True, record_shapes=True) as prof:
+    model(inputs)
+
+print(prof.key_averages().table(sort_by="self_cpu_memory_usage", row_limit=10))
+```
+
+</hfoption>
+<hfoption id="Accelerate">
+
+```python
+model = models.resnet18()
+inputs = torch.randn(5, 3, 224, 224)
+
+profile_kwargs = ProfileKwargs(
+    activities=["cpu"],
+    profile_memory=True,
+    record_shapes=True
+)
+
+accelerator = Accelerator(cpu=True, kwargs_handlers=[profile_kwargs])
+model = accelerator.prepare(model)
+
+with accelerator.profile() as prof:
+    model(inputs)
+
+print(prof.key_averages().table(sort_by="self_cpu_memory_usage", row_limit=10))
+```
+
+</hfoption>
+</hfoptions>
+
+The resulting table output (omitting some columns):
+
+```
+---------------------------------  ------------  ------------  ------------  
+                             Name       CPU Mem  Self CPU Mem    # of Calls  
+---------------------------------  ------------  ------------  ------------  
+                      aten::empty      94.85 Mb      94.85 Mb           205  
+    aten::max_pool2d_with_indices      11.48 Mb      11.48 Mb             1  
+                      aten::addmm      19.53 Kb      19.53 Kb             1  
+                       aten::mean      10.00 Kb      10.00 Kb             1  
+              aten::empty_strided         492 b         492 b             5  
+                        aten::cat         240 b         240 b             6  
+                        aten::abs         480 b         240 b             4  
+              aten::masked_select         120 b         112 b             1  
+                         aten::ne          61 b          53 b             3  
+                         aten::eq          30 b          30 b             1  
+---------------------------------  ------------  ------------  ------------  
+Self CPU time total: 69.332ms
+```
+
+
+## Exporting chrome trace
+
+You can examine the sequence of profiled operators and CUDA kernels in Chrome trace viewer (`chrome://tracing`):
+
+![profile_export](https://github.com/huggingface/accelerate/assets/100389977/5acb193f-6d11-4f7b-9873-c600c19e8172)
+
+<hfoptions id="exporting chrome trace">
+<hfoption id="PyTorch">
+
+```python
+model = models.resnet18().cuda()
+inputs = torch.randn(5, 3, 224, 224).cuda()
+
+with profile(activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA]) as prof:
+    model(inputs)
+
+prof.export_chrome_trace("trace.json")
+```
+
+</hfoption>
+<hfoption id="Accelerate">
+
+```python
+model = models.resnet18()
+inputs = torch.randn(5, 3, 224, 224).cuda()
+profile_kwargs = ProfileKwargs(
+    activities=["cpu", "cuda"],
+    output_trace_dir="trace"
+)
+
+accelerator = Accelerator(kwargs_handlers=[profile_kwargs])
+model = accelerator.prepare(model)
+
+with accelerator.profile() as prof:
+    model(inputs)
+
+# The trace will be saved to the specified directory
+```
+For other hardware accelerators, e.g. XPU, you can change `cuda` to `xpu` in the above example code.
+
+</hfoption>
+</hfoptions>
+
+## Using Profiler to Analyze Long-Running Jobs
+
+Profiler offers an additional API to handle long-running jobs (such as training loops). Tracing all of the execution can be slow and result in very large trace files. To avoid this, use optional arguments:
+
+- `schedule_option`: Scheduling options allow you to control when profiling is active. This is useful for long-running jobs to avoid collecting too much data. Available keys are `wait`, `warmup`, `active`, `repeat` and `skip_first`. The profiler will skip the first `skip_first` steps, then wait for `wait` steps, then do the warmup for the next `warmup` steps, then do the active recording for the next `active` steps and then repeat the cycle starting with `wait` steps. The optional number of cycles is specified with the `repeat` parameter, the zero value means that the cycles will continue until the profiling is finished.
+- `on_trace_ready`: specifies a function that takes a reference to the profiler as an input and is called by the profiler each time the new trace is ready.
+
+To illustrate how the API works, consider the following example:
+
+<hfoptions id="custom handler">
+<hfoption id="PyTorch">
+
+```python
+from torch.profiler import schedule
+
+my_schedule = schedule(
+    skip_first=1,
+    wait=5,
+    warmup=1,
+    active=3,
+    repeat=2
+)
+
+def trace_handler(p):
+    output = p.key_averages().table(sort_by="self_cuda_time_total", row_limit=10)
+    print(output)
+    p.export_chrome_trace("/tmp/trace_" + str(p.step_num) + ".json")
+
+with profile(
+    activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
+    schedule=my_schedule,
+    on_trace_ready=trace_handler
+) as p:
+    for idx in range(8):
+        model(inputs)
+        p.step()
+```
+
+</hfoption>
+<hfoption id="Accelerate">
+
+```python
+def trace_handler(p):
+    output = p.key_averages().table(sort_by="self_cuda_time_total", row_limit=10)
+    print(output)
+    p.export_chrome_trace("/tmp/trace_" + str(p.step_num) + ".json")
+
+profile_kwargs = ProfileKwargs(
+    activities=["cpu", "cuda"],
+    schedule_option={"wait": 5, "warmup": 1, "active": 3, "repeat": 2, "skip_first": 1},
+    on_trace_ready=trace_handler
+)
+
+accelerator = Accelerator(kwargs_handlers=[profile_kwargs])
+model = accelerator.prepare(model)
+
+with accelerator.profile() as prof:
+    for idx in range(8):
+        model(inputs)
+        prof.step()
+```
+
+</hfoption>
+</hfoptions>
+
+## FLOPS
+
+Use formula to estimate the FLOPs (floating point operations) of specific operators (matrix multiplication and 2D convolution).
+
+To measure floating-point operations (FLOPS):
+
+<hfoptions id="FLOPS">
+<hfoption id="PyTorch">
+
+```python
+with profile(
+    activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
+    with_flops=True
+) as prof:
+    model(inputs)
+
+print(prof.key_averages().table(sort_by="flops", row_limit=10))
+```
+
+</hfoption>
+<hfoption id="Accelerate">
+
+```python
+profile_kwargs = ProfileKwargs(
+    with_flops=True
+)
+accelerator = Accelerator(kwargs_handlers=[profile_kwargs])
+
+with accelerator.profile() as prof:
+    model(inputs)
+
+print(prof.key_averages().table(sort_by="flops", row_limit=10))
+```
+
+</hfoption>
+</hfoptions>
+
+The resulting table output (omitting some columns):
+
+```
+-------------------------------------------------------  ------------  ------------  ------------  
+                                                   Name      Self CPU     Self CUDA    Total FLOPs  
+-------------------------------------------------------  ------------  ------------  ------------  
+                                           aten::conv2d     197.000us       0.000us  18135613440.000  
+                                            aten::addmm     103.000us      17.000us     5120000.000  
+                                              aten::mul      29.000us       2.000us          30.000  
+                                      aten::convolution     409.000us       0.000us            --  
+                                     aten::_convolution     253.000us       0.000us            --  
+                                aten::cudnn_convolution       5.465ms       2.970ms            --  
+                                        cudaEventRecord     138.000us       0.000us            --  
+                                  cudaStreamIsCapturing      43.000us       0.000us            --  
+                                  cudaStreamGetPriority      40.000us       0.000us            --  
+                       cudaDeviceGetStreamPriorityRange      10.000us       0.000us            --  
+-------------------------------------------------------  ------------  ------------  ------------  
+Self CPU time total: 21.938ms
+Self CUDA time total: 4.165ms
+```
+
+
+
+## Conclusion and Further Information
+
+PyTorch Profiler is a powerful tool for analyzing the performance of your models. By integrating it with Accelerate, you can easily profile your models and gain insights into their performance, helping you to optimize and improve them.
+
+For more detailed information, refer to the [PyTorch Profiler documentation](https://pytorch.org/docs/stable/profiler.html).

docs/source/usage_guides/quantization.md

@@ -13,13 +13,13 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Quantization
+# Model quantization
 
 ## `bitsandbytes` Integration
 
-🤗 Accelerate brings `bitsandbytes` quantization to your model. You can now load any pytorch model in 8-bit or 4-bit with a few lines of code.
+Accelerate brings `bitsandbytes` quantization to your model. You can now load any pytorch model in 8-bit or 4-bit with a few lines of code.
 
-If you want to use 🤗 Transformers models with `bitsandbytes`, you should follow this [documentation](https://huggingface.co/docs/transformers/main_classes/quantization). 
+If you want to use Transformers models with `bitsandbytes`, you should follow this [documentation](https://huggingface.co/docs/transformers/main_classes/quantization). 
 
 To learn more about how the `bitsandbytes` quantization works, check out the blog posts on [8-bit quantization](https://huggingface.co/blog/hf-bitsandbytes-integration) and [4-bit quantization](https://huggingface.co/blog/4bit-transformers-bitsandbytes).
 
@@ -30,6 +30,8 @@ You will need to install the following requirements:
 ```bash
 pip install bitsandbytes
 ```
+For non-cuda devices, you can refer to the bitsandbytes installation guide [here](https://huggingface.co/docs/bitsandbytes/main/en/installation#multi-backend).
+
 - Install latest `accelerate` from source
 ```bash
 pip install git+https://github.com/huggingface/accelerate.git
@@ -127,10 +129,10 @@ device_map = {
 
 It is not possible to perform pure 8bit or 4bit training on these models. However, you can train these models by leveraging parameter efficient fine tuning methods (PEFT) and train for example adapters on top of them. Please have a look at [peft](https://github.com/huggingface/peft) library for more details.
 
-Currently, you can't add adapters on top of any quantized model. However, with the official support of adapters with 🤗 Transformers models, you can fine-tune quantized models. If you want to finetune a 🤗 Transformers model , follow this [documentation](https://huggingface.co/docs/transformers/main_classes/quantization) instead. Check out this [demo](https://colab.research.google.com/drive/1VoYNfYDKcKRQRor98Zbf2-9VQTtGJ24k?usp=sharing) on how to fine-tune a 4-bit 🤗 Transformers model. 
+Currently, you can't add adapters on top of any quantized model. However, with the official support of adapters with Transformers models, you can fine-tune quantized models. If you want to finetune a Transformers model , follow this [documentation](https://huggingface.co/docs/transformers/main_classes/quantization) instead. Check out this [demo](https://colab.research.google.com/drive/1VoYNfYDKcKRQRor98Zbf2-9VQTtGJ24k?usp=sharing) on how to fine-tune a 4-bit Transformers model. 
 
 Note that you don’t need to pass `device_map` when loading the model for training. It will automatically load your model on your GPU. Please note that `device_map=auto` should be used for inference only.
 
 ### Example demo - running GPT2 1.5b on a Google Colab
 
-Check out the Google Colab [demo](https://colab.research.google.com/drive/1T1pOgewAWVpR9gKpaEWw4orOrzPFb3yM?usp=sharing) for running quantized models on a GTP2 model. The GPT2-1.5B model checkpoint is in FP32 which uses 6GB of memory. After quantization, it uses 1.6GB with 8-bit modules and 1.2GB with 4-bit modules.
+Check out the Google Colab [demo](https://colab.research.google.com/drive/1T1pOgewAWVpR9gKpaEWw4orOrzPFb3yM?usp=sharing) for running quantized models on a GPT2 model. The GPT2-1.5B model checkpoint is in FP32 which uses 6GB of memory. After quantization, it uses 1.6GB with 8-bit modules and 1.2GB with 4-bit modules.

docs/source/usage_guides/sagemaker.md

@@ -23,17 +23,16 @@ make it easier than ever to train Hugging Face Transformer models in [Amazon Sag
 ### Setup & Installation
 
 
-Before you can run your 🤗 Accelerate scripts on Amazon SageMaker you need to sign up for an AWS account. If you do not
+Before you can run your Accelerate scripts on Amazon SageMaker you need to sign up for an AWS account. If you do not
 have an AWS account yet learn more [here](https://docs.aws.amazon.com/sagemaker/latest/dg/gs-set-up.html).
 
-After you have your AWS Account you need to install the `sagemaker` sdk for 🤗 Accelerate with:
+After you have your AWS Account you need to install the `sagemaker` sdk for Accelerate with:
 
 ```bash
 pip install "accelerate[sagemaker]" --upgrade
 ```
 
-🤗 Accelerate currently uses the 🤗 DLCs, with `transformers`, `datasets` and `tokenizers` pre-installed. 🤗
-Accelerate is not in the DLC yet (will soon be added!) so to use it within Amazon SageMaker you need to create a
+Accelerate currently uses the DLCs, with `transformers`, `datasets` and `tokenizers` pre-installed. Accelerate is not in the DLC yet (will soon be added!) so to use it within Amazon SageMaker you need to create a
 `requirements.txt` in the same directory where your training script is located and add it as dependency:
 
 ```
@@ -43,25 +42,25 @@ accelerate
 You should also add any other dependencies you have to this `requirements.txt`.
 
 
-### Configure 🤗 Accelerate
+### Configure Accelerate
 
 You can configure the launch configuration for Amazon SageMaker the same as you do for non SageMaker training jobs with
-the 🤗 Accelerate CLI:
+the Accelerate CLI:
 
 ```bash
 accelerate config
 # In which compute environment are you running? ([0] This machine, [1] AWS (Amazon SageMaker)): 1
 ```
 
-🤗 Accelerate will go through a questionnaire about your Amazon SageMaker setup and create a config file you can edit.
+Accelerate will go through a questionnaire about your Amazon SageMaker setup and create a config file you can edit.
 
 <Tip>
 
-    🤗 Accelerate is not saving any of your credentials.
+    Accelerate is not saving any of your credentials.
 
 </Tip>
 
-### Prepare a 🤗 Accelerate fine-tuning script
+### Prepare a Accelerate fine-tuning script
 
 The training script is very similar to a training script you might run outside of SageMaker, but to save your model
 after training you need to specify either `/opt/ml/model` or use `os.environ["SM_MODEL_DIR"]` as your save
@@ -82,7 +81,7 @@ directory. After training, artifacts in this directory are uploaded to S3:
 
 ### Launch Training
 
-You can launch your training with 🤗 Accelerate CLI with:
+You can launch your training with Accelerate CLI with:
 
 ```
 accelerate launch path_to_script.py --args_to_the_script
@@ -146,8 +145,8 @@ image_uri: null
 mixed_precision: fp16
 num_machines: 1
 profile: xxxxx
-py_version: py38
-pytorch_version: 1.10.2
+py_version: py10
+pytorch_version: 2.5.0
 region: us-east-1
 transformers_version: 4.17.0
 use_cpu: false
@@ -159,7 +158,7 @@ use_cpu: false
 
 ### Python packages and dependencies
 
-🤗 Accelerate currently uses the 🤗 DLCs, with `transformers`, `datasets` and `tokenizers` pre-installed. If you
+Accelerate currently uses the DLCs, with `transformers`, `datasets` and `tokenizers` pre-installed. If you
 want to use different/other Python packages you can do this by adding them to the `requirements.txt`. These packages
 will be installed before your training script is started.
 
@@ -198,7 +197,7 @@ additional_args:
   max_wait: 86400
 ```
 
-*Note: Spot Instances are subject to be terminated and training to be continued from a checkpoint. This is not handled in 🤗 Accelerate out of the box. Contact us if you would like this feature.*
+*Note: Spot Instances are subject to be terminated and training to be continued from a checkpoint. This is not handled in Accelerate out of the box. Contact us if you would like this feature.*
 
 ### Remote scripts: Use scripts located on Github
 

docs/source/usage_guides/tracking.md

@@ -13,19 +13,22 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Tracking
+# Experiment trackers
 
-There are a large number of experiment tracking API's available, however getting them all to work with in a multi-processing environment can oftentimes be complex.
-🤗 Accelerate provides a general tracking API that can be used to log useful items during your script through [`Accelerator.log`]
+There are a large number of experiment tracking APIs available, however getting them all to work in a multi-processing environment can oftentimes be complex.
+Accelerate provides a general tracking API that can be used to log useful items during your script through [`Accelerator.log`]
 
 ## Integrated Trackers
 
-Currently `Accelerate` supports four trackers out-of-the-box:
+Currently `Accelerate` supports seven trackers out-of-the-box:
 
 - TensorBoard
 - WandB
 - CometML
+- Aim
 - MLFlow
+- ClearML
+- DVCLive
 
 To use any of them, pass in the selected type(s) to the `log_with` parameter in [`Accelerate`]:
 ```python
@@ -68,12 +71,12 @@ config = {
 
 accelerator.init_trackers("example_project", config=config)
 
-my_model, my_optimizer, my_training_dataloader = accelerate.prepare(my_model, my_optimizer, my_training_dataloader)
+my_model, my_optimizer, my_training_dataloader = accelerator.prepare(my_model, my_optimizer, my_training_dataloader)
 device = accelerator.device
 my_model.to(device)
 
-for iteration in config["num_iterations"]:
-    for step, batch in my_training_dataloader:
+for iteration in range(config["num_iterations"]):
+    for step, batch in enumerate(my_training_dataloader):
         my_optimizer.zero_grad()
         inputs, targets = batch
         inputs = inputs.to(device)
@@ -181,7 +184,7 @@ wandb_tracker = accelerator.get_tracker("wandb")
 From there you can interact with `wandb`'s `run` object like normal:
 
 ```python
-wandb_run.log_artifact(some_artifact_to_log)
+wandb_tracker.log_artifact(some_artifact_to_log)
 ```
 
 <Tip>
@@ -195,7 +198,7 @@ achieve the same outcome with:
 
 ```python
 wandb_tracker = accelerator.get_tracker("wandb", unwrap=True)
-with accelerator.on_main_process:
+if accelerator.is_main_process:
     wandb_tracker.log_artifact(some_artifact_to_log)
 ```
 
@@ -205,10 +208,10 @@ with accelerator.on_main_process:
 If a library has an API that does not follow a strict `.log` with an overall dictionary such as Neptune.AI, logging can be done manually under an `if accelerator.is_main_process` statement:
 ```diff
   from accelerate import Accelerator
-+ import neptune.new as neptune
++ import neptune
 
   accelerator = Accelerator()
-+ run = neptune.init(...)
++ run = neptune.init_run(...)
 
   my_model, my_optimizer, my_training_dataloader = accelerate.prepare(my_model, my_optimizer, my_training_dataloader)
   device = accelerator.device

docs/source/usage_guides/training_zoo.md

@@ -15,7 +15,7 @@ rendered properly in your Markdown viewer.
 
 # Example Zoo
 
-Below contains a non-exhuastive list of tutorials and scripts showcasing 🤗 Accelerate
+Below contains a non-exhaustive list of tutorials and scripts showcasing Accelerate.
 
 ## Official Accelerate Examples:
 
@@ -68,10 +68,15 @@ These examples showcase every feature in Accelerate at once that was shown in "F
 
 ## Integration Examples 
 
-These are tutorials from libraries that integrate with 🤗 Accelerate: 
+These are tutorials from libraries that integrate with Accelerate: 
 
 > Don't find your integration here? Make a PR to include it!
 
+### Amphion
+- [Training Text-to-Speech Models with Amphion](https://github.com/open-mmlab/Amphion/blob/main/egs/tts/README.md)
+- [Training Singing Voice Conversion Models with Amphion](https://github.com/open-mmlab/Amphion/blob/main/egs/svc/README.md)
+- [Training Vocoders with Amphion](https://github.com/open-mmlab/Amphion/blob/main/egs/vocoder/README.md)
+
 ### Catalyst
 
 - [Distributed training tutorial with Catalyst](https://catalyst-team.github.io/catalyst/tutorials/ddp.html)
@@ -80,7 +85,7 @@ These are tutorials from libraries that integrate with 🤗 Accelerate:
 
 - [Fine-tuning DALLE2](https://github.com/lucidrains/DALLE2-pytorch#usage)
 
-### 🤗 diffusers
+### Diffusers
 
 - [Performing textual inversion with diffusers](https://github.com/huggingface/diffusers/tree/main/examples/textual_inversion)
 - [Training DreamBooth with diffusers](https://github.com/huggingface/diffusers/tree/main/examples/dreambooth)
@@ -129,7 +134,7 @@ These are tutorials from libraries that integrate with 🤗 Accelerate:
 
 ## In Science
 
-Below contains a non-exhaustive list of papers utilizing 🤗 Accelerate. 
+Below contains a non-exhaustive list of papers utilizing Accelerate. 
 
 > Don't find your paper here? Make a PR to include it!
 
@@ -154,12 +159,12 @@ Below contains a non-exhaustive list of papers utilizing 🤗 Accelerate.
 * Puijin Cheng, Li Lin, Yijin Huang, Huaqing He, Wenhan Luo, Xiaoying Tang: “Learning Enhancement From Degradation: A Diffusion Model For Fundus Image Enhancement”, 2023; [arXiv:2303.04603](http://arxiv.org/abs/2303.04603).
 * Shun Shao, Yftah Ziser, Shay Cohen: “Erasure of Unaligned Attributes from Neural Representations”, 2023; [arXiv:2302.02997](http://arxiv.org/abs/2302.02997).
 * Seonghyeon Ye, Hyeonbin Hwang, Sohee Yang, Hyeongu Yun, Yireun Kim, Minjoon Seo: “In-Context Instruction Learning”, 2023; [arXiv:2302.14691](http://arxiv.org/abs/2302.14691).
-* Shikun Liu, Linxi Fan, Edward Johns, Zhiding Yu, Chaowei Xiao, Anima Anandkumar: “Prismer: A Vision-Language Model with An Ensemble of Experts”, 2023; [arXiv:2303.02506](http://arxiv.org/abs/2303.02506 ).
+* Shikun Liu, Linxi Fan, Edward Johns, Zhiding Yu, Chaowei Xiao, Anima Anandkumar: “Prismer: A Vision-Language Model with An Ensemble of Experts”, 2023; [arXiv:2303.02506](http://arxiv.org/abs/2303.02506).
 * Haoyu Chen, Zhihua Wang, Yang Yang, Qilin Sun, Kede Ma: “Learning a Deep Color Difference Metric for Photographic Images”, 2023; [arXiv:2303.14964](http://arxiv.org/abs/2303.14964).
 * Van-Hoang Le, Hongyu Zhang: “Log Parsing with Prompt-based Few-shot Learning”, 2023; [arXiv:2302.07435](http://arxiv.org/abs/2302.07435).
 * Keito Kudo, Yoichi Aoki, Tatsuki Kuribayashi, Ana Brassard, Masashi Yoshikawa, Keisuke Sakaguchi, Kentaro Inui: “Do Deep Neural Networks Capture Compositionality in Arithmetic Reasoning?”, 2023; [arXiv:2302.07866](http://arxiv.org/abs/2302.07866).
 * Ruoyao Wang, Peter Jansen, Marc-Alexandre Côté, Prithviraj Ammanabrolu: “Behavior Cloned Transformers are Neurosymbolic Reasoners”, 2022; [arXiv:2210.07382](http://arxiv.org/abs/2210.07382).
-* Martin Wessel, Tomáš Horych, Terry Ruas, Akiko Aizawa, Bela Gipp, Timo Spinde: “Introducing MBIB -- the first Media Bias Identification Benchmark Task and Dataset Collection”, 2023; [arXiv:2304.13148](http://arxiv.org/abs/2304.13148 ). DOI: [https://dx.doi.org/10.1145/3539618.3591882 10.1145/3539618.3591882].
+* Martin Wessel, Tomáš Horych, Terry Ruas, Akiko Aizawa, Bela Gipp, Timo Spinde: “Introducing MBIB -- the first Media Bias Identification Benchmark Task and Dataset Collection”, 2023; [arXiv:2304.13148](http://arxiv.org/abs/2304.13148). DOI: [https://dx.doi.org/10.1145/3539618.3591882 10.1145/3539618.3591882].
 * Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, Daniel Cohen-Or: “Attend-and-Excite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models”, 2023; [arXiv:2301.13826](http://arxiv.org/abs/2301.13826).
 * Marcio Fonseca, Yftah Ziser, Shay B. Cohen: “Factorizing Content and Budget Decisions in Abstractive Summarization of Long Documents”, 2022; [arXiv:2205.12486](http://arxiv.org/abs/2205.12486).
 * Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, Daniel Cohen-Or: “TEXTure: Text-Guided Texturing of 3D Shapes”, 2023; [arXiv:2302.01721](http://arxiv.org/abs/2302.01721).
@@ -172,4 +177,4 @@ Below contains a non-exhaustive list of papers utilizing 🤗 Accelerate.
 * Zhiruo Wang, Shuyan Zhou, Daniel Fried, Graham Neubig: “Execution-Based Evaluation for Open-Domain Code Generation”, 2022; [arXiv:2212.10481](http://arxiv.org/abs/2212.10481).
 * Minh-Long Luu, Zeyi Huang, Eric P. Xing, Yong Jae Lee, Haohan Wang: “Expeditious Saliency-guided Mix-up through Random Gradient Thresholding”, 2022; [arXiv:2212.04875](http://arxiv.org/abs/2212.04875).
 * Jun Hao Liew, Hanshu Yan, Daquan Zhou, Jiashi Feng: “MagicMix: Semantic Mixing with Diffusion Models”, 2022; [arXiv:2210.16056](http://arxiv.org/abs/2210.16056).
-* Yaqing Wang, Subhabrata Mukherjee, Xiaodong Liu, Jing Gao, Ahmed Hassan Awadallah, Jianfeng Gao: “LiST: Lite Prompted Self-training Makes Parameter-Efficient Few-shot Learners”, 2021; [arXiv:2110.06274](http://arxiv.org/abs/2110.06274).
+* Yaqing Wang, Subhabrata Mukherjee, Xiaodong Liu, Jing Gao, Ahmed Hassan Awadallah, Jianfeng Gao: “LiST: Lite Prompted Self-training Makes Parameter-Efficient Few-shot Learners”, 2021; [arXiv:2110.06274](http://arxiv.org/abs/2110.06274).

examples/README.md

@@ -28,6 +28,7 @@ pip install datasets evaluate transformers
 
 The same script can be run in any of the following configurations:
 - single CPU or single GPU
+- multi CPUs
 - multi GPUs (using PyTorch distributed mode)
 - (multi) TPUs
 - fp16 (mixed-precision) or fp32 (normal precision)
@@ -58,15 +59,27 @@ To run it in each of these various modes, use the following commands:
     * from any server with Accelerate launcher
         ```bash
         accelerate launch --mixed_precision fp16 ./nlp_example.py
+- multi CPUs (requires Open MPI, Intel MPI, or MVAPICH)
+    * With Accelerate config and launcher, execute the following from node 0:
+        ```bash
+        accelerate config  # Select to have accelerate launch mpirun
+        accelerate launch ./nlp_example.py  # This will run the script on each server
+        ```
+    * With Intel MPI:
+        ```bash
+        export CCL_WORKER_COUNT=1
+        export MASTER_ADDR=xxx.xxx.xxx.xxx #node0 ip
+        mpirun -f hostfile -n 16 -ppn 4 python ./nlp_example.py
+        ```
 - multi GPUs (using PyTorch distributed mode)
     * With Accelerate config and launcher
         ```bash
         accelerate config  # This will create a config file on your server
         accelerate launch ./nlp_example.py  # This will run the script on your server
         ```
-    * With traditional PyTorch launcher (`torch.distributed.launch` can be used with older versions of PyTorch)
+    * With traditional PyTorch launcher (`python -m torch.distributed.run` can be used instead of `torchrun`)
         ```bash
-        python -m torchrun --nproc_per_node 2 --use_env ./nlp_example.py
+        torchrun --nproc_per_node 2 ./nlp_example.py
         ```
 - multi GPUs, multi node (several machines, using PyTorch distributed mode)
     * With Accelerate config and launcher, on each machine:
@@ -74,18 +87,15 @@ To run it in each of these various modes, use the following commands:
         accelerate config  # This will create a config file on each server
         accelerate launch ./nlp_example.py  # This will run the script on each server
         ```
-    * With PyTorch launcher only (`torch.distributed.launch` can be used in older versions of PyTorch)
+    * With PyTorch launcher only (`python -m torch.distributed.run` can be used instead of `torchrun`). Run this command on each node:
         ```bash
-        python -m torchrun --nproc_per_node 2 \
-            --use_env \
-            --node_rank 0 \
-            --master_addr master_node_ip_address \
-            ./nlp_example.py  # On the first server
-        python -m torchrun --nproc_per_node 2 \
-            --use_env \
-            --node_rank 1 \
-            --master_addr master_node_ip_address \
-            ./nlp_example.py  # On the second server
+        torchrun \ # python -m torch.distributed.run 
+            --nproc_per_node 2 \
+            --nnodes 2 \
+            --rdzv_id 2299 \ # A unique job id 
+            --rdzv_backend c10d \
+            --rdzv_endpoint master_node_ip_address:29500 \
+            ./nlp_example.py
         ```
 - (multi) TPUs
     * With Accelerate config and launcher
@@ -103,6 +113,7 @@ The [cv_example.py](./cv_example.py) script is a simple example to fine-tune a R
 
 The same script can be run in any of the following configurations:
 - single CPU or single GPU
+- multi CPUs
 - multi GPUs (using PyTorch distributed mode)
 - (multi) TPUs
 - fp16 (mixed-precision) or fp32 (normal precision)
@@ -146,40 +157,49 @@ To run it in each of these various modes, use the following commands:
     * from any server with Accelerate launcher
         ```bash
         accelerate launch --mixed_precison fp16 ./cv_example.py --data_dir path_to_data
+- multi CPUs (requires Open MPI, Intel MPI, or MVAPICH)
+    * With Accelerate config and launcher, run the following from node 0:
+        ```bash
+        accelerate config --config_file config.yaml  # Select to have accelerate launch mpirun
+        accelerate launch ./cv_example.py --data_dir path_to_data # This will run the script on each server
+        ```
+    * With Intel MPI, execute mpirun from node 0:
+        ```bash
+        export CCL_WORKER_COUNT=1
+        export MASTER_ADDR=xxx.xxx.xxx.xxx #node0 ip
+        mpirun -f hostfile -n 16 -ppn 4 python ./cv_example.py --data_dir path_to_data
+        ```
 - multi GPUs (using PyTorch distributed mode)
     * With Accelerate config and launcher
         ```bash
-        accelerate config  # This will create a config file on your server
-        accelerate launch ./cv_example.py --data_dir path_to_data  # This will run the script on your server
+        accelerate config --config_file config.yaml  # This will create a config file on your server to `config.yaml`
+        accelerate launch --config_file config.yaml ./cv_example.py --data_dir path_to_data  # This will run the script on your server
         ```
-    * With traditional PyTorch launcher (`torch.distributed.launch` can be used with older versions of PyTorch)
+    * With traditional PyTorch launcher (`python -m torch.distributed.run` can be used instead of `torchrun`)
         ```bash
-        python -m torchrun --nproc_per_node 2 --use_env ./cv_example.py --data_dir path_to_data
+        torchrun --nproc_per_node 2 ./cv_example.py --data_dir path_to_data
         ```
 - multi GPUs, multi node (several machines, using PyTorch distributed mode)
     * With Accelerate config and launcher, on each machine:
         ```bash
-        accelerate config  # This will create a config file on each server
-        accelerate launch ./cv_example.py --data_dir path_to_data  # This will run the script on each server
+        accelerate config --config_file config.yaml  # This will create a config file on your server to `config.yaml`
+        accelerate launch --config_file config.yaml ./cv_example.py --data_dir path_to_data  # This will run the script on each server
         ```
-    * With PyTorch launcher only (`torch.distributed.launch` can be used with older versions of PyTorch)
+    * With PyTorch launcher only (`python -m torch.distributed.run` can be used instead of `torchrun`). Run this command on each node:
         ```bash
-        python -m torchrun --nproc_per_node 2 \
-            --use_env \
-            --node_rank 0 \
-            --master_addr master_node_ip_address \
-            ./cv_example.py --data_dir path_to_data  # On the first server
-        python -m torchrun --nproc_per_node 2 \
-            --use_env \
-            --node_rank 1 \
-            --master_addr master_node_ip_address \
-            ./cv_example.py --data_dir path_to_data  # On the second server
+        torchrun \ # python -m torch.distributed.run
+            --nproc_per_node 2 \
+            --nnodes 2 \
+            --rdzv_id 2299 \ # A unique job id 
+            --rdzv_backend c10d \
+            --rdzv_endpoint master_node_ip_address:29500 \
+            ./cv_example.py --data_dir path_to_data
         ```
 - (multi) TPUs
     * With Accelerate config and launcher
         ```bash
-        accelerate config  # This will create a config file on your TPU server
-        accelerate launch ./cv_example.py --data_dir path_to_data  # This will run the script on each server
+        accelerate config --config_file config.yaml  # This will create a config file on your server to `config.yaml`
+        accelerate launch --config_file config.yaml ./cv_example.py --data_dir path_to_data  # This will run the script on each server
         ```
     * In PyTorch:
         Add an `xmp.spawn` line in your script as you usually do.
@@ -188,11 +208,40 @@ To run it in each of these various modes, use the following commands:
 
 - [huggan project](https://github.com/huggingface/community-events/tree/main/huggan)
 
+
 ### Using AWS SageMaker integration
 - [Examples showcasing AWS SageMaker integration of 🤗 Accelerate.](https://github.com/pacman100/accelerate-aws-sagemaker)
 
+## Configuration zoo
+In [/config_yaml_templates](./config_yaml_templates/) we have a variety of *minimal* `config.yaml` templates and examples to help you learn
+how to create your own configuration files depending on the scenario. 
 
-## Simple Multi-GPU Hardware Launcher
+## SLURM Scripts 
+In [/slurm/submit_multigpu.sh](./slurm/submit_multigpu.sh) and [/slurm/submit_multinode.sh](./slurm/submit_multinode.sh) we present two scripts for running the examples on a machine with [SLURM](https://slurm.schedmd.com/documentation.html) workload manager. 
+
+In [/slurm/submit_multigpu.sh](./slurm/submit_multigpu.sh) the only parameter in the launcher that needs to be modified is `--num_processes`, which determines the number of GPUs we will use. In this case, using the environment variable `$SLURM_GPUS`, we indicate that we want to utilize all the GPUs available on the node we have requested. 
+
+In [/slurm/submit_multinode.sh](./slurm/submit_multinode.sh) we must specify the number of nodes that will be part of the training (`--num_machines`), how many GPUs we will use in total (`--num_processes`), the [`backend`](https://pytorch.org/docs/stable/elastic/run.html#note-on-rendezvous-backend), `--main_process_ip` which will be the address the master node and the `--main_process_port`.
+
+In [/slurm/submit_multicpu.sh](./slurm/submit_multicpu.sh) we must specify the number of nodes that will be part of the training (`--num_machines`), how many CPU processes we will use in total (`--num_processes`), the [`backend`](https://pytorch.org/docs/stable/elastic/run.html#note-on-rendezvous-backend), `--main_process_ip` which will be the address the master node and the `--main_process_port`. `mpirun_hostfile` specifies to run the job using MPIRun.
+
+In both scripts, we run `activateEnviroment.sh` at the beginning. This script should contain the necessary instructions to initialize the environment for execution. Below, we show an example that loads the necessary libraries ([Environment modules](https://github.com/cea-hpc/modules)), activates the Python environment, and sets up various environment variables, most of them to run the scripts in offline mode in case we don't have internet connection from the cluster.
+
+```bash
+# activateEnvironment.sh 
+module purge
+module load anaconda3/2020.02 cuda/10.2 cudnn/8.0.5 nccl/2.9.9 arrow/7.0.0 openmpi
+source activate /home/nct01/nct01328/pytorch_antoni_local
+
+export HF_HOME=/gpfs/projects/nct01/nct01328/
+export HF_LOCAL_HOME=/gpfs/projects/nct01/nct01328/HF_LOCAL
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+export PYTHONPATH=/home/nct01/nct01328/transformers-in-supercomputers:$PYTHONPATH 
+export GPUS_PER_NODE=4
+```
+
+## Simple Multi-GPU Hardware Launcher (using an external platform)
 
 [multigpu_remote_launcher.py](./multigpu_remote_launcher.py) is a minimal script that demonstrates launching accelerate
 on multiple remote GPUs, and with automatic hardware environment and dependency setup for reproducibility. You can

examples/by_feature/README.md

@@ -88,4 +88,34 @@ These arguments should be added at the end of any method for starting the python
 accelerate launch ./local_sgd.py --local_sgd_steps 4
 ```
 
+### DDP Communication Hook (`ddp_comm_hook.py`)
 
+- Shows how to use DDP Communication Hooks to control and optimize gradient communication across workers in a DistributedDataParallel setup.
+- Arguments available:
+  - `ddp_comm_hook`, the type of DDP communication hook to use. Choose between `no`, `fp16`, `bf16`, `power_sgd`, and `batched_power_sgd`.
+
+These arguments should be added at the end of any method for starting the python script (such as `accelerate launch`, `python -m torch.distributed.run`), such as:
+
+```bash
+accelerate launch ./ddp_comm_hook.py --mixed_precision fp16 --ddp_comm_hook power_sgd
+```
+
+### Profiler (`profiler.py`)
+
+- Shows how to use the profiling capabilities of `Accelerate` to profile PyTorch models during training.
+- Uses the `ProfileKwargs` handler to customize profiling options, including activities, scheduling, and additional profiling options.
+- Can generate and save profiling traces in JSON format for visualization in Chrome's tracing tool.
+
+Arguments available:
+- `--record_shapes`: If passed, records shapes for profiling.
+- `--profile_memory`: If passed, profiles memory usage.
+- `--with_stack`: If passed, profiles stack traces.
+- `--with_flops`: If passed, profiles floating point operations (FLOPS).
+- `--output_trace_dir`: If specified, saves the profiling trace to the given dir in JSON format.
+- `--cpu`: If passed, trains on the CPU instead of GPU.
+
+These arguments should be added at the end of any method for starting the Python script (such as `python`, `accelerate launch`, `python -m torchrun`), such as:
+
+```bash
+accelerate launch ./profiler.py --record_shapes --profile_memory --with_flops --output_trace_dir "profiler"
+```