refactor: nit change for get_parameters_from_modules (code debt) (#3815)

* refactor: nit change for get_parameters_from_modules

Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com>

* fix: quality check

Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com>

---------

Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.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: @muellerzr
-- DeepSpeed: @muellerzr
-- Command Line Interface: @muellerzr
-- Documentation: @muellerzr
-- Core parts of the library: @muellerzr @BenjaminBossan @SunMarc
-- Maintained examples: @muellerzr or @SunMarc
+- Fully-Sharded Data Parallism: @SunMarc @zach-huggingface
+- DeepSpeed: @SunMarc @zach-huggingface
+- Command Line Interface: @SunMarc @zach-huggingface
+- Documentation: @SunMarc @zach-huggingface
+- Core parts of the library: @BenjaminBossan @SunMarc @zach-huggingface 
+- Maintained examples: @SunMarc or  @zach-huggingface
 
  -->

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

@@ -15,7 +15,7 @@ jobs:
     outputs:
       version: ${{ steps.step1.outputs.version }}
     steps:
-      - uses: actions/checkout@v3.1.0
+      - uses: actions/checkout@v4
       - id: step1
         run: echo "version=$(python setup.py --version)" >> $GITHUB_OUTPUT
 

.github/workflows/build_and_run_tests.yml

@@ -16,13 +16,13 @@ jobs:
     outputs:
       changed: ${{ steps.was_changed.outputs.changed }}
     steps:
-      - uses: actions/checkout@v3.1.0
+      - uses: actions/checkout@v4
         with: 
           fetch-depth: "2"
       
       - name: Get changed files
         id: changed-files
-        uses: tj-actions/changed-files@v41
+        uses: tj-actions/changed-files@3f54ebb830831fc121d3263c1857cfbdc310cdb9 #v42
       
       - name: Was setup changed 
         id: was_changed
@@ -47,4 +47,4 @@ jobs:
   run-integration-tests:
     needs: build-docker-containers
     if: always()
-    uses: ./.github/workflows/self_hosted_integration_tests.yml
+    uses: ./.github/workflows/self_hosted_integration_tests.yml

.github/workflows/build_docker_images.yml

@@ -102,9 +102,15 @@ jobs:
         id: date
         run: |
           echo "date=$(date '+%Y-%m-%d')" >> $GITHUB_ENV
+          # Get the previous month
+          echo "base_year=$(date -d 'last month' '+%y')" >> $GITHUB_ENV
+          echo "base_month=$(date -d 'last month' '+%m')" >> $GITHUB_ENV
       - name: Build and Push GPU
         uses: docker/build-push-action@v4
         with:
           file: benchmarks/fp8/transformer_engine/Dockerfile
           push: true
-          tags: huggingface/accelerate:gpu-fp8-transformerengine-nightly-${{ env.date }}
+          tags: huggingface/accelerate:gpu-fp8-transformerengine-nightly-${{ env.date }}
+          build-args: |
+            BASE_YEAR=${{ env.base_year }}
+            BASE_MONTH=${{ env.base_month }}

.github/workflows/fp8_runner.yml

@@ -0,0 +1,37 @@
+name: Test FP8 Runner
+
+on:
+  workflow_dispatch:
+
+env:
+  GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+jobs:
+  set-prev-day:
+    runs-on: ubuntu-latest
+    outputs:
+      prev-day: ${{ steps.set-prev-day.outputs.prev-day }}
+    steps:
+      - name: Set PREV_DAY
+        id: set-prev-day
+        run: |
+          PREV_DAY=$(date -d "yesterday" '+%Y-%m-%d')
+          echo "prev-day=$PREV_DAY" >> $GITHUB_OUTPUT
+  run-fp8-tests:
+    needs: set-prev-day
+    runs-on:
+      group: aws-g6e-12xlarge
+    container:
+      image: huggingface/accelerate:gpu-fp8-transformerengine-nightly-${{ needs.set-prev-day.outputs.prev-day }}
+      options: --gpus all --shm-size "16gb"
+    steps:
+      - uses: actions/checkout@v3
+      - name: Install the library
+        run: |
+            pip install -e .[test_prod,test_fp8]
+      - name: Show installed libraries
+        run: |
+          pip freeze
+      - name: Run TE FP8 tests
+        run: |
+          python -m pytest -s -v ./tests/test_fp8.py
+

.github/workflows/gaudi3_scheduled.yml

@@ -1,23 +1,22 @@
-name: Gaudi1 tests (scheduled)
+name: Gaudi3 tests (scheduled)
 
 on:
   workflow_dispatch:
-  schedule:
-    - cron: "0 2 * * *"
+  schedule: # every day at 6 AM UTC
+    - cron: "0 6 * * *"
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
-  run_gaudi1_tests:
-    name: Test on Gaudi1
+  run-gaudi3-tests:
     runs-on:
-      group: aws-dl1-24xlarge
+      group: itac-bm-emr-gaudi3-dell-2gaudi
 
     container:
-      image: docker://vault.habana.ai/gaudi-docker/1.20.0/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
-      options: --runtime=habana --shm-size=64G --cap-add=sys_nice --env HABANA_VISIBLE_DEVICES=0,1
+      image: docker://vault.habana.ai/gaudi-docker/1.21.1/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
+      options: --runtime=habana --shm-size=64G --cap-add=sys_nice --env HABANA_VISIBLE_DEVICES
       env:
         OMPI_MCA_btl_vader_single_copy_mechanism: none
         PT_ENABLE_INT64_SUPPORT: 1
@@ -50,28 +49,39 @@ jobs:
         run: |
           pip install -e .[testing] \
             git+https://github.com/HabanaAI/DeepSpeed.git@1.20.0 \
-            git+https://github.com/huggingface/transformers.git@hpu-support
+            git+https://github.com/huggingface/transformers.git
 
       - name: Run CLI tests
+        if: ${{ !cancelled() && (success() || failure()) }}
         run: |
           make test_cli
 
       - name: Run Core tests
+        if: ${{ !cancelled() && (success() || failure()) }}
         run: |
           make test_core
 
       - name: Run Big Modeling tests
+        if: ${{ !cancelled() && (success() || failure()) }}
         run: |
           make test_big_modeling
 
-      - name: Run FSDP integration tests
-        run: |
-          make test_fsdp
-
       - name: Run DeepSpeed integration tests
+        if: ${{ !cancelled() && (success() || failure()) }}
         run: |
           make test_deepspeed
 
+      - name: Run FSDP integration tests
+        if: ${{ !cancelled() && (success() || failure()) }}
+        run: |
+          make test_fsdp
+
+      - name: Run TP integration tests
+        if: ${{ !cancelled() && (success() || failure()) }}
+        run: |
+          make test_tp
+
       - name: Run Examples tests
+        if: ${{ !cancelled() && (success() || failure()) }}
         run: |
           make test_examples

.github/workflows/integration_tests.yml

@@ -26,11 +26,11 @@ jobs:
     strategy:
       fail-fast: false
     steps:
-    - uses: actions/checkout@v3.1.0
-    - name: Set up python 3.9
-      uses: actions/setup-python@v3
+    - uses: actions/checkout@v4
+    - name: Set up python 3.10
+      uses: actions/setup-python@v5
       with:
-        python-version: 3.9
+        python-version: '3.10'
         cache: 'pip'
         cache-dependency-path: 'setup.py'
 

.github/workflows/pr_style_bot.yml

@@ -0,0 +1,19 @@
+# To run this bot, comment "@bot /style" on a PR
+name: Style Bot
+
+on:
+  issue_comment:
+    types: [created]
+
+permissions:
+  contents: write
+  pull-requests: write
+
+jobs:
+  style:
+    uses: huggingface/huggingface_hub/.github/workflows/style-bot-action.yml@main
+    with:
+      python_quality_dependencies: "[quality]"
+      style_command_type: "default"
+    secrets:
+      bot_token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/quality.yml

@@ -6,11 +6,11 @@ jobs:
   quality:
     runs-on: ubuntu-latest
     steps:
-    - uses: actions/checkout@v3.1.0
-    - name: Set up Python 3.9
-      uses: actions/setup-python@v3
+    - uses: actions/checkout@v4
+    - name: Set up Python 3.10
+      uses: actions/setup-python@v5
       with:
-        python-version: 3.9
+        python-version: '3.10'
         cache: 'pip'
         cache-dependency-path: 'setup.py'
     - name: Install Python dependencies

.github/workflows/self_hosted_integration_tests.yml

@@ -112,7 +112,7 @@ jobs:
           cd skorch;
           git config --global --add safe.directory '*'
           git checkout master && git pull
-          pip install .[testing]
+          pip install .[test]
           pip install flaky
 
       - name: Show installed libraries

.github/workflows/stale.yml

@@ -16,12 +16,12 @@ jobs:
     env:
       GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
     steps:
-    - uses: actions/checkout@v3.1.0
+    - uses: actions/checkout@v4
     
     - name: Setup Python
-      uses: actions/setup-python@v3
+      uses: actions/setup-python@v5
       with:
-        python-version: 3.9
+        python-version: '3.10'
         cache: 'pip'
         cache-dependency-path: 'setup.py'
     

.github/workflows/test.yml

@@ -38,11 +38,11 @@ jobs:
           test_rest
         ]
     steps:
-    - uses: actions/checkout@v3.1.0
-    - name: Set up python 3.9
-      uses: actions/setup-python@v3
+    - uses: actions/checkout@v4
+    - name: Set up python 3.10
+      uses: actions/setup-python@v5
       with:
-        python-version: 3.9
+        python-version: '3.10'
         cache: 'pip'
         cache-dependency-path: 'setup.py'
     
@@ -52,7 +52,7 @@ jobs:
         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.pytorch-version }} = minimum ]]; then pip install torchvision==0.18.1 torch==2.3.1; fi
-        pip install pytest-reportlog tabulate setuptools
+        pip install pytest-reportlog tabulate setuptools importlib_metadata
 
     - name: Show installed libraries
       run: |

.github/workflows/test_imports.yml

@@ -26,11 +26,11 @@ jobs:
           minimum,
         ]
     steps:
-    - uses: actions/checkout@v3.1.0
-    - name: Set up python 3.9
-      uses: actions/setup-python@v3
+    - uses: actions/checkout@v4
+    - name: Set up python 3.10
+      uses: actions/setup-python@v5
       with:
-        python-version: 3.9
+        python-version: '3.10'
         cache: 'pip'
         cache-dependency-path: 'setup.py'
     

Makefile

@@ -8,31 +8,35 @@ extra_quality_checks:
 	python utils/check_copies.py
 	python utils/check_dummies.py
 	python utils/check_repo.py
-	doc-builder style src/accelerate docs/source --max_len 119
 
 # this target runs checks on all files
 quality:
 	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:
 	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_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",)
-
 test_core:
-	python -m pytest -s -v ./tests/ --ignore=./tests/test_examples.py --ignore=./tests/deepspeed --ignore=./tests/test_big_modeling.py \
-	--ignore=./tests/fsdp --ignore=./tests/tp --ignore=./tests/test_cli.py $(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_core.log",)
+	python -m pytest -s -v ./tests/ \
+	--ignore=./tests/test_big_modeling.py \
+	--ignore=./tests/test_modeling_utils.py \
+	--ignore=./tests/test_examples.py \
+	--ignore=./tests/test_cli.py \
+	--ignore=./tests/deepspeed \
+	--ignore=./tests/fsdp \
+	--ignore=./tests/tp \
+	$(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_core.log",)
 
 test_cli:
 	python -m pytest -s -v ./tests/test_cli.py $(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_cli.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",)
+
 test_deepspeed:
 	python -m pytest -s -v ./tests/deepspeed $(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_deepspeed.log",)
 
@@ -57,7 +61,7 @@ test_examples:
 
 # Broken down example tests for the CI runners
 test_integrations:
-	python -m pytest -s -v ./tests/deepspeed ./tests/fsdp ./tests/tp $(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_integrations.log",)
+	python -m pytest -s -v ./tests/fsdp ./tests/tp ./tests/deepspeed $(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_integrations.log",)
 
 test_example_differences:
 	python -m pytest -s -v ./tests/test_examples.py::ExampleDifferenceTests $(if $(IS_GITHUB_CI),--report-log "$(PYTORCH_VERSION)_example_diff.log",)
@@ -83,7 +87,7 @@ prepare_release:
 # Make sure this is ran in a fresh venv of some form
 install_test_release:
 	pip uninstall accelerate -y
-	pip install -i https://testpypi.python.org/pypi --extra-index-url https://pypi.org/simple accelerate
+	pip install -i https://testpypi.python.org/pypi --extra-index-url https://pypi.org/simple accelerate$(if $(version),==$(version),)
 
 # Run as `make target=testpypi upload_release`
 upload_release:

benchmarks/big_model_inference/README.md

@@ -13,7 +13,7 @@ pip install transformers
 To reproduce or test a new setup, run
 
 ```py
-python inference_acc.py model_name
+python big_model_inference.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`.
@@ -43,4 +43,4 @@ Note on the results:
 
 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.
+- 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

@@ -18,6 +18,12 @@ import time
 import psutil
 import torch
 
+from accelerate.test_utils.testing import get_backend
+
+
+torch_device_type, _, _ = get_backend()
+torch_accelerator_module = getattr(torch, torch_device_type, torch.cuda)
+
 
 class PeakCPUMemory:
     def __init__(self):
@@ -54,16 +60,16 @@ def start_measure():
     measures = {"time": time.time()}
 
     gc.collect()
-    torch.cuda.empty_cache()
+    torch_accelerator_module.empty_cache()
 
     # CPU mem
     measures["cpu"] = psutil.Process().memory_info().rss
     cpu_peak_tracker.start()
 
     # GPU mem
-    for i in range(torch.cuda.device_count()):
-        measures[str(i)] = torch.cuda.memory_allocated(i)
-    torch.cuda.reset_peak_memory_stats()
+    for i in range(torch_accelerator_module.device_count()):
+        measures[str(i)] = torch_accelerator_module.memory_allocated(i)
+    torch_accelerator_module.reset_peak_memory_stats()
 
     return measures
 
@@ -73,16 +79,16 @@ def end_measure(start_measures):
     measures = {"time": time.time() - start_measures["time"]}
 
     gc.collect()
-    torch.cuda.empty_cache()
+    torch_accelerator_module.empty_cache()
 
     # CPU mem
     measures["cpu"] = (psutil.Process().memory_info().rss - start_measures["cpu"]) / 2**20
     measures["cpu-peak"] = (cpu_peak_tracker.stop() - start_measures["cpu"]) / 2**20
 
     # GPU mem
-    for i in range(torch.cuda.device_count()):
-        measures[str(i)] = (torch.cuda.memory_allocated(i) - start_measures[str(i)]) / 2**20
-        measures[f"{i}-peak"] = (torch.cuda.max_memory_allocated(i) - start_measures[str(i)]) / 2**20
+    for i in range(torch_accelerator_module.device_count()):
+        measures[str(i)] = (torch_accelerator_module.memory_allocated(i) - start_measures[str(i)]) / 2**20
+        measures[f"{i}-peak"] = (torch_accelerator_module.max_memory_allocated(i) - start_measures[str(i)]) / 2**20
 
     return measures
 
@@ -90,9 +96,9 @@ def end_measure(start_measures):
 def log_measures(measures, description):
     print(f"{description}:")
     print(f"- Time: {measures['time']:.2f}s")
-    for i in range(torch.cuda.device_count()):
-        print(f"- GPU {i} allocated: {measures[str(i)]:.2f}MiB")
+    for i in range(torch_accelerator_module.device_count()):
+        print(f"- {torch_device_type} {i} allocated: {measures[str(i)]:.2f}MiB")
         peak = measures[f"{i}-peak"]
-        print(f"- GPU {i} peak: {peak:.2f}MiB")
+        print(f"- {torch_device_type} {i} peak: {peak:.2f}MiB")
     print(f"- CPU RAM allocated: {measures['cpu']:.2f}MiB")
     print(f"- CPU RAM peak: {measures['cpu-peak']:.2f}MiB")

benchmarks/fp8/ms_amp/ddp.py

@@ -62,12 +62,12 @@ def train_baseline(opt_level="O2"):
 
     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"]}'
+    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
 
@@ -95,12 +95,12 @@ def train_integration(opt_level="O2"):
 
     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"]}'
+    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
 
@@ -109,15 +109,15 @@ 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"]}'
+        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

@@ -90,12 +90,12 @@ def train_baseline(zero_stage: int = 1, opt_level: str = "O1"):
     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"]}'
+    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
 
@@ -129,12 +129,12 @@ def train_integration(zero_stage: int = 1, opt_level: str = "O1"):
     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"]}'
+    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
@@ -145,17 +145,17 @@ if __name__ == "__main__":
         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"]}'
+            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/non_distributed.py

@@ -56,12 +56,12 @@ def train_baseline(opt_level="O2"):
 
     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"]}'
+    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
 
@@ -89,12 +89,12 @@ def train_integration(opt_level="O2"):
 
     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"]}'
+    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
 
@@ -104,15 +104,15 @@ if __name__ == "__main__":
         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"]}'
+        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/torchao/ddp.py

@@ -96,12 +96,12 @@ def train_baseline():
 
     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"]}'
+    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
 
@@ -128,12 +128,12 @@ def train_integration():
 
     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"]}'
+    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
 
@@ -142,17 +142,17 @@ 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"]}'
+    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/torchao/distrib_deepspeed.py

@@ -126,12 +126,12 @@ def train_baseline(zero_stage: int = 1):
 
     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"]}'
+    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']}"
+    )
 
     del config
     return base_model_results, trained_model_results, model_outputs, data
@@ -180,12 +180,12 @@ def train_integration(zero_stage: int = 1):
 
     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"]}'
+    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']}"
+    )
 
     del config
     return base_model_results, trained_model_results, model_outputs, data
@@ -197,17 +197,17 @@ if __name__ == "__main__":
         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"]}'
+        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']}"
+        )
         AcceleratorState()._reset_state(True)
     torch.distributed.destroy_process_group()

benchmarks/fp8/torchao/fsdp.py

@@ -106,12 +106,12 @@ def train_baseline():
 
     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"]}'
+    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
 
@@ -143,12 +143,12 @@ def train_integration():
 
     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"]}'
+    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
 
@@ -157,17 +157,17 @@ 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"]}'
+    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/torchao/non_distributed.py

@@ -87,12 +87,12 @@ def train_baseline():
 
     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"]}'
+    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
 
@@ -117,12 +117,12 @@ def train_integration():
 
     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"]}'
+    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
 
@@ -131,15 +131,15 @@ if __name__ == "__main__":
     baseline_not_trained, baseline_trained = train_baseline()
     AcceleratorState._reset_state(True)
     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"]}'
+    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

@@ -1,10 +1,13 @@
-FROM nvcr.io/nvidia/pytorch:24.07-py3
+ARG BASE_YEAR=25
+ARG BASE_MONTH=03
+
+FROM nvcr.io/nvidia/pytorch:${BASE_YEAR}.${BASE_MONTH}-py3
 
 RUN pip install transformers evaluate datasets
 RUN git clone https://github.com/huggingface/accelerate.git
 
 RUN cd accelerate && \
-    pip install -e . && \
+    pip install -e .[deepspeed] && \
     cd benchmarks/fp8
 
 RUN /bin/bash

benchmarks/fp8/transformer_engine/ddp.py

@@ -79,12 +79,12 @@ def train_baseline():
 
     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"]}'
+    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
 
@@ -114,12 +114,12 @@ def train_integration():
 
     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"]}'
+    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
 
@@ -128,17 +128,17 @@ 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"]}'
+    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

@@ -113,12 +113,12 @@ def train_baseline(zero_stage: int = 1):
 
     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"]}'
+    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
 
@@ -159,12 +159,12 @@ def train_integration(zero_stage: int = 1):
 
     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"]}'
+    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
 
@@ -175,17 +175,17 @@ if __name__ == "__main__":
         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"]}'
+        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/fsdp.py

@@ -91,12 +91,12 @@ def train_baseline():
 
     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"]}'
+    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
 
@@ -131,12 +131,12 @@ def train_integration():
 
     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"]}'
+    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
 
@@ -145,17 +145,17 @@ 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"]}'
+    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

@@ -70,12 +70,12 @@ def train_baseline():
 
     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"]}'
+    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
 
@@ -104,12 +104,12 @@ def train_integration():
 
     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"]}'
+    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
 
@@ -118,15 +118,15 @@ 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"]}'
+    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/fsdp2/README.md

@@ -0,0 +1,74 @@
+# FSDP2 Benchmarks
+
+This benchmark showcases `FSDP2` in 🤗 `accelerate` and compares it to `torch` baseline.
+
+## Overview
+
+This benchmark consists of two parts:
+- `main.py` is the main script that runs the benchmark
+- `visualize.py` is the script that visualizes the results (if `--output_dir` was specified for the previous command)
+
+## Motivation
+
+We want to showcase that 🤗 `accelerate`'s integration of `FSDP2` is on par raw PyTorch, and highlight a "broken" part in PyTorch that creating an optimizer before applying `FSDP2` **doesn't result in a working training loop**. (more on this later)
+This script showcases **matching memory usage and convergence between `accelerate` and `torch`'s baseline.**
+To deal with this breaking change (and maintain backward compatibility with FSDP1 in terms of an API), `accelerate` had to come up with a workaround since `accelerate` assumes that the user will nearly always create a model, optimizer, scheduler, etc beforehand and bring them themselves. This lead to an issue of a stark increase in memory as well as the model not even training if the user creates an optimizer beforehand. 
+To workaround this, we replace the parameters inside the optimizer with the newly created FSDP2 sharded ones. More about this can be found in this [blog post (TBD)](TODO)
+> [!WARNING]
+> This script is intended to fit on 2x 24GB GPUs, though on so few GPUs it's not possible to see the memory difference (discrepancies in grad allocation result in lower memory usage in the non-fixed case), only the difference in convergence. Below are attached results from 8x H100 GPUs where the difference is visible.
+> TLDR: more GPUs = bigger memory difference between fixed and non-fixed cases.
+
+## Results
+
+Here are the results from running the benchmark on 8x H100 GPUs:
+
+<p align="center">
+  <img src="imgs/allocated_memory.png" width="80%" alt="Allocated Memory Usage">
+</p>
+<p align="center">
+  <img src="imgs/reserved_memory.png" width="80%" alt="Reserved Memory Usage">
+</p>
+
+As you can see, the memory usage of `accelerate` and `torch_post_shard` (the **intended** way) are very similar, while `torch_pre_shard_not_fixed` uses significantly more memory. Our fix in `torch_pre_shard_fixed` brings the memory usage back in line with the **intended** approach.
+
+> [!WARNING]
+> Timing discrepancies are due to the benchmarks being ran in 1 script.
+
+
+## Running
+
+To run the benchmark, you can either use `accelerate launch` or `torchrun`:
+```bash
+accelerate launch main.py
+```
+```bash
+# For two GPUs
+torchrun --nproc_per_node 2 main.py
+```
+
+This supports multiple configurable options, you can learn about them by running:
+```bash
+python3 main.py --help
+```
+
+This script will run 4 different benchmarks:
+- `torch_optimizer_after_fsdp`: `torch` baseline where optimizer is created after applying `FSDP2`, this is the **intended** way to do it
+- `torch_optimizer_before_fsdp_not_fixed`: `torch` baseline where optimizer is created before applying `FSDP2` without fixing the optimizer parameters
+- `torch_optimizer_before_fsdp_fixed`: `torch` baseline where optimizer is created before applying `FSDP2` with our fix to the optimizer
+- `accelerate`: `accelerate`'s own integration of `FSDP2` where optimizer is created before applying `FSDP2`, but we apply our fix to the optimizer
+
+Memory results are saved in a folder specified by `--output_dir` argument.
+Optionally, you can specify `--save_memory_snapshot` to save the torch memory snapshot, which can then be viewed using [`torch memory viz`](https://pytorch.org/memory_viz)
+
+## Visualizing results
+
+To visualize the results, you can run:
+
+```bash
+python3 visualize.py --dir <path_to_output_dir>
+```
+
+This will then create two plots, showcasing allocated and reserved memory usage between all the different benchmarks discussed above.
+
+
+

benchmarks/fsdp2/main.py

@@ -0,0 +1,122 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import functools
+from typing import Callable
+
+import torch
+
+from accelerate import Accelerator
+from utils import parse_args, prepare_accelerate, prepare_torch
+
+
+MODEL_NAME = "Qwen/Qwen2.5-1.5B-Instruct"
+LEARNING_RATE = 3e-5
+
+CONFIG = {
+    "model_name": MODEL_NAME,
+    "learning_rate": LEARNING_RATE,
+}
+
+
+def train(
+    model: torch.nn.Module,
+    optimizer: torch.optim.Optimizer,
+    train_dataloader: torch.utils.data.DataLoader,
+    accelerator: Accelerator,
+) -> torch.Tensor:
+    losses = []
+    for batch in train_dataloader:
+        optimizer.zero_grad()
+        outputs = model(**batch, use_cache=False)
+
+        loss = outputs.loss
+        losses.append(loss.item())
+        accelerator.backward(loss)
+        optimizer.step()
+
+    return torch.tensor(losses)
+
+
+def evaluate(args, config: dict, init_fn: Callable, run_name: str) -> torch.Tensor:
+    model, optimizer, dataloader, accelerator, memory_tracker = init_fn(args, config)
+
+    loss = train(model, optimizer, dataloader, accelerator)
+
+    memory_tracker.stop()
+    msg = f"""Results for {run_name} (rank 0):
+Loss: {loss[-1].item()}
+Peak Allocated Memory: {float(memory_tracker.peak_allocated_memory):.2f} MB
+Peak Reserved Memory: {float(memory_tracker.peak_reserved_memory):.2f} MB
+{"-" * 34}"""
+    accelerator.print(msg)
+    return loss
+
+
+def main():
+    args = parse_args()
+    evaluations = [
+        functools.partial(
+            evaluate,
+            init_fn=functools.partial(prepare_torch, post_shard_optimizer=False, apply_optimizer_fix=True),
+            run_name="Optimizer Before FSDP (w/ fix)",
+        ),
+        functools.partial(
+            evaluate,
+            init_fn=functools.partial(prepare_torch, post_shard_optimizer=False, apply_optimizer_fix=False),
+            run_name="Optimizer Before FSDP (w/o fix)",
+        ),
+        functools.partial(
+            evaluate,
+            init_fn=functools.partial(prepare_torch, post_shard_optimizer=True),
+            run_name="Optimizer After FSDP",
+        ),
+        functools.partial(evaluate, init_fn=prepare_accelerate, run_name="Accelerate"),
+    ]
+    labels = [
+        "Optimizer Before FSDP (w/ fix)",
+        "Optimizer Before FSDP (w/o fix)",
+        "Optimizer After FSDP",
+        "Accelerate",
+    ]
+
+    results = {}
+    torch.use_deterministic_algorithms(True)
+
+    for evaluation, label in zip(evaluations, labels):
+        results[label] = evaluation(args, CONFIG)
+
+    torch.testing.assert_close(
+        results["Optimizer After FSDP"],
+        results["Optimizer Before FSDP (w/ fix)"],
+        msg="Optimizer After FSDP and Optimizer Before FSDP (w/ fix) should be the same",
+    )
+
+    torch.testing.assert_close(
+        results["Optimizer After FSDP"],
+        results["Accelerate"],
+        msg="Optimizer After FSDP and Accelerate should be the same",
+    )
+
+    torch.testing.assert_close(
+        results["Accelerate"],
+        results["Optimizer Before FSDP (w/ fix)"],
+        msg="Accelerate and Optimizer Before FSDP (w/ fix) should be the same",
+    )
+
+    torch.distributed.destroy_process_group()
+
+
+if __name__ == "__main__":
+    main()

benchmarks/fsdp2/measure_utils.py

@@ -0,0 +1,130 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import gc
+import json
+import os
+import threading
+import time
+
+import psutil
+import torch
+
+from accelerate import PartialState
+
+
+class MemoryTracker:
+    def __init__(
+        self,
+        device: torch.device,
+        output_directory: str,
+        run_name: str,
+        save_memory_snapshot: bool,
+        log_interval: float = 0.01,
+    ):
+        """Class for tracking gpu and cpu memory usage of the process.
+
+        Args:
+            device (`torch.device`):
+                PyTorch device to monitor.
+            output_directory (`str`):
+                Directory to save the memory usage data to, will be created if it doesn't exist.
+            run_name (`str`):
+                Name of the run, will be used to name the output files.
+            save_memory_snapshot (`bool`):
+                Whether to also save `torch.cuda.memory._dump_snapshot` to the output directory.
+            log_interval (`float`, *optional*):
+                Interval in seconds between memory measurements. Defaults to 0.01.
+        """
+        self.log_interval = log_interval
+        self.save_memory_snapshot = save_memory_snapshot
+        self.output_directory = output_directory
+        self.run_name = run_name
+
+        self.timestamps = []
+        self.allocated_memory = []
+        self.reserved_memory = []
+        self.virtual_memory = []
+
+        self.start_time = None
+        self.running = False
+
+        self._thread = None
+        self._state = PartialState()
+        self._process = psutil.Process()
+        self._device = device
+        self.torch_accelerator_module = getattr(torch, device.type, torch.cuda)
+
+    def _monitor(self):
+        self.start_time = time.time()
+
+        while self.running:
+            allocated = self.torch_accelerator_module.memory_allocated(self._device) / (1024 * 1024)
+            reserved = self.torch_accelerator_module.memory_reserved(self._device) / (1024 * 1024)
+            virtual_memory = self._process.memory_info().rss / (1024 * 1024)
+
+            self.allocated_memory.append(allocated)
+            self.reserved_memory.append(reserved)
+            self.virtual_memory.append(virtual_memory)
+            self.timestamps.append(time.time() - self.start_time)
+
+            time.sleep(self.log_interval)
+
+    def start(self):
+        gc.collect()
+        self.torch_accelerator_module.empty_cache()
+
+        if self.output_directory:
+            os.makedirs(self.output_directory, exist_ok=True)
+
+        if self.save_memory_snapshot:
+            self.torch_accelerator_module.memory._record_memory_history()
+
+        self.running = True
+        self._thread = threading.Thread(target=self._monitor)
+        self._thread.daemon = True
+        self._thread.start()
+
+    def stop(self):
+        self.running = False
+        if self._thread:
+            self._thread.join()
+
+        if self.save_memory_snapshot and self._state.is_main_process and self.output_directory:
+            output_file = os.path.join(self.output_directory, f"{self.run_name}_memory_snapshot.pkl")
+            self.torch_accelerator_module.memory._dump_snapshot(output_file)
+
+        if self._state.is_main_process and self.output_directory:
+            path = os.path.join(self.output_directory, f"{self.run_name}_memory_usage.json")
+            with open(path, "w") as f:
+                json.dump(
+                    {
+                        "timestamps": self.timestamps,
+                        "allocated_memory": self.allocated_memory,
+                        "reserved_memory": self.reserved_memory,
+                        "virtual_memory": self.virtual_memory,
+                    },
+                    f,
+                )
+        if self.save_memory_snapshot:
+            self.torch_accelerator_module.memory._record_memory_history(False)
+        self.torch_accelerator_module.empty_cache()
+
+    @property
+    def peak_allocated_memory(self):
+        return max(self.allocated_memory)
+
+    @property
+    def peak_reserved_memory(self):
+        return max(self.reserved_memory)

benchmarks/fsdp2/utils.py

@@ -0,0 +1,290 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+from types import MethodType
+from typing import Union
+
+import torch
+from datasets import load_dataset
+from measure_utils import MemoryTracker
+from torch.distributed.fsdp import MixedPrecisionPolicy, fully_shard
+from torch.optim import AdamW
+from torch.utils.data import DataLoader
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, DataCollatorForLanguageModeling
+from transformers.models.qwen2.modeling_qwen2 import Qwen2DecoderLayer
+
+from accelerate import Accelerator, FullyShardedDataParallelPlugin
+from accelerate.state import AcceleratorState, is_initialized
+from accelerate.utils import convert_outputs_to_fp32, set_seed
+
+
+SEED = 421
+
+
+def get_named_parameters(model: torch.nn.Module, drop_refs: bool = False) -> dict[str, Union[torch.Tensor, int]]:
+    """
+    This function returns a dictionary mapping the parameter names to their data pointers or
+    the original parameters if `drop_refs` is `False`.
+    It is used to get the original parameter names before `fully_shard` is applied.
+
+    We only return the data pointers, so we drop the references to the original parameters
+    and `fully_shard` will then trigger a new allocation for the sharded ones.
+
+    Args:
+        model (`torch.nn.Module`): Model instance to get the named parameters from
+        drop_refs (`bool`, *optional*, defaults to `False`): Whether to drop the references to the original parameters
+
+    Returns:
+        `dict[str, Union[torch.Tensor, int]]`: Dictionary mapping the parameter names to their data pointers or the original parameters if `drop_refs` is `False`
+    """
+    named_parameters = {}
+    for n, p in model.named_parameters():
+        # We only preserve the data pointers to have the unique 1:1 mapping between the original and the sharded parameters
+        named_parameters[n] = p.data_ptr() if drop_refs else p
+    return named_parameters
+
+
+def replace_optimizer_params(optimizer: torch.optim.Optimizer):
+    """
+    This function is called before using `fully_shard` on the model. It replaces the parameters of the optimizer with
+    empty tensors, so `fully_shard` can trigger a new allocation for the sharded ones. After this, we swap the parameters
+    `data_ptr` to the original one, so we can reuse that later to map the sharded parameters to the original ones.
+    This function modifies the optimizer in-place.
+
+    Args:
+        optimizer (torch.optim.Optimizer): Optimizer instance which contains the original model parameters
+    """
+
+    for param_group in optimizer.param_groups:
+        for i, p in enumerate(param_group["params"]):
+            # We drop a reference to the original param here, so that _move_states_to_device triggers a reallocation
+            # This is required or else the `fully_shard` -> `_move_states_to_device` uses the original memory address
+            # for the sharded parameters, and we get a weird/undefined behavior.
+            param_group["params"][i] = torch.empty_like(p)
+
+            # We save the original data_ptr, so we can swap back the parameters later
+            param_group["params"][i].data_ptr = p.data_ptr()
+
+
+def swap_back_optimizer_params(
+    model: torch.nn.Module, optimizer: torch.optim.Optimizer, old_named_parameter_pointers: dict[str, int]
+):
+    """
+    This function is the counterpart of `replace_optimizer_params`. It is called after `fully_shard` being applied to
+    the model. It swaps the parameters of the optimizer to their sharded counterparts.
+    It is done using the `data_ptr` mapping prepared in `replace_optimizer_params` and `get_named_parameters`.
+
+    Args:
+        model (`torch.nn.Module`): Model instance to get the new named parameters from
+        optimizer (`torch.optim.Optimizer`): Optimizer instance to swap the parameters of
+        old_named_parameter_pointers (`dict[str, int]`): Dictionary mapping the original parameter names: data_ptrs to the new ones
+    """
+    # We get the new named parameters after `fully_shard` being applied
+    # We don't drop the references as we need the sharded parameters now
+    new_named_parameters = get_named_parameters(model, drop_refs=False)
+
+    # We create a mapping from the original data_ptr to the new sharded param corresponding to it
+    mapping = {p: new_named_parameters[n] for n, p in old_named_parameter_pointers.items()}
+
+    for param_group in optimizer.param_groups:
+        # We swap the parameters of the optimizer to the new sharded ones
+        param_group["params"] = [mapping[p.data_ptr] for p in param_group["params"]]
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        help="Directory to save the benchmarking results.",
+    )
+    parser.add_argument(
+        "--save_memory_snapshot",
+        action="store_true",
+        default=False,
+        help="If True, `torch.cuda.memory._dump_snapshot` will be used to additionaly save the memory trace.",
+    )
+    ######################
+    # Training arguments #
+    ######################
+    parser.add_argument(
+        "--batch_size",
+        type=int,
+        default=2,
+        help="Batch size for the training loop.",
+    )
+    parser.add_argument(
+        "--block_size",
+        type=int,
+        default=128,
+        help="The maximum sequence length to use with the model.",
+    )
+    parser.add_argument(
+        "--dataset_fraction",
+        type=float,
+        default=1.0,
+        help="Fraction of the dataset to use.",
+    )
+    return parser.parse_args()
+
+
+def prepare_dataloader(tokenizer, args, accelerator: Accelerator) -> DataLoader:
+    dataset = load_dataset("tiny_shakespeare", split="train", trust_remote_code=True)
+
+    def tokenize_function(example):
+        return tokenizer(
+            example["text"],
+        )
+
+    dataset = dataset.map(
+        tokenize_function,
+        batched=True,
+        remove_columns=["text"],
+    )
+
+    block_size = min(tokenizer.model_max_length, args.block_size)
+
+    def group_texts(examples):
+        concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
+        total_length = len(concatenated_examples[list(examples.keys())[0]])
+
+        total_length = (total_length // block_size) * block_size
+
+        result = {
+            k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
+            for k, t in concatenated_examples.items()
+        }
+
+        result["labels"] = result["input_ids"].copy()
+        return result
+
+    dataset = dataset.map(group_texts, batched=True)
+    dataset = dataset.select(range(int(len(dataset) * args.dataset_fraction)))
+
+    def collate_fn(examples):
+        return DataCollatorForLanguageModeling(
+            tokenizer=tokenizer,
+            mlm=False,
+        )(examples)
+
+    dataloader = DataLoader(
+        dataset,
+        batch_size=args.batch_size,
+        collate_fn=collate_fn,
+    )
+    dataloader = accelerator.prepare(dataloader)
+    return dataloader
+
+
+def get_model(model_name: str):
+    # We reguire model to be loaded in fp32, otherwise benchmarks don't match as accelerate does upcasting of parameters to fp32
+    config = AutoConfig.from_pretrained(model_name, trust_remote_code=True, torch_dtype=torch.float32)
+    model = AutoModelForCausalLM.from_config(config)
+    return model
+
+
+def get_tokenizer(model_name: str):
+    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
+    tokenizer.pad_token = tokenizer.eos_token
+    return tokenizer
+
+
+def prepare_torch(
+    args, config: dict, post_shard_optimizer: bool = False, apply_optimizer_fix: bool = False
+) -> tuple[torch.nn.Module, torch.optim.Optimizer, torch.utils.data.DataLoader, Accelerator]:
+    mp_policy = MixedPrecisionPolicy(
+        param_dtype=torch.bfloat16,
+        reduce_dtype=torch.bfloat16,
+        output_dtype=torch.bfloat16,
+    )
+
+    accelerator = Accelerator(mixed_precision="bf16")
+    set_seed(SEED)
+    is_fixed = "fixed" if apply_optimizer_fix else "not_fixed"
+    is_post_shard = "optimizer_after_fsdp" if post_shard_optimizer else "optimizer_before_fsdp"
+    run_name = f"torch_{is_post_shard}" if post_shard_optimizer else f"torch_{is_post_shard}_{is_fixed}"
+
+    tokenizer = get_tokenizer(config["model_name"])
+    train_dataloader = prepare_dataloader(tokenizer, args, accelerator)
+
+    memory_tracker = MemoryTracker(accelerator.device, args.output_dir, run_name, args.save_memory_snapshot)
+    memory_tracker.start()
+
+    model = get_model(config["model_name"])
+    optimizer = None
+
+    if not post_shard_optimizer:
+        optimizer = AdamW(model.parameters(), lr=config["learning_rate"])
+
+        if apply_optimizer_fix:
+            # We drop the references to the original parameters, so that `fully_shard` can trigger a new allocation
+            # Then we get the `module_name: data_ptr` mapping, so we can swap back the parameters later
+            old_named_parameters = get_named_parameters(model, drop_refs=True)
+
+            # We replace the parameters of the optimizer with empty tensors, so that `fully_shard` can trigger a new allocation
+            # We also change the `data_ptr` of the parameters to the original ones, so we can swap back the parameters later
+            replace_optimizer_params(optimizer)
+
+    for module in model.modules():
+        if isinstance(module, Qwen2DecoderLayer):
+            fully_shard(module, mp_policy=mp_policy)
+    fully_shard(model, mp_policy=mp_policy)
+
+    # We do this to imitate how accelerate forces outputs to be in fp32 via `convert_outputs_to_fp32`
+    autocast_context = torch.autocast(device_type=accelerator.state.device.type, dtype=torch.bfloat16)
+    model_forward_func = model.forward.__func__
+    new_forward = autocast_context(model_forward_func)
+    model.forward = MethodType(new_forward, model)
+    model.forward = MethodType(convert_outputs_to_fp32(model.forward.__func__), model)
+
+    if post_shard_optimizer:
+        optimizer = AdamW(model.parameters(), lr=config["learning_rate"])
+
+    if not post_shard_optimizer and apply_optimizer_fix:
+        # We swap back the parameters of the optimizer to the original ones
+        swap_back_optimizer_params(model, optimizer, old_named_parameters)
+
+    return model, optimizer, train_dataloader, accelerator, memory_tracker
+
+
+def prepare_accelerate(
+    args, config: dict
+) -> tuple[torch.nn.Module, torch.optim.Optimizer, torch.utils.data.DataLoader, Accelerator]:
+    if is_initialized():
+        AcceleratorState()._reset_state(True)
+
+    fsdp_plugin = FullyShardedDataParallelPlugin(
+        fsdp_version=2,
+        auto_wrap_policy="transformer_based_wrap",
+        transformer_cls_names_to_wrap=["Qwen2DecoderLayer"],
+    )
+    accelerator = Accelerator(
+        fsdp_plugin=fsdp_plugin,
+        mixed_precision="bf16",
+    )
+    set_seed(SEED)
+
+    tokenizer = get_tokenizer(config["model_name"])
+    train_dataloader = prepare_dataloader(tokenizer, args, accelerator)
+
+    memory_tracker = MemoryTracker(accelerator.device, args.output_dir, "accelerate", args.save_memory_snapshot)
+    memory_tracker.start()
+
+    model = get_model(config["model_name"])
+    optimizer = AdamW(model.parameters(), lr=config["learning_rate"])
+
+    model, optimizer = accelerator.prepare(model, optimizer)
+
+    return model, optimizer, train_dataloader, accelerator, memory_tracker

benchmarks/fsdp2/visualize.py

@@ -0,0 +1,114 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import json
+
+import matplotlib.pyplot as plt
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--dir", type=str, help="Directory containing the memory usage data")
+    parser.add_argument(
+        "--memory_threshold",
+        type=int,
+        default=0,
+        help="Memory threshold to filter data that is below this value (only filters 1st `--filter_partition` of the points which should roughtly correspond to the model loading)",
+    )
+    parser.add_argument(
+        "--filter_partition",
+        type=float,
+        default=1 / 3,
+        help="Partition to drop data from that are below the memory threshold",
+    )
+    return parser.parse_args()
+
+
+def filter_data(data, memory_threshold, filter_partition, key):
+    timestamps = data["timestamps"]
+    memory = data[key]
+
+    mid_point = int(len(timestamps) * filter_partition)
+    filtered_times = []
+    filtered_memory = []
+    for i, (t, m) in enumerate(zip(timestamps, memory)):
+        if i < mid_point and m < memory_threshold:
+            continue
+        filtered_times.append(t)
+        filtered_memory.append(m)
+    return filtered_times, filtered_memory
+
+
+def compare_memory_usage(data, labels, memory_threshold, filter_partition):
+    plt.style.use("seaborn-v0_8")
+    colors = ["#2ecc71", "#e74c3c", "#3498db", "#f1c40f"]
+
+    fig1, ax1 = plt.subplots(figsize=(15, 5))
+    for data_item, label, color in zip(data, labels, colors):
+        timestamps, allocated = filter_data(data_item, memory_threshold, filter_partition, "allocated_memory")
+        ax1.plot(timestamps, allocated, label=label, color=color, linewidth=2)
+
+    ax1.set_xlabel("Time (s)", fontsize=12)
+    ax1.set_ylabel("Allocated Memory (GB)", fontsize=12)
+    ax1.set_title("Allocated Memory Usage Over Time", fontsize=14, pad=15)
+    ax1.grid(True, linestyle="--", alpha=0.7)
+    ax1.legend(frameon=True, fancybox=True, shadow=True, fontsize=10)
+    ax1.spines["top"].set_visible(False)
+    ax1.spines["right"].set_visible(False)
+    plt.tight_layout()
+
+    fig2, ax2 = plt.subplots(figsize=(15, 5))
+    for data_item, label, color in zip(data, labels, colors):
+        timestamps, reserved = filter_data(data_item, memory_threshold, filter_partition, "reserved_memory")
+        ax2.plot(timestamps, reserved, label=label, color=color, linewidth=2)
+
+    ax2.set_xlabel("Time (s)", fontsize=12)
+    ax2.set_ylabel("Reserved Memory (GB)", fontsize=12)
+    ax2.set_title("Reserved Memory Usage Over Time", fontsize=14, pad=15)
+    ax2.grid(True, linestyle="--", alpha=0.7)
+    ax2.legend(frameon=True, fancybox=True, shadow=True, fontsize=10)
+    ax2.spines["top"].set_visible(False)
+    ax2.spines["right"].set_visible(False)
+    plt.tight_layout()
+
+    return fig1, fig2
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    DIR = args.dir
+    with open(f"{DIR}/torch_optimizer_before_fsdp_not_fixed_memory_usage.json") as f:
+        optimizer_before_fsdp_not_fixed = json.load(f)
+
+    with open(f"{DIR}/torch_optimizer_after_fsdp_memory_usage.json") as f:
+        optimizer_after_fsdp = json.load(f)
+
+    with open(f"{DIR}/torch_optimizer_before_fsdp_fixed_memory_usage.json") as f:
+        optimizer_before_fsdp_fixed = json.load(f)
+
+    with open(f"{DIR}/accelerate_memory_usage.json") as f:
+        accelerate = json.load(f)
+
+    data = [optimizer_before_fsdp_not_fixed, optimizer_before_fsdp_fixed, optimizer_after_fsdp, accelerate]
+    labels = [
+        "Optimizer Before FSDP (w/o fix)",
+        "Optimizer Before FSDP (w/ fix)",
+        "Optimizer After FSDP",
+        "Accelerate",
+    ]
+
+    fig1, fig2 = compare_memory_usage(data, labels, args.memory_threshold, args.filter_partition)
+    fig1.savefig(f"{DIR}/allocated_memory.png")
+    fig2.savefig(f"{DIR}/reserved_memory.png")

benchmarks/torch.compile/README.md

@@ -0,0 +1,111 @@
+# Regional Compilation Benchmark
+
+This benchmark compares different compilation strategies using PyTorch's `torch.compile` and Accelerate's `compile_regions` utility, which is based on the recipe in [PyTorch documentation](https://pytorch.org/tutorials/recipes/regional_compilation.html).
+
+## Overview
+
+The benchmark evaluates three approaches:
+
+- **Baseline**: No compilation, standard PyTorch eager execution.
+- **Full compilation**: Using PyTorch's `torch.compile()` on the entire model.
+- **Regional compilation**: Using `accelerate.utils.compile_regions()` which targets specific blocks of the model to optimize compilation time.
+
+Each approach is tested with different batch sizes (1 and 4) and sequence lengths (128) on various LLaMA-based models ranging from 1B to 13B parameters. We purposefully run the forward pass outside of the `torch.no_grad()` context to simulate performance in a training environment, where gradients are needed.
+
+## Usage
+
+To run this benchmark:
+
+```bash
+python regional_compilation.py
+```
+
+The script will automatically download the model configurations, create models, and benchmark both compilation and inference times across different scenarios.
+
+## Requirements
+
+- Suitable GPU memory for the models being tested.
+- PyTorch with CUDA support.
+- Transformers library.
+- Accelerate library.
+
+## Results
+
+The benchmark results are summarized in the following figures:
+
+- Compilation time is how long it takes to run the first forward pass.
+- Speedup factor is the ratio of non-compiled baseline inference time to the fully/regionally compiled inference time.
+
+<p align="center">
+  <img src="imgs/compilation_time.png" width="80%" alt="Compilation Time">
+</p>
+<p align="center">
+  <img src="imgs/speedup_factor.png" width="80%" alt="Speedup Factor">
+</p>
+
+Full results are available in the tables below:
+
+```markdown
+[-------------------------------------------------- NousResearch/Llama-3.2-1B ---------------------------------------------------]
+                            |  Inference time (1x128)  |  Inference time (4x128)  |  Compile time (1x128)  |  Compile time (4x128)
+1 threads: -----------------------------------------------------------------------------------------------------------------------
+      Baseline              |           18.3           |           18.4           |                        |                      
+      Full compilation      |            6.3           |           10.0           |        10696.4         |        10248.0       
+      Regional compilation  |            9.7           |           10.0           |         1952.7         |         2903.9       
+
+Times are in milliseconds (ms).
+
+[---------------------------------------------- NousResearch/Hermes-3-Llama-3.2-3B ----------------------------------------------]
+                            |  Inference time (1x128)  |  Inference time (4x128)  |  Compile time (1x128)  |  Compile time (4x128)
+1 threads: -----------------------------------------------------------------------------------------------------------------------
+      Baseline              |           33.4           |           33.6           |                        |                      
+      Full compilation      |           11.2           |           23.9           |        17857.5         |        17736.5       
+      Regional compilation  |           17.3           |           23.7           |         2993.2         |         2478.8       
+
+Times are in milliseconds (ms).
+
+[---------------------------------------------- NousResearch/Hermes-3-Llama-3.1-8B ----------------------------------------------]
+                            |  Inference time (1x128)  |  Inference time (4x128)  |  Compile time (1x128)  |  Compile time (4x128)
+1 threads: -----------------------------------------------------------------------------------------------------------------------
+      Baseline              |           40.3           |           59.5           |                        |                      
+      Full compilation      |           18.9           |           54.4           |        20437.8         |        20152.3       
+      Regional compilation  |           19.7           |           54.0           |         2903.1         |         2438.0       
+
+Times are in milliseconds (ms).
+
+[--------------------------------------------- NousResearch/Nous-Hermes-Llama2-13b ----------------------------------------------]
+                            |  Inference time (1x128)  |  Inference time (4x128)  |  Compile time (1x128)  |  Compile time (4x128)
+1 threads: -----------------------------------------------------------------------------------------------------------------------
+      Baseline              |           45.5           |          100.4           |                        |                      
+      Full compilation      |           29.4           |           89.7           |        23099.4         |        22885.9       
+      Regional compilation  |           29.4           |           87.5           |         2945.5         |         2526.2       
+
+Times are in milliseconds (ms).
+```
+
+## Results Summary
+
+### Compilation Time
+
+Regional compilation provides significantly faster compilation times compared to full model compilation:
+
+- **Full compilation**: Takes ~10-23 seconds depending on model size.
+- **Regional compilation**: Takes only ~2-3 seconds across all model sizes.
+- **Speed improvement**: Regional compilation is **5-9x faster** to compile.
+
+### Inference Time
+
+Regional compilation delivers inference performance close to full compilation:
+
+- For batch size 1:
+  - For smaller models (1B-3B): Full compilation has a slight edge over regional compilation.
+  - For larger models (8B-13B): Regional compilation performs similarly to full compilation.
+- For batch size 4: Regional compilation performs similarly to full compilation across all models.
+
+## Key Takeaways
+
+1. **Comparable Performance**: Regional compilation delivers performance speedups similar to full compilation, especially for larger models.
+2. **Faster Compilation**: Regional compilation significantly reduces the time taken to compile models, making it a more efficient choice for deployment.
+3. **Batch Size Impact**: At batch size 4, full compilation and regional compilation perform nearly identically.
+4. **Model Size Impact**: Even with a small batch size, full compilation and regional compilation perform similarly for larger models (8B-13B).
+5. **Practical Application**: For real-world applications, regional compilation is a practical choice for optimizing training cold start times, especially when working with large models.

benchmarks/torch.compile/regional_compilation.py

@@ -0,0 +1,77 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from torch.utils.benchmark import Compare, Timer
+from transformers import AutoConfig, AutoModelForCausalLM
+
+from accelerate.test_utils.testing import get_backend
+from accelerate.utils import compile_regions
+
+
+torch.set_float32_matmul_precision("high")
+
+COMPILE_ITERS = 2
+INFERENCE_ITERS = 100
+
+BASELINE = "Baseline"
+COMPILE_TIME = "Compile time"
+INFRENCE_TIME = "Inference time"
+FULL_COMPILATION = "Full compilation"
+REGIONAL_COMPILATION = "Regional compilation"
+
+INFRENCE_STMT = "model(input_ids, use_cache=False)"
+COMPILE_STMT = f"torch._dynamo.reset(); torch._inductor.utils.clear_inductor_caches(); {INFRENCE_STMT}"
+
+torch_device_type, _, _ = get_backend()
+
+results = []
+for model_id in [
+    # non-gated llama models
+    "NousResearch/Llama-3.2-1B",
+    "NousResearch/Hermes-3-Llama-3.2-3B",
+    "NousResearch/Hermes-3-Llama-3.1-8B",
+    "NousResearch/Nous-Hermes-Llama2-13b",
+]:
+    with torch.device(torch_device_type):
+        config = AutoConfig.from_pretrained(model_id)
+        model = AutoModelForCausalLM.from_config(config).to(dtype=torch.float16).eval()
+
+    full_compilation_model = torch.compile(model)
+    regional_compilation_model = compile_regions(model)
+
+    for model, sub_label, description, stmt, iters in [
+        (model, BASELINE, INFRENCE_TIME, INFRENCE_STMT, INFERENCE_ITERS),
+        (full_compilation_model, FULL_COMPILATION, COMPILE_TIME, COMPILE_STMT, COMPILE_ITERS),
+        (full_compilation_model, FULL_COMPILATION, INFRENCE_TIME, INFRENCE_STMT, INFERENCE_ITERS),
+        (regional_compilation_model, REGIONAL_COMPILATION, COMPILE_TIME, COMPILE_STMT, COMPILE_ITERS),
+        (regional_compilation_model, REGIONAL_COMPILATION, INFRENCE_TIME, INFRENCE_STMT, INFERENCE_ITERS),
+    ]:
+        for batch_size, sequence_length in [(1, 128), (4, 128)]:
+            input_ids = torch.randint(
+                0, 1000, size=(batch_size, sequence_length), dtype=torch.int64, device=torch_device_type
+            )
+            results.append(
+                Timer(
+                    label=model_id,
+                    sub_label=sub_label,
+                    description=f"{description} ({batch_size}x{sequence_length})",
+                    globals={"model": model, "input_ids": input_ids},
+                    stmt=stmt,
+                ).timeit(number=iters)
+            )
+
+compare = Compare(results)
+compare.colorize()
+compare.print()

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:3.9-slim as compile-image
+FROM python:3.10-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:3.9-slim AS build-image
+FROM python:3.10-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

@@ -4,7 +4,6 @@
 # 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 && \
@@ -25,12 +24,12 @@ 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
+    --extra-index-url https://download.pytorch.org/whl/cu126
 
 RUN python3 -m pip install --no-cache-dir bitsandbytes
 
 # Stage 2
-FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu20.04 AS build-image
+FROM nvidia/cuda:12.6.3-cudnn-devel-ubuntu22.04 AS build-image
 COPY --from=compile-image /opt/conda /opt/conda
 ENV PATH /opt/conda/bin:$PATH
 

docker/accelerate-gpu/Dockerfile

@@ -4,7 +4,7 @@
 # Use base conda image to reduce time
 FROM continuumio/miniconda3:latest AS compile-image
 # Specify py version
-ENV PYTHON_VERSION=3.9
+ENV PYTHON_VERSION=3.10
 # Install apt libs
 RUN apt-get update && \
     apt-get install -y curl git wget && \
@@ -24,12 +24,12 @@ 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] \
-    --extra-index-url https://download.pytorch.org/whl/cu117
+    --extra-index-url https://download.pytorch.org/whl/cu126
 
 RUN python3 -m pip install --no-cache-dir bitsandbytes
 
 # Stage 2
-FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu20.04 AS build-image
+FROM nvidia/cuda:12.6.3-cudnn-devel-ubuntu22.04  AS build-image
 COPY --from=compile-image /opt/conda /opt/conda
 ENV PATH /opt/conda/bin:$PATH
 

docs/source/_toctree.yml

@@ -62,8 +62,12 @@
       title: Amazon SageMaker
     - local: usage_guides/mps
       title: Apple M1 GPUs
-    - local: usage_guides/ipex
-      title: IPEX training with CPU
+    - local: usage_guides/intel_cpu
+      title: Intel CPU
+    - local: usage_guides/gaudi
+      title: Intel Gaudi
+    - local: usage_guides/compilation
+      title: Compilation
     title: Training
   - isExpanded: true
     sections:
@@ -86,12 +90,16 @@
     title: Gradient synchronization
   - local: concept_guides/fsdp_and_deepspeed
     title: FSDP vs DeepSpeed
+  - local: concept_guides/fsdp1_vs_fsdp2
+    title: FSDP1 vs FSDP2
+  - local: concept_guides/context_parallelism
+    title: Context parallelism
   - local: concept_guides/low_precision_training
     title: Low precision training methods
   - local: concept_guides/training_tpu
     title: Training on TPUs
   title: Concepts and fundamentals
-- sections: 
+- sections:
   - local: package_reference/accelerator
     title: Accelerator
   - local: package_reference/state

docs/source/basic_tutorials/notebook.md

@@ -26,7 +26,7 @@ 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, an 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
@@ -52,7 +52,7 @@ os._exit(00)  # Restart the notebook
 
 ## Preparing the Dataset and Model
 
-Next you should prepare your dataset. As mentioned at earlier, great care should be taken when preparing the `DataLoaders` and model to make sure that **nothing** is put on *any* GPU. 
+Next you should prepare your dataset. As mentioned earlier, great care should be taken when preparing the `DataLoaders` and model to make sure that **nothing** is put on *any* GPU. 
 
 If you do, it is recommended to put that specific code into a function and call that from within the notebook launcher interface, which will be shown later. 
 

docs/source/basic_tutorials/troubleshooting.md

@@ -153,7 +153,7 @@ To use [`find_executable_batch_size`], restructure your training function to inc
 
 <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.
+The inner function **must** take batch size as the first parameter, but we do not pass one to it when called. The wrapper will handle 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>
 

docs/source/concept_guides/context_parallelism.md

@@ -0,0 +1,204 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ 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.
+-->
+
+# Context Parallel in 🤗`accelerate`
+
+This guide will cover basics of using context parallelism in 🤗`accelerate`, for the more curious readers, we will also cover some technicalities in the later sections.
+
+## Why context parallelism?
+
+With the advent of large language models, and recently reasoning models, the sequence length has been growing rapidly. This, combined with quadratic memory complexity of attention, has led to a need for more efficient ways to train models with long sequences.
+With sequence length of 128k, the memory requirement of the attention matrix is `128k * 128k * 2 bytes * num_heads = ~32 GB * num_heads` for `bf16` precision, given vanilla attention implementation. Granted, with usage of `flash attention` or `SDPA` which do not materialize these attention weights, this decreases drastically, but the growth in memory requirements is still considerable.
+
+Context parallelism allows us to shard the inputs to the attention computation along the sequence dimension and compute the attention in parallel on multiple GPUs. With this, we can train models with long sequences, scaling potentially to 1M+ sequence length.
+
+## How to use context parallelism?
+
+```diff
+from accelerate.utils import ParallelismConfig, TorchContextParallelConfig
+
++ cp_config = TorchContextParallelConfig(
++       cp_comm_strategy="alltoall", # no need to use cp_config at all, if you want to use the default "allgather"
++ )
+
++ parallelism_config = ParallelismConfig(
++     cp_size=8,
++     cp_handler=cp_config,  # or just cp_size=8, if you want to use the default "allgather"
++ )
+
+accelerator = Accelerator(
+    ...,
+    parallelism_config=parallelism_config,
+)
+```
+
+As with any other feature in 🤗`accelerate`, you can enable context parallelism also by passing the corresponding flags to `accelerate launch`.
+In this case, it's no different:
+
+```bash
+accelerate launch --parallelism-config-cp-size 8 --parallelism-config-cp-comm-strategy [allgather|alltoall] ...
+```
+
+> [!Tip]
+> You can also set the `cp_size` and `cp_comm_strategy` in the `accelerate config` command, which will save them in your `accelerate` configuration file, so you don't have to pass them every time you launch your script.
+
+> [!Tip]
+> Context parallelism is compatible with other parallelism strategies, such as data parallelism, tensor parallelism and FSDP2.
+> You can simply combine them by setting your parallelism sizes to the desired values, e.g. `--parallelism-config-dp-size 8 --parallelism-config-tp-size 2 --parallelism-config-cp-size 8`. Or you can use the `ParallelismConfig` class to set them programmatically.
+
+> [!Warning]
+> Context parallelism is tightly coupled  with `FSDP2`, which you can learn more about in the [FSDP2 introduction](fsdp1_vs_fsdp2.md). Meaning, context parallelism only works if you use `FullyShardedDataParallelPlugin` or `--use-fsdp` with version set to 2 to your
+> program. If no `FSDP2` is used, error will be raised.
+
+> [!Warning]
+> Context parallelism works only with [SDPA](https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html) and only with no mask or causal mask. We can't properly detect this for you, so it's your responsibility to ensure that you are using `SDPA` with no mask or causal mask. If you use any other attention implementation, it will raise an error.
+
+After enabling context parallelism with the methods mentioned above, you can then apply it to your training loop. We provide a thin wrapper around [`torch.distributed.tensor.experimental.context_parallel`](https://docs.pytorch.org/docs/stable/distributed.tensor.html#torch.distributed.tensor.experimental.context_parallel) that you can use in your training loop, that abstracts some of the complexity of using it (more on this later). To minimize the changes you have to do in your training loop, we provide a context manager that is a `noop` if context parallelism is not enabled, and applies the context parallelism if it is enabled. This way, you can use it in your training loop without changing any code based on your parallelism configuration.
+You can use it as follows:
+
+```python
+for batch in dataloader:
+    with accelerator.maybe_context_parallel(
+        buffers=[batch["input_ids"], batch["attention_mask"]],
+        buffer_seq_dims=[1, 1],
+        no_restore_buffers={batch["input_ids"], batch["labels"]},
+    ):
+        outputs = model(**batch)
+        ...
+```
+
+> [!Warning]
+> This context manager has to be recreated with each training step, as shown in the example above. It's crucial to do so.
+
+This can scale your context size to 1M+ sequence length potentially. Below, we showcase speed and memory usage of context parallelism for up-to 256k context size. We can see that when we double the context size and number of GPUs, we can achieve consistent memory usage, potentially enabling endless context length scaling.
+
+<p align="center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/examples/fsdp2/cp_perf.png" alt="context parallelism memory usage" />
+  <br>
+  <em>Figure 1: Memory usage and speed of context parallelism for up-to 256k context size.</em>
+</p>
+
+> [!Tip]
+> These examples were created with a script you can find [in the examples folder](https://github.com/huggingface/accelerate/blob/main/examples/fsdp2/nd_parallel.py). To run the example on 8 H100 GPUs (128k sequence length), you can use the following command:
+> ```bash
+> accelerate launch --use-fsdp --fsdp-activation-checkpointing=TRUE examples/fsdp2/nd_parallel.py --cp-size=8 --sequence-length=128000
+> ```
+
+
+## Accelerate's interface
+
+The context manager takes a few arguments, that are used to configure the context parallelism.
+
+- `buffers`: This is a list of tensors that are to be sharded across the sequence dimension. These tensors are usually input ids, labels and attention mask.
+- `buffer_seq_dims`: This is a list of integers, that specify the sequence dimension of the buffers, in the order of the `buffers` list. If you pass `buffers=[input_ids, shift_labels]` with both having shape `[batch_size, sequence_length]`, you would pass `buffer_seq_dims=[1, 1]`.
+                     as the sequence dimension is the second dimension of the tensors. This is required for correct computation of the model outputs.
+- `no_restore_buffers`: The implementation of context parallelism modifies the buffers in-place, converting them to `torch.distributed.tensor.Dtensor`s. After the context manager exits, a communication kernel would need to be launched to restore the buffers to their original state (usually all-gather). This takes some time, so it is recommended to pass the same tensors as in the `buffers` argument, to avoid unnecessary communication, unless you are sure that you need to use the buffers after the context manager exits.
+
+
+> [!Warning]
+> Context parallelism is not compatible with `labels` that are a copy of `input_ids`, which models from 🤗 transformers can shift to enable causal language modeling themselves.
+> Imagine this case:
+> labels = [l1, l2, l3, l4, ... li]
+> if we apply context parallelism, each rank would end up with a part of labels, such as this:
+> labels_rank_0 = [l1, l2], labels_rank_1 = [l3, l4], ...
+> after transformers modelling code shifts the labels, it would end up with:
+> labels_rank_0 = [l2, PAD], labels_rank_1 = [l3, PAD], ...
+> where `PAD` is a padding token. This would result in incorrect loss computation, as the labels are not aligned with the inputs anymore.
+> Because of this, you need to manually shift the labels before passing them in the model
+
+
+## Configurable options
+Accelerate provides only a single option to configure context parallelism (except for `cp_size`)
+
+- `cp_comm_strategy`: The rotation method to use for the shards. We strongly recommend keeping this as `"allgather"`, as it's very likely it will outperform `"alltoall"` in most cases.
+
+Context parallel size is rather self-explanatory, it's the number of ranks across which the inputs are to be-sharded.
+Context parallel shard rotation defines how the shards of the inputs are rotated across ranks. We'll cover the 2 options in more detail in the next section.
+
+You can see an end-to-end example in the [ND parallel example](https://github.com/huggingface/accelerate/blob/main/examples/fsdp2/nd_parallel.py) file, where you can train an 8B model with up-to 128k context length on a single 8xH100 node. Using multi-node training, you can scale this to 1M+ sequence length on multiple GPUs. You can also seamlessly combine it with other parallelism strategies to fit your needs.
+
+## Technical details
+
+> [!Tip]
+> This section is fairly technical, so if you don't need to learn the internals of context parallelism, you can skip it and start building 🚀
+
+We're going to be using word `shard` extensively in the following sections, so let's define it first. If we call tensor `sharded` across `Dth` dimension, across `N` ranks, we mean that this tensor is split into `N` parts, where each part of the tensor has shape `[..., D//N, ...]`.
+
+
+## So how does it work?
+
+Context parallelism works on sharding the `Q, K and V` matrices across the sequence dimension. Each rank has its assigned shard of `Q`, let's call it `Q_i`. This matrix stays only on this rank, during the whole computation. Similarly, each rank has its own shard of `K` and `V`, let's call them `K_i` and `V_i`. Then, each rank calculates attention with its own shard of `Q_i`, `K_i` and `V_i`, let's call it `attn_i`. During this computation, a communication kernel is launched to gather the `Ks` and `Vs` from all other ranks. What communication primitive is used, depends on the `context_parallel_shard_rotation` option.
+This way, each rank gets to calculate local attention, first with `Q_i`, `K_i` and `V_i`, then with `K_j` and `V_j` from all other ranks. As each rank holds `Q, K and V` matrices that are sharded across the sequence dimension, the resulting matrices are smaller and can fit on a single GPU.
+
+We can formalize this in the following pseudocode:
+```python
+comm_kernel = {"allgather": allgather, "alltoall": alltoall}[context_parallel_shard_rotation]
+Qi, Ki, Vi = shard(Q, K, V, seq_dim)
+attn[i] = attn(Qi, Ki, Vi)
+for j in range(context_parallel_size):
+    Kj, Vj = comm_kernel()
+    attn[j] = attn(Qi, Kj, Vj) # [batch, num_heads, seq_len // context_parallel_size, head_dim]
+
+final_attn = combine(attn)
+```
+
+## all-to-all vs all-gather
+
+### all-gather
+So what's the difference between all-to-all and all-gather? With all-gather, the communication is very simple. After (well, before, as it usually takes longer) we compute the local attention `attn_i` we launch an all-gather to gather all other `Ks` and `Vs` from all other ranks. As this communication is done, each rank has all the `Ks` and `Vs` from all other ranks, and can compute the attention with them sequentially.
+In ideal scenario, all-gather finishes in the exact moment as the calculation of `attn_i` is done. However, this never happens in practice, so the ideal real overlap is achieved when the full `attn_i` is overlapped with a part of the communication, then to start the computation with `K_j` and `V_j`, we wait for the all-gather to finish.
+
+### all-to-all
+All-to-all, or sometimes called `ring-rotation` utilizes a ring-like communication pattern. After concluding `attn_i` computation, an all-to-all is launched to send `K_i` and `V_i` to the neighbouring ranks. We then repeat this `context_parallel_size-1` times, so that each rank sees all the shards of `K` and `V` from all other ranks once. In ideal scenario, we prefetch shards `K_i+1` and `V_i+1` from the neighbouring rank and this communication is exactly overlapped with computation of our current `attn_i`. Again, realistically, this perfect overlap doesn't ever happen. Given the nature of this approach, if we don't achieve perfect overlap, the penalty is way larger than with all-gather.
+
+## How to choose the right rotation method?
+In theory, all-to-all should be the better choice. Though in practice, it rarely is. Therefore, we default to all-gather, as it's more likely to achieve better performance. Extensive [benchmarks](https://discuss.pytorch.org/t/distributed-w-torchtitan-breaking-barriers-training-long-context-llms-with-1m-sequence-length-in-pytorch-using-context-parallel/215082) from the `torchtitan` team also show that all-to-all rarely outperforms all-gather. Though, we still provide both options, as you might find one to be better for your use case.
+
+You can directly see this issue in the profiler output in the image below:
+<p align="center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/examples/fsdp2/cp_all_to_all.png" alt="all-to-all profiler output" />
+  <br>
+  <em>Figure 1: In red you can see the idle time, while we wait for the all-to-all kernel to finish. Highlighted in the first blue bar, you can see that it takes ~250us to finish, which is repeated N-1 times for each attention call, where N is the context parallel size.</em>
+</p>
+
+
+## Why only FSDP2?
+
+We only support context parallelism with `FSDP2`, as we create a joint mesh of `context_parallel_size` and `dp_shard_size` to 
+utilize its full potential. 
+How it works is: we shard the model across the joint mesh of size `cp_size*dp_shard_size`, which maximizes the memory savings.
+This is a "free lunch" of sorts, as `FSDP` communication is fully overlapped with the computation of attention, as shown in the images below.
+
+<p align="center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/examples/fsdp2/cp_why_fsdp2.png" alt="why FSDP2+CP" />
+  <br>
+  <em>Figure 2: In blue rectangles (Stream 23), you can see that the pre-fetch of `FSDP` shard is fully overlapped with the computation of attention (Stream 7), while in red rectangles (Stream 24), you can see that the all-gather kernel results in a bubble of idle time, in which our compute stream (7) is idle.</em>
+</p>
+
+In the figure above, you can also note the difference between all-to-all and all-gather. While in all-to-all (Figure 1), we launch a communication kernel N-1 times for each attention call, in all-gather (Figure 2), we launch a communication kernel only once. This results in a bigger bubble, but it only happens once per attention call, while in all-to-all, it happens N-1 times.
+
+## Data dispatching in joint mesh
+
+We make sure to dispatch the same batch of data to the whole `cp` subgroup, so that the results are correct. (Meaning each rank in `cp` subgroup gets the same batch of data.) However, we also dispatch different batches to each rank of `dp_shard` group.
+Imagine it like this:
+```
+# 8 GPUS, --dp_shard_size 4, --cp_size 2
+# mesh = [[0, 1], [2, 3], [4, 5], [6, 7]]
+# model is sharded across the whole mesh (each GPU holds 1/8 of the model)
+# GPUs 0,1 = batch 0
+# GPUs 2,3 = batch 1
+... and so on.
+```
+

docs/source/concept_guides/fsdp1_vs_fsdp2.md

@@ -0,0 +1,105 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ 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.
+-->
+
+# FSDP1 vs FSDP2
+
+This guide explains the key differences between `FSDP1` and `FSDP2` and helps you migrate your existing code to use `FSDP2` with minimal changes.
+
+## How is FSDP2 better than FSDP1?
+
+First, we want to understand how `FSDP1` and `FSDP2` work internally to understand the differences between them. This also helps us understand the limitations of `FSDP1` and how `FSDP2` solves them.
+
+We'll be discussing a scenario where we have a single `Layer` that contains 3 `Linear` layers and is wrapped using `FSDP` to be sharded across 2 GPUs.
+
+<div align="center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/layer.png" alt="Layer">
+</div>
+
+### FSDP1
+First, we have to understand the original `FSDP1` and the limitations it brings. It represents each `FSDP` module as a single `FlatParameter` which is a single 1D tensor that contains all of the module parameters, which then get sharded across ranks. I.e. if you wrap the `Layer` with `FSDP1`, you'd achieve something as such:
+
+<div align="center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/fsdp1.png" alt="FSDP1">
+</div>
+
+You might notice a problem. The whole `Layer` gets flattened into a single `FlatParameter`, which then gets sharded across ranks. But if it's a single `FlatParameter` object, how do we store metadata? That is one of the limitations. Properly storing per-parameter metadata such as `dtype`, `requires_grad`, etc. is not possible without some ugly hacks.
+
+### FSDP2
+This is why `FSDP2` was introduced. It doesn't use `FlatParameter`, instead it uses `DTensor` which is short for "Distributed Tensor". Each `DTensor` basically represents a vanilla `torch.Tensor` that has been sharded across ranks. It contains metadata about the original `torch.Tensor` and how it's sharded, what is the [placement type](https://pytorch.org/docs/stable/distributed.tensor.html#module-torch.distributed.tensor.placement_types) and so on. This is why it's called `per-parameter sharding`. The following figure shows the difference:
+
+<div align="center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/fsdp2.png" alt="FSDP2">
+</div>
+
+Each Parameter of the original `Layer` is sharded across the 0th dimension, and split between 2 GPUs. Now, each `Linear` layer is a separate `DTensor` and storing metadata per-parameter is possible and straightforward.
+
+
+> [!TIP] 
+> In the image above, the tensors were sharded across the 1st dimension for the sake of fitting the image on the screen, in reality, they are sharded across the 0th dimension as stated above
+
+## What does FSDP2 offer?
+
+`FSDP2` is a new and improved version of PyTorch's fully-sharded data parallel training API. Its main advantage is using `DTensor` to represent sharded parameters. Compared to `FSDP1`, it offers:
+- Simpler internal implementation, where each `Parameter` is a separate `DTensor`
+- Enables simple partial parameter freezing because of the above, which makes methods as [`LORA`](https://arxiv.org/abs/2106.09685) work out of the box
+- With `DTensor`, `FSDP2` supports mixing `fp8` and other parameter types in the same model out of the box
+- Faster and simpler checkpointing without extra communication across ranks using `SHARDED_STATE_DICT` and [`torch.distributed.checkpoint`](https://pytorch.org/docs/stable/distributed.checkpoint.html), this way, each rank only saves its own shard and corresponding metadata
+- For loading, it uses a `state_dict` of the sharded model to directly load the sharded parameters
+- Support for asynchronous checkpointing, where parameters are first copied to CPU memory, after this, main thread continues training while another thread stores the parameters on disk
+- Memory efficiency and deterministic memory usage, `FSDP2` doesn't use `recordStream` anymore and uses stream-to-stream synchronization (for more technical details see [this forum post](https://dev-discuss.pytorch.org/t/fsdp-cudacachingallocator-an-outsider-newb-perspective/1486) and [this issue](https://github.com/pytorch/pytorch/issues/114299))
+- In the future, optimizations of the communication patterns via `torch.compile` are planned, further improving the performance and memory efficiency
+
+
+## API Differences
+
+We have already discussed the internal differences, now let's discuss the differences, you, as a user, will need to know. 
+
+Here are the main changes in configuration options when using `FSDP2` through the `accelerate` CLI:
+
+Previous (`FSDP1`) | New (`FSDP2`) | What Changed
+-- | -- | --
+`--fsdp_sharding_strategy` | `--fsdp_reshard_after_forward` | replaces `--fsdp_sharding_strategy`, changed to `true` (previously `FULL_SHARD`) or `false` (previously `SHARD_GRAD_OP`)
+`--fsdp_backward_prefetch` | \*\***REMOVED**\*\* | `FSDP2` uses previous `BACKWARD_PRE` option by default, as only this allows communication and computation overlap
+`--fsdp_forward_prefetch` | \*\***NOT YET IMPLEMENTED**\*\* | How to implement this is under active discussion, for now it is not supported in `FSDP2`
+`--fsdp_sync_module_states` | \*\***REMOVED**\*\* | with `FSDP2`, this parameter becomes redundant
+`--fsdp_cpu_ram_efficient_loading` | `--fsdp_cpu_ram_efficient_loading` | if `true`, `FSDP2` will similarly load the model only on rank 0, and then parameters get synced to other ranks, this is the same behavior as `FSDP1`, however, setting `--fsdp_sync_module_states` isn't required anymore
+`--fsdp_state_dict_type` | `--fsdp_state_dict_type` | `LOCAL_STATE_DICT` becomes obsolete and with `FSDP2` `SHARDED_STATE_DICT` is the default option, which results in no extra communication and each rank saving its own shard, other possible option is `FULL_STATE_DICT` which results in extra communication and spike in memory usage but saves the full model from rank 0.
+`--fsdp_use_orig_params` | \*\***REMOVED**\*\* | `FSDP2` uses a `DTensor` class on the background, which means it *always* uses the original parameters by default
+\*\***NEW**\*\* | `--fsdp_version` | `1` is the default option, to not break existing code, set to `2` to use `FSDP2`
+
+For all other options that remain unchanged, see the [`FSDP` documentation](../usage_guides/fsdp.md).
+
+## How to Switch to FSDP2
+
+### If using Python code:
+Simply set `fsdp_version=2` when creating your plugin and replace options according to the table above.
+
+```python
+from accelerate import FullyShardedDataParallelPlugin, Accelerator
+
+fsdp_plugin = FullyShardedDataParallelPlugin(
+    fsdp_version=2
+    # other options...
+)
+accelerator = Accelerator(fsdp_plugin=fsdp_plugin)
+```
+
+### If using YAML config:
+Use our conversion tool:
+```bash
+accelerate to-fsdp2 --config_file config.yaml --output_file new_config.yaml
+```
+
+This will automatically convert all FSDP1 settings to their FSDP2 equivalents. Use `--overwrite` to update the existing file instead of creating a new one.

docs/source/concept_guides/fsdp_and_deepspeed.md

@@ -109,7 +109,7 @@ While FSDP require an explicit `--fsdp_cpu_ram_efficient_loading true` to activa
 <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.
+    For RAM efficient loading the weights will be loaded only in a single rank, and thus requires `sync_module_states` to broadcast weights to other ranks.
 
 </Tip>
 
@@ -125,7 +125,7 @@ FSDP requires an explicit `--fsdp_auto_wrap_policy` for the algorithm to decide
 
 ### 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.
+FSDP requires an explicit `--fsdp_use_orig_params` flag if using `torch.compile`, see [the pytorch documentation](https://pytorch.org/docs/stable/fsdp.html#module-torch.distributed.fsdp). For DeepSpeed this is transparent to the user.
 
 <Tip>
 
@@ -147,7 +147,7 @@ Deepspeed requires explicit `--gradient_accumulation_steps` and `--gradient_clip
 
 ## 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.
+To discuss 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>
 
@@ -166,7 +166,7 @@ Optimizer (Actual Step) | ✅ | FSDP<br>DeepSpeed  | occurs in `torch_dtype` <br
 
 <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.
+    Therefore when using DeepSpeed a small number of GPUs, be aware of potentially significant memory overheads due to the upcasting during preparation.
 
 </Tip>
 

docs/source/concept_guides/internal_mechanism.md

@@ -71,4 +71,4 @@ setting the same seed in the main random number generator in all processes.
 
 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).
+For more details about the internals, see the [Internals page](../package_reference/torch_wrappers).

docs/source/package_reference/big_modeling.md

@@ -63,6 +63,10 @@ rendered properly in your Markdown viewer.
 
 [[autodoc]] hooks.SequentialHook
 
+### LayerwiseCastingHook
+
+[[autodoc]] hooks.LayerwiseCastingHook
+
 ## Adding Hooks
 
 ### add_hook_to_module
@@ -81,6 +85,10 @@ rendered properly in your Markdown viewer.
 
 [[autodoc]] hooks.attach_align_device_hook_on_blocks
 
+### attach_layerwise_casting_hooks
+
+[[autodoc]] big_modeling.attach_layerwise_casting_hooks
+
 ## Removing Hooks
 
 ### remove_hook_from_module
@@ -99,4 +107,4 @@ rendered properly in your Markdown viewer.
 
 ### align_module_device
 
-[[autodoc]] utils.align_module_device
+[[autodoc]] utils.align_module_device

docs/source/package_reference/cli.md

@@ -139,7 +139,7 @@ values. They can also be passed in manually.
 * `--cpu` (`bool`) -- Whether or not to force the training on the CPU.
 * `--multi_gpu` (`bool`) -- Whether or not this should launch a distributed GPU training.
 * `--tpu` (`bool`) -- Whether or not this should launch a TPU training.
-* `--ipex` (`bool`) -- Whether or not this should launch an Intel Pytorch Extension (IPEX) training.
+* `--ipex` (`bool`) -- Whether or not this should launch an Intel Pytorch Extension (IPEX) training. **This argument is deprecated, will be removed in Accelerate v1.10**
 
 **Resource Selection Arguments**:
 
@@ -158,13 +158,13 @@ The following arguments are useful for selecting which training paradigm to use.
 * `--use_deepspeed` (`bool`) -- Whether or not to use DeepSpeed for training.
 * `--use_fsdp` (`bool`) -- Whether or not to use FullyShardedDataParallel for training.
 * `--use_megatron_lm` (`bool`) -- Whether or not to use Megatron-LM for training.
-* `--use_xpu` (`bool`) -- Whether to use IPEX plugin to speed up training on XPU specifically.
+* `--use_xpu` (`bool`) -- Whether to use IPEX plugin to speed up training on XPU specifically. **This argument is deprecated and ignored, will be removed in Accelerate v1.10**
 
 **Distributed GPU Arguments**:
 
 The following arguments are only useful when `multi_gpu` is passed or multi-gpu training is configured through `accelerate config`: 
 
-* `--gpu_ids` (`str`) -- What GPUs (by id) should be used for training on this machine as a comma-seperated list
+* `--gpu_ids` (`str`) -- What GPUs (by id) should be used for training on this machine as a comma-separated list
 * `--same_network` (`bool`) -- Whether all machines used for multinode training exist on the same local network.
 * `--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.

docs/source/package_reference/fsdp.md

@@ -30,3 +30,17 @@ rendered properly in your Markdown viewer.
 ## FullyShardedDataParallelPlugin
 
 [[autodoc]] utils.FullyShardedDataParallelPlugin
+
+## fsdp2_load_full_state_dict
+
+[[autodoc]] utils.fsdp2_load_full_state_dict
+
+## fsdp2_switch_optimizer_parameters
+
+[[autodoc]] utils.fsdp2_switch_optimizer_parameters
+
+## fsdp2_prepare_model
+
+[[autodoc]] utils.fsdp2_prepare_model
+
+## fsdp2_prepare_auto_wrap_policy

docs/source/package_reference/tracking.md

@@ -29,6 +29,11 @@ rendered properly in your Markdown viewer.
 [[autodoc]] tracking.WandBTracker
     - __init__
 
+## Trackio
+
+[[autodoc]] tracking.TrackioTracker
+    - __init__
+
 ## CometMLTracker
 
 [[autodoc]] tracking.CometMLTracker
@@ -48,3 +53,8 @@ rendered properly in your Markdown viewer.
 
 [[autodoc]] tracking.ClearMLTracker
     - __init__
+
+## SwanLabTracker
+
+[[autodoc]] tracking.SwanLabTracker
+    - __init__

docs/source/package_reference/utilities.md

@@ -208,6 +208,7 @@ These utilities relate to interacting with PyTorch models
 
 [[autodoc]] utils.set_module_tensor_to_device
 
+[[autodoc]] utils.get_module_children_bottom_up
 
 ## Parallel
 

docs/source/usage_guides/compilation.md

@@ -0,0 +1,76 @@
+# Compilation
+
+## Overview
+
+Pytorch 2.0 introduced `torch.compile`, a powerful feature that makes PyTorch code run faster by JIT-compiling PyTorch code into optimized kernels. Key features of `torch.compile` include:
+
+- **Performance Improvement**: Significantly speeds up model execution by optimizing the computation graph.
+- **Ease of Use**: Requires minimal code changes to implement, making it highly accessible.
+- **Compatibility**: Works seamlessly with existing PyTorch code and models.
+
+When used with Accelerate, `torch.compile` integrates smoothly into distributed training workflows, allowing you to benefit from both distributed execution and compilation optimizations simultaneously.
+
+The first execution of compiled code typically takes longer as it includes the compilation time, but subsequent runs are significantly faster. For optimal performance in different scenarios, `torch.compile` offers various modes like `"default"`, `"reduce-overhead"` (which uses CUDA graphs to further reduce overhead), and `"max-autotune"` (which performs extensive autotuning to find the best kernels for your model).
+
+## Using `torch.compile` with Accelerate
+
+Accelerate provides `TorchDynamoPlugin` for easy and seemless integration of `torch.compile` into your training scripts.
+
+```python
+from accelerate import Accelerator
+from accelerate.utils import TorchDynamoPlugin
+
+# Configure the compilation backend
+dynamo_plugin = TorchDynamoPlugin(
+    backend="inductor",  # Options: "inductor", "aot_eager", "aot_nvfuser", etc.
+    mode="default",      # Options: "default", "reduce-overhead", "max-autotune"
+    fullgraph=True,
+    dynamic=False
+)
+
+# Initialize accelerator with the plugin
+accelerator = Accelerator(dynamo_plugin=dynamo_plugin)
+# This will apply torch.compile to your model
+model = accelerator.prepare(model)
+```
+
+It is compatible with all other features and plugins of Accelerate, including mixed precision, distributed training (DDP, FSDP, Deepspeed), etc.
+
+## Regional Compilation
+
+Instead of trying to compile the whole model, which usually has a big problem space for optimization. Regional compilation targets repeated blocks of the same class and compiles them sequentially to hit the compiler's cache. For example, in `GPT2LMHeadModel`, the repeated block/class is `GPT2Block`, and can be accessed as `model.transformer.h[0]`. The rest of the model (e.g model.lm_head) is compiled separately.
+
+This allows us to speed up the compilation overhead / cold start of models like LLMs and Transformers in general.
+See <https://pytorch.org/tutorials/recipes/regional_compilation.html> for more details.
+
+### How to Use Regional Compilation
+
+It can be enabled by setting `use_regional_compilation=True` in the `TorchDynamoPlugin` configuration:
+
+```python
+# Configure the compilation backend
+dynamo_plugin = TorchDynamoPlugin(
+    use_regional_compilation=True,
+    ... # other parameters
+)
+# Initialize accelerator with the plugin
+accelerator = Accelerator(dynamo_plugin=dynamo_plugin)
+# This will apply compile_regions to your model
+model = accelerator.prepare(model)
+```
+
+You could also use the `accelerate.utils.compile_regions` utility directly the same way you would use `torch.compile`.
+
+### Benefits of Regional Compilation
+
+We have conducted extensive benchmarks comparing full compilation and regional compilation using the `torch.compile` feature in PyTorch. The full results are available in the [accelerate repository](https://github.com/huggingface/accelerate/tree/main/benchmarks/torch.compile/regional_compilation). The key findings from our benchmarks are:
+
+1. **Comparable Performance**: Regional compilation delivers performance speedups similar to full compilation, especially for larger models.
+2. **Faster Compilation**: Regional compilation significantly reduces the time taken to compile models, making it a more efficient choice for deployment.
+3. **Batch Size Impact**: The performance difference between compilation strategies diminishes with larger batch sizes, indicating that the overhead of compilation is less impactful in those scenarios.
+4. **Model Size Consideration**: The benefits of regional compilation are more pronounced in larger models, where the compilation time savings can be substantial.
+5. **Practical Application**: For real-world applications, regional compilation is a practical choice for optimizing training cold start times, especially when working with large models.
+
+## Conclusion
+
+Both full and regional compilation can significantly speed up your models. Regional compilation offers a practical balance between compilation time and runtime performance, especially for training large models with substantial batch sizes.

docs/source/usage_guides/ddp_comm_hook.md

@@ -34,6 +34,10 @@ In this tutorial, you will see how to quickly set up DDP communication hooks and
 import torch
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.distributed.algorithms.ddp_comm_hooks import default_hooks
+from accelerate.test_utils.testing import get_backend
+
+device_type, _, _ = get_backend()
+device_id = getattr(torch, device_type, torch.cuda).current_device()
 
 class MyModel(torch.nn.Module):
     def __init__(self):
@@ -44,7 +48,7 @@ class MyModel(torch.nn.Module):
         return self.layer(x)
 
 model = MyModel()
-model = DDP(model, device_ids=[torch.cuda.current_device()])
+model = DDP(model, device_ids=[device_id])
 model.register_comm_hook(state=None, hook=default_hooks.fp16_compress_hook)
 
 # Training loop
@@ -108,6 +112,10 @@ BF16 Compression Hook API is experimental, and it requires NCCL version later th
 import torch
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.distributed.algorithms.ddp_comm_hooks import default_hooks
+from accelerate.test_utils.testing import get_backend
+
+device_type, _, _ = get_backend()
+device_id = getattr(torch, device_type, torch.cuda).current_device()
 
 class MyModel(torch.nn.Module):
     def __init__(self):
@@ -118,7 +126,7 @@ class MyModel(torch.nn.Module):
         return self.layer(x)
 
 model = MyModel()
-model = DDP(model, device_ids=[torch.cuda.current_device()])
+model = DDP(model, device_ids=[device_id])
 model.register_comm_hook(state=None, hook=default_hooks.bf16_compress_hook)
 
 # Training loop
@@ -182,6 +190,10 @@ PowerSGD typically requires extra memory of the same size as the model’s gradi
 import torch
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.distributed.algorithms.ddp_comm_hooks import powerSGD_hook
+from accelerate.test_utils.testing import get_backend
+
+device_type, _, _ = get_backend()
+device_id = getattr(torch, device_type, torch.cuda).current_device()
 
 class MyModel(torch.nn.Module):
     def __init__(self):
@@ -192,7 +204,7 @@ class MyModel(torch.nn.Module):
         return self.layer(x)
 
 model = MyModel()
-model = DDP(model, device_ids=[torch.cuda.current_device()])
+model = DDP(model, device_ids=[device_id])
 state = powerSGD_hook.PowerSGDState(process_group=None)
 model.register_comm_hook(state=state, hook=powerSGD_hook.powerSGD_hook)
 

docs/source/usage_guides/deepspeed.md

@@ -167,7 +167,7 @@ Currently, `Accelerate` supports following config through the CLI:
 `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_multinode_launcher`: DeepSpeed multi-node launcher to use, e.g. `pdsh`, `standard`, `openmpi`, `mvapich`, `mpich`, `slurm`, `nossh` (requires DeepSpeed >= 0.14.5). 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.
@@ -194,7 +194,7 @@ For instance, here is how you would run the NLP example `examples/by_feature/dee
 ```bash
 compute_environment: LOCAL_MACHINE
 deepspeed_config:
- deepspeed_config_file: /home/ubuntu/accelerate/examples/configs/deepspeed_config_templates/zero_stage2_config.json
+ deepspeed_config_file: /home/ubuntu/accelerate/examples/deepspeed_config_templates/zero_stage2_config.json
  zero3_init_flag: true
 distributed_type: DEEPSPEED
 fsdp_config: {}
@@ -275,7 +275,7 @@ accelerate launch examples/by_feature/deepspeed_with_config_support.py \
 ```bash
 compute_environment: LOCAL_MACHINE
 deepspeed_config:
- deepspeed_config_file: /home/ubuntu/accelerate/examples/configs/deepspeed_config_templates/zero_stage3_offload_config.json
+ deepspeed_config_file: /home/ubuntu/accelerate/examples/deepspeed_config_templates/zero_stage3_offload_config.json
  zero3_init_flag: true
 distributed_type: DEEPSPEED
 fsdp_config: {}
@@ -710,6 +710,13 @@ model, eval_dataloader = accelerator.prepare(model, eval_dataloader)
 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.
 
+## Multi-node DeepSpeed
+DeepSpeed supports multi-node inference and training over a variety of different launchers. You can specify a different launcher by setting the `deepspeed_multinode_launcher` config in the CLI or in the DeepSpeed config file.
+
+Currently, accelerate supports passing configuration for the following DeepSpeed multi-node launchers: `pdsh` (default), `standard`, `openmpi`, `mvapich`, `mpich`, `slurm`, `nossh` (requires DeepSpeed >= 0.14.5).
+
+Please read the [DeepSpeed documentation](https://www.deepspeed.ai/getting-started/#resource-configuration-multi-node) for more information on the different launchers. By default, DeepSpeed will attempt to use passwordless SSH from the main machine node to the other nodes to perform the launcher command. In this configuration, the accelerate launch command only needs to be run on the main node. If using the `nossh` launcher, you will need to run the accelerate launch command on every node using copied configuration. 
+
 ## DeepSpeed Resources
 
 The documentation for the internals related to deepspeed can be found [here](../package_reference/deepspeed).

docs/source/usage_guides/gaudi.md

@@ -0,0 +1,38 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ 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.
+-->
+
+# Intel Gaudi
+
+Users can take advantage of Intel Gaudi AI accelerators for significantly faster and cost-effective model training and inference.
+The Intel Gaudi AI accelerator family currently includes three product generations: [Intel Gaudi 1](https://habana.ai/products/gaudi/), [Intel Gaudi 2](https://habana.ai/products/gaudi2/), and [Intel Gaudi 3](https://habana.ai/products/gaudi3/). Each server is equipped with 8 devices, known as Habana Processing Units (HPUs), providing 128GB of memory on Gaudi 3, 96GB on Gaudi 2, and 32GB on the first-gen Gaudi. For more details on the underlying hardware architecture, check out the [Gaudi Architecture Overview](https://docs.habana.ai/en/latest/Gaudi_Overview/Gaudi_Architecture.html).
+
+## How it works out of the box
+
+It is enabled by default if an Intel Gaudi device is detected.
+To disable it, pass `--cpu` flag to `accelerate launch` command or answer the corresponding question when answering the `accelerate config` questionnaire.
+
+You can directly run the following script to test it out on Intel Gaudi:
+
+```bash
+accelerate launch /examples/cv_example.py --data_dir images
+```
+
+## Limitations
+
+The following features are not part of the Accelerate library and requires [Optimum for Intel Gaudi](https://huggingface.co/docs/optimum/main/en/habana/index):
+
+- `fast_ddp` which implements DDP by applying an all-reduce on gradients instead of the Torch DDP wrapper.
+- `minimize_memory` which is used for fp8 training and enables keeping fp8 weights in memory between the forward and backward passes, leading to a smaller memory footprint at the cost of additional fp8 casts.
+- `context_parallel_size` which is used for Context/Sequence Parallelism (CP/SP) and partitions the network inputs and activations along sequence dimension to reduce memory footprint and increase throughput.

docs/source/usage_guides/gradient_accumulation.md

@@ -245,7 +245,7 @@ As was pointed out in this [blog-post](https://huggingface.co/blog/gradient_accu
 
 >  [...] 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.
+In other words, some adjustments must be made on losses that operate on a token-level basis.
 
 ### Skeleton code
 
@@ -282,7 +282,7 @@ for update_step in range(total_updates):
         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
+            # if we perform gradient accumulation in a multi-devices set-up, we want to avoid unnecessary 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
@@ -294,7 +294,7 @@ for update_step in range(total_updates):
             with ctx():
                 inputs, targets = batch
                 outputs = model(inputs)
-                loss = loss_function(outputs, targets) # the loss function shoud sum over samples rather than averaging
+                loss = loss_function(outputs, targets) # the loss function should 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
@@ -394,7 +394,7 @@ for update_step in range(total_gradient_updates):
         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
+            # if we perform gradient accumulation in a multi-devices set-up, we want to avoid unnecessary 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

docs/source/usage_guides/intel_cpu.md

@@ -13,34 +13,11 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 -->
 
-# Intel® Extension for PyTorch
-
-[IPEX](https://github.com/intel/intel-extension-for-pytorch) is optimized for CPUs with AVX-512 or above, and functionally works for CPUs with only AVX2. So, it is expected to bring performance benefit for Intel CPU generations with AVX-512 or above while CPUs with only AVX2 (e.g., AMD CPUs or older Intel CPUs) might result in a better performance under IPEX, but not guaranteed. IPEX provides performance optimizations for CPU training with both Float32 and BFloat16. The usage of BFloat16 is the main focus of the following sections.
-
-Low precision data type BFloat16 has been natively supported on the 3rd Generation Xeon® Scalable Processors (aka Cooper Lake) with AVX512 instruction set and will be supported on the next generation of Intel® Xeon® Scalable Processors with Intel® Advanced Matrix Extensions (Intel® AMX) instruction set with further boosted performance. The Auto Mixed Precision for CPU backend has been enabled since PyTorch-1.10. At the same time, the support of Auto Mixed Precision with BFloat16 for CPU and BFloat16 optimization of operators has been massively enabled in Intel® Extension for PyTorch, and partially upstreamed to PyTorch master branch. Users can get better performance and user experience with IPEX Auto Mixed Precision.
-
-## IPEX installation:
-
-IPEX release is following PyTorch, to install via pip:
-
-| PyTorch Version   | IPEX version   |
-| :---------------: | :----------:   |
-| 2.0               |  2.0.0         |
-| 1.13              |  1.13.0        |
-| 1.12              |  1.12.300      |
-| 1.11              |  1.11.200      |
-| 1.10              |  1.10.100      |
-
-```
-pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
-```
-
-Check more approaches for [IPEX installation](https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/installation.html).
-
+# Training on Intel CPU
 
 ## 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 full support for Intel CPU, all you need to do is enabling it through the config.
 
 **Scenario 1**: Acceleration of No distributed CPU training
 
@@ -55,7 +32,6 @@ This machine
 Which type of machine are you using?
 No distributed training
 Do you want to run your training on CPU only (even if a GPU / Apple Silicon device is available)? [yes/NO]:yes
-Do you want to use Intel PyTorch Extension (IPEX) to speed up training on CPU? [yes/NO]:yes
 Do you wish to optimize your script with torch dynamo?[yes/NO]:NO
 Do you want to use DeepSpeed? [yes/NO]: NO
 -----------------------------------------------------------------------------------------------------------------------------------------------------------
@@ -69,15 +45,12 @@ default options when doing
 accelerate launch my_script.py --args_to_my_script
 ```
 
-For instance, here is how you would run the NLP example `examples/nlp_example.py` (from the root of the repo) with IPEX enabled.
-default_config.yaml that is generated after `accelerate config`
+For instance, here is how you would run the NLP example `examples/nlp_example.py` (from the root of the repo) with `default_config.yaml` which is generated by `accelerate config`
 
 ```bash
 compute_environment: LOCAL_MACHINE
 distributed_type: 'NO'
 downcast_bf16: 'no'
-ipex_config:
-  ipex: true
 machine_rank: 0
 main_training_function: main
 mixed_precision: bf16
@@ -94,6 +67,9 @@ use_cpu: true
 accelerate launch examples/nlp_example.py
 ```
 
+> [!CAUTION]
+> `accelerator.prepare` can currently only handle simultaneously preparing multiple models (and no optimizer) OR a single model-optimizer pair for training. Other attempts (e.g., two model-optimizer pairs) will raise a verbose error. To work around this limitation, consider separately using `accelerator.prepare` for each model-optimizer pair.
+
 **Scenario 2**: Acceleration of distributed CPU training
 we use Intel oneCCL for communication, combined with Intel® MPI library to deliver flexible, efficient, scalable cluster messaging on Intel® architecture. you could refer the [here](https://huggingface.co/docs/transformers/perf_train_cpu_many) for the installation guide
 
@@ -114,7 +90,6 @@ What is the rank of this machine?
 What is the IP address of the machine that will host the main process? 36.112.23.24
 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
@@ -126,13 +101,11 @@ bf16
 ```
 For instance, here is how you would run the NLP example `examples/nlp_example.py` (from the root of the repo) with IPEX enabled for distributed CPU training.
 
-default_config.yaml that is generated after `accelerate config`
+`default_config.yaml` which is generated by `accelerate config`
 ```bash
 compute_environment: LOCAL_MACHINE
 distributed_type: MULTI_CPU
 downcast_bf16: 'no'
-ipex_config:
-  ipex: true
 machine_rank: 0
 main_process_ip: 36.112.23.24
 main_process_port: 29500
@@ -153,8 +126,10 @@ 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.
+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 let 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
@@ -162,18 +137,18 @@ xxx.xxx.xxx.xxx #node1 ip
 xxx.xxx.xxx.xxx #node2 ip
 xxx.xxx.xxx.xxx #node3 ip
 ```
-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.
+
+Before executing `accelerate launch` command, you need source the oneCCL bindings `setvars.sh` to get your Intel MPI environment properly. Note that both the python script and environment need to be available 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.
+
+You can also directly launch distributed training with `mpirun` command, you need to 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 available 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
@@ -182,11 +157,3 @@ export MASTER_ADDR=xxx.xxx.xxx.xxx #node0 ip
 export CCL_ATL_TRANSPORT=ofi
 mpirun -f hostfile -n 16 -ppn 4 accelerate launch examples/nlp_example.py
 ```
-
-## Related Resources
-
-- [Project's github](https://github.com/intel/intel-extension-for-pytorch)
-- [API docs](https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/api_doc.html)
-- [Tuning guide](https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/performance_tuning/tuning_guide.html)
-- [Blogs & Publications](https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/blogs_publications.html)
-

docs/source/usage_guides/low_precision_training.md

@@ -39,7 +39,7 @@ 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 one of the `RecipeKwargs` dataclasses such as [`utils.AORecipeKwargs`], [`utils.TERecipeKwargs`], or [`utils.MSAMPRecipeKwargs`]; you can also nclarify it in your config `yaml`/during `accelerate launch`:
+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 one of the `RecipeKwargs` dataclasses such as [`utils.AORecipeKwargs`], [`utils.TERecipeKwargs`], or [`utils.MSAMPRecipeKwargs`]; you can also clarify it in your config `yaml`/during `accelerate launch`:
 
 ```{python}
 from accelerate import Accelerator

docs/source/usage_guides/megatron_lm.md

@@ -19,7 +19,7 @@ rendered properly in your Markdown viewer.
 [Megatron-LM](https://github.com/NVIDIA/Megatron-LM) enables training large transformer language models at scale.
 It provides efficient tensor, pipeline and sequence based model parallelism for pre-training transformer based
 Language Models such as [GPT](https://arxiv.org/abs/2005.14165) (Decoder Only), [BERT](https://arxiv.org/pdf/1810.04805.pdf) (Encoder Only) and [T5](https://arxiv.org/abs/1910.10683) (Encoder-Decoder).
-For detailed information and how things work behind the scene please refer the github [repo](https://github.com/NVIDIA/Megatron-LM).
+For detailed information and how things work behind the scene please refer to the github [repo](https://github.com/NVIDIA/Megatron-LM).
 
 ## What is integrated?
 
@@ -30,7 +30,7 @@ a. **Tensor Parallelism (TP)**: Reduces memory footprint without much additional
 Each tensor is split into multiple chunks with each shard residing on separate GPU. At each step, the same mini-batch of data is processed
 independently and in parallel by each shard followed by syncing across all GPUs (`all-reduce` operation). 
 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
+For more details, please refer to the 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).
 
@@ -45,7 +45,7 @@ this section of blogpost [The Technology Behind BLOOM Training](https://huggingf
 
 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 
-post `all-reduce` by replacing then with `reduce-scatter` and `no-op` operation would be replaced by `all-gather`. 
+post `all-reduce` by replacing them with `reduce-scatter` and `no-op` operation would be replaced by `all-gather`. 
 As `all-reduce = reduce-scatter + all-gather`, this saves a ton of activation memory at no added communication cost. 
 To put it simply, it shards the outputs of each transformer layer along sequence dimension, e.g., 
 if the sequence length is `1024` and the TP size is `4`, each GPU will have `256` tokens (1024/4) for each sample. 
@@ -56,7 +56,7 @@ d. **Data Parallelism (DP)** via Distributed Optimizer: Reduces the memory footp
 (versus the traditional method of replicating the optimizer state across data parallel ranks). 
 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
+For more details, please refer to the research paper [ZeRO: Memory Optimizations Toward Training Trillion
 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).
 
@@ -66,7 +66,7 @@ For example, for GPT-3, this leads to 70% reduction in required memory for activ
 only 2.7% FLOPs overhead for recomputation of activations. For more details, please refer to the research paper 
 [Reducing Activation Recomputation in Large Transformer Models](https://arxiv.org/pdf/2205.05198.pdf).
 
-f. **Fused Kernels**: Fused Softmax, Mixed Precision Fused Layer Norm and  Fused gradient accumulation to weight gradient computation of linear layer.
+f. **Fused Kernels**: Fused Softmax, Mixed Precision Fused Layer Norm and Fused gradient accumulation to weight gradient computation of linear layer.
 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.
@@ -445,7 +445,7 @@ python checkpoint_utils/megatgron_gpt2/checkpoint_reshaping_and_interoperability
 ## Megatron-LM GPT models support returning logits and `megatron_generate` function for text generation
 
 1. Returning logits require setting `require_logits=True` in MegatronLMPlugin as shown below. 
-These would be available on the in the last stage of pipeline.
+These would be available in the last stage of pipeline.
 ```python
 megatron_lm_plugin = MegatronLMPlugin(return_logits=True)
 ```
@@ -569,7 +569,7 @@ 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.
+8. Below are the mapping from Megatron-LM model architectures to 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) : 

docs/source/usage_guides/tracking.md

@@ -20,10 +20,11 @@ Accelerate provides a general tracking API that can be used to log useful items
 
 ## Integrated Trackers
 
-Currently `Accelerate` supports seven trackers out-of-the-box:
+Currently `Accelerate` supports eight trackers out-of-the-box:
 
 - TensorBoard
-- WandB
+- WandB 
+- Trackio
 - CometML
 - Aim
 - MLFlow

examples/README.md

@@ -225,7 +225,7 @@ In [/slurm/submit_multinode.sh](./slurm/submit_multinode.sh) we must specify the
 
 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.
+In both scripts, we run `activateEnvironment.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 

examples/by_feature/cross_validation.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import argparse
-from typing import List
 
 import evaluate
 import numpy as np
@@ -61,7 +60,7 @@ EVAL_BATCH_SIZE = 32
 
 
 def get_fold_dataloaders(
-    accelerator: Accelerator, dataset: DatasetDict, train_idxs: List[int], valid_idxs: List[int], batch_size: int = 16
+    accelerator: Accelerator, dataset: DatasetDict, train_idxs: list[int], valid_idxs: list[int], batch_size: int = 16
 ):
     """
     Gets a set of train, valid, and test dataloaders for a particular fold

examples/by_feature/deepspeed_with_config_support.py

@@ -218,7 +218,7 @@ def parse_args():
         default="all",
         help=(
             'The integration to report the results and logs to. Supported platforms are `"tensorboard"`,'
-            ' `"wandb"`, `"comet_ml"`, and `"dvclive"`. Use `"all"` (default) to report to all integrations.'
+            ' `"wandb"`, `"comet_ml"`, `"dvclive"`, and `"swanlab"`. Use `"all"` (default) to report to all integrations.'
             "Only applicable when `--with_tracking` is passed."
         ),
     )

examples/by_feature/megatron_lm_gpt_pretraining.py

@@ -215,7 +215,7 @@ def parse_args():
         default="all",
         help=(
             'The integration to report the results and logs to. Supported platforms are `"tensorboard"`,'
-            ' `"wandb"`, `"comet_ml"`, and `"dvclive"`. Use `"all"` (default) to report to all integrations.'
+            ' `"wandb"`, `"comet_ml"`, and `"dvclive"`, and `"swanlab"`. Use `"all"` (default) to report to all integrations.'
             "Only applicable when `--with_tracking` is passed."
         ),
     )
@@ -611,7 +611,7 @@ def main():
 
             if isinstance(checkpointing_steps, int):
                 if completed_steps % checkpointing_steps == 0:
-                    output_dir = f"step_{completed_steps }"
+                    output_dir = f"step_{completed_steps}"
                     if args.output_dir is not None:
                         output_dir = os.path.join(args.output_dir, output_dir)
                     accelerator.save_state(output_dir)

examples/by_feature/profiler.py

@@ -31,8 +31,8 @@ from accelerate.utils import ProfileKwargs
 #
 # This example trains a Bert base model on GLUE MRPC
 # in any of the following settings (with the same script):
-#   - single CPU or single GPU
-#   - multi GPUS (using PyTorch distributed mode)
+#   - single CPU or single device (CUDA GPU, Intel XPU etc.)
+#   - multi devices (using PyTorch distributed mode)
 #   - (multi) TPUs
 #   - fp16 (mixed-precision) or fp32 (normal precision)
 #
@@ -183,7 +183,8 @@ def training_function(config, args):
         # New Code #
         accelerator.print(
             prof.key_averages().table(
-                sort_by="self_cpu_time_total" if args.cpu else "self_cuda_time_total", row_limit=-1
+                sort_by="self_cpu_time_total" if args.cpu else f"self_{accelerator.device.type}_time_total",
+                row_limit=-1,
             )
         )
 
@@ -215,7 +216,7 @@ def main():
         choices=["no", "fp16", "bf16", "fp8"],
         help="Whether to use mixed precision. Choose"
         "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10."
-        "and an Nvidia Ampere GPU.",
+        "and an Nvidia Ampere GPU or an Intel XPU.",
     )
     # New Code #
     parser.add_argument(

examples/config_yaml_templates/deepspeed.yaml

@@ -8,7 +8,7 @@ deepspeed_config:
   # `transformers` uses the right `init` function
   zero3_init_flag: false # true 
 
-# Finally we need to specify the number of GPUs to use
+# Finally we need to specify the number of accelerators to use
 num_processes: 2
 # Optionally we can set the mixed precision now instead of in the deepspeed config file,
 # however this requires the `fp16` and `bf16` options to be set to `auto` in the deepspeed config file

examples/config_yaml_templates/fp8.yaml

@@ -11,8 +11,8 @@ fp8_config:
   fp8_format: E4M3
   interval: 1
   margin: 0
-  override_linear_precision: (false, false, false)
+  override_linear_precision: [false, false, false]
   # Generally this should always be set to `false` to have the most realistic fp8 eval performance
   use_autocast_during_eval: false
   # If using MS-AMP, we ignore all of the prior and set a opt_level
-  #opt_level: O1
+  #opt_level: O1

examples/config_yaml_templates/fsdp.yaml

@@ -1,8 +1,8 @@
-# Since we are doing FSDP (even though it's multi-GPU), we need to specify the distributed type as FSDP
+# Since we are doing FSDP (even though it's multi-accelerator), we need to specify the distributed type as FSDP
 distributed_type: FSDP
 # Can be one of "no", "fp16", or "bf16" (see `transformer_engine.yaml` for `fp8`, but it works for FSDP as well)
 mixed_precision: 'bf16'
-# Specify the number of GPUs to use
+# Specify the number of accelerators to use
 num_processes: 2
 # Then we can specify the FSDP config
 fsdp_config:

examples/config_yaml_templates/multi_xpu.yaml

@@ -0,0 +1,6 @@
+# Specify distributed_type as `MULTI_XPU` for DDP
+distributed_type: "MULTI_XPU"
+# Can be one of "no", "fp16", or "bf16" (see `transformer_engine.yaml` for `fp8`)
+mixed_precision: "bf16"
+# Specify the number of XPUs to use
+num_processes: 2

examples/config_yaml_templates/single_accelerator.yaml

@@ -1,4 +1,4 @@
-# Since this is single GPU, we don't need distributed training
+# Since this is single GPU/XPU, we don't need distributed training
 distributed_type: "NO"
 # Can be one of "no", "fp16", or "bf16" (see `transformer_engine.yaml` for `fp8`)
-mixed_precision: "bf16"
+mixed_precision: "bf16"

examples/inference/pippy/bert.py

@@ -21,10 +21,7 @@ from accelerate.test_utils import torch_device
 from accelerate.utils import set_seed
 
 
-if torch_device == "hpu":
-    synchronize_func = torch.hpu.synchronize
-else:
-    synchronize_func = torch.cuda.synchronize
+synchronize_func = getattr(torch, torch_device, torch.cuda).synchronize
 
 # Set the random seed to have reproducable outputs
 set_seed(42)

examples/inference/pippy/gpt2.py

@@ -21,11 +21,7 @@ from accelerate.test_utils import torch_device
 from accelerate.utils import set_seed
 
 
-if torch_device == "hpu":
-    synchronize_func = torch.hpu.synchronize
-else:
-    synchronize_func = torch.cuda.synchronize
-
+synchronize_func = getattr(torch, torch_device, torch.cuda).synchronize
 
 # Set the random seed to have reproducable outputs
 set_seed(42)

examples/nlp_example.py

@@ -177,6 +177,7 @@ def training_function(config, args):
                 outputs = model(**batch)
             predictions = outputs.logits.argmax(dim=-1)
             predictions, references = accelerator.gather_for_metrics((predictions, batch["labels"]))
+            print(f"=====  {predictions}")
             metric.add_batch(
                 predictions=predictions,
                 references=references,

examples/requirements.txt

@@ -1,5 +1,5 @@
 accelerate # used to be installed in Amazon SageMaker environment
 evaluate
-datasets==2.3.2
+datasets
 schedulefree
 huggingface_hub>=0.20.0

examples/slurm/submit_multicpu.sh

@@ -8,7 +8,7 @@
 #SBATCH --error=E-%x.%j
 
 ######################
-### Set enviroment ###
+### Set environment ###
 ######################
 source activateEnvironment.sh
 

examples/slurm/submit_multigpu.sh

@@ -11,7 +11,7 @@
 #SBATCH --time=01:59:00             # maximum execution time (HH:MM:SS)
 
 ######################
-### Set enviroment ###
+### Set environment ###
 ######################
 source activateEnvironment.sh
 export GPUS_PER_NODE=4

examples/slurm/submit_multinode.sh

@@ -11,7 +11,7 @@
 #SBATCH --time=01:59:00             # maximum execution time (HH:MM:SS)
 
 ######################
-### Set enviroment ###
+### Set environment ###
 ######################
 source activateEnvironment.sh
 export GPUS_PER_NODE=4

examples/slurm/submit_multinode_fsdp.sh

@@ -11,7 +11,7 @@
 #SBATCH --time=01:59:00             # maximum execution time (HH:MM:SS)
 
 ######################
-### Set enviroment ###
+### Set environment ###
 ######################
 source activateEnvironment.sh
 export GPUS_PER_NODE=4
@@ -25,7 +25,7 @@ head_node_ip=$(scontrol show hostnames $SLURM_JOB_NODELIST | head -n 1)
 export ACCELERATE_DIR="${ACCELERATE_DIR:-/accelerate}"
 
 export LAUNCHER="accelerate launch \
-    --config ${ACCELERATE_DIR}/examples/slurm/fsdp_config.yaml \
+    --config_file ${ACCELERATE_DIR}/examples/slurm/fsdp_config.yaml \
     --num_processes $((SLURM_NNODES * GPUS_PER_NODE)) \
     --num_machines $SLURM_NNODES \
     --rdzv_backend c10d \

examples/torch_native_parallelism/README.md

@@ -0,0 +1,76 @@
+## Torch Native Parallelism
+
+With recent versions of Torch, there have been steady improvements in native parallelism using `DeviceMesh` and `DTensor`. 🤗 accelerate allows you to use these with our `ParallelismConfig` abstraction and/or `FullyShardedDataParallelPlugin(fsdp_version=2)`
+This folder contains various examples of such use-cases: such as composing multiple parallelism strategies, low-bit training etc.
+
+### ND Parallelism
+
+With `ParallelismConfig`, you can use 🤗 accelerate to train models with n-dimensional parallelism. This builds on top of 🤗 transformers, which we utilize for tensor parallelism sharding.
+Accelerate then takes care of everything else, such as data parallelism, FSDP or context parallelism.
+Script `nd_parallel.py` showcases this. We enable you to configure 4 different parallel dimensions (for now 👀):
+- dp_replicate_size: how many replicas of the model to create, each replica is trained on a different subset of the data and averaged at the end of each step, same as DDP in Torch
+- dp_shard_size: across how many devices is the model sharded, this is utilizing FSDP2 to shard the model across devices, so each device has a different part of the model
+- tp_size: how many devices to use for tensor parallelism, this is utilizing the tensor parallelism from 🤗 transformers
+- cp_size: how many devices to use for context parallelism, this will also shard the model, optimizer and gradients using `FSDP2` across
+the same group of devices, to further optimize memory usage (this comes with no slowdown)
+
+For example, with 8 nodes, you can run the script as such:
+```bash
+accelerate launch --num-processes 8 nd_parallel.py \
+    --dp-replicate-size 2 \
+    --dp-shard-size 2 \
+    --tp-size 2
+```
+
+> [!Tip]
+> Only use TP intra-node - therefore max TP size you should need is 8. You can also use a lower size, as FSDP (`--dp-shard-size`) can be faster on smaller models with shorter sequence lengths. If you cannot fit your model into memory, utilize `--dp-shard-size` as much as you can. Afterwards, to scale up and utilize all your resources, use `--dp-replicate-size`. This is only a general guideline, you can (and should) experiment with different parallelism configurations to find the best one for your model and hardware. You can learn more about the general strategies for parallelism in our [blog](https://huggingface.co/blog/accelerate-nd-parallel), or if you really want to dive deep, read the [Ultra-Scale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook).
+
+
+This feature is also fully integrated into 🤗 transformers `Trainer`. To use it, simply launch your script with path to your accelerate configuration file. You can see a minimal example of such script in `nd_parallel_trainer.py`.
+We provide 2 pre-configured configuration files:
+
+#### HSDP + TP (3D parallelism)
+
+```bash
+accelerate launch --config-file configs/tp_hsdp.yaml nd_parallel_trainer.py
+```
+
+#### Context parallelism (128k sequence length)
+
+```bash
+accelerate launch --config-file configs/cp.yaml nd_parallel_trainer.py --sequence-length=128000
+```
+
+  ### FSDP2 + ao Float8Linear
+
+In file `fsdp2_fp8.py` we use `Float8Linear` from `ao` to train a model partially in FP8 precision. We utilize `AORecipeKwargs` to pass the `Float8LinearConfig` to the accelerator, 
+which replaces the default `torch.nn.Linear` with `Float8Linear`. We also utilize `TorchDynamoPlugin` together with regional compilation to compile the model,
+gaining even more speed and memory savings, as `ao` doesn't ship with any kernels by default, so we have to gain the performance from compiling the model.
+
+Replacing linear layers with `Float8Linear` can greatly improve performance, if used correctly and on hardware that supports FP8 tensor cores. This highly depends on the model dimensions and sequence length used for training.
+You can view the performance of `Float8Linear` as a function of matrix dimensions in [this document](https://github.com/pytorch/ao/blob/main/torchao/float8/README.md#performance). 
+
+In our example, we use a 8B Llama3.1 model, which has a hidden dimension of 4096 and we train on sequence length of 8192. In the below images, we can see that this improves performance by ~25% compared to `bf16`, reaching ~10000 tokens per second, per device on 8x H100 GPUs, compared to ~8000 tokens per second using `bf16`, while loss function stays roughly the same. We can also see that the FLOPS rise by using FP8.
+
+<div style="display: flex; gap: 25px;">
+  <div style="text-align: center; width: 49%;">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/examples/fsdp2/fp8_tps.png" alt="tps" style="width: 100%;">
+    <p style="text-align: center; margin-top: 8px;">TPS per device, BF16 vs FP8</p>
+  </div>
+  <div style="text-align: center; width: 49%;">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/examples/fsdp2/fp8_tflops.png" alt="tflops" style="width: 100%;">
+    <p style="text-align: center; margin-top: 8px;">TFLOPS per device, BF16 vs FP8. We cannot really compare MFU as FP8 tensor cores are used as well.</p>
+  </div>
+  
+  <div style="text-align: center; width: 49%;">  
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/examples/fsdp2/fp8_loss.png" alt="loss" style="width: 100%; max-width: 900px;">
+    <p style="text-align: center; margin-top: 8px;">Loss curve, BF16 vs FP8, it's hard to see the difference as the curves mostly overlap</p>
+  </div>
+</div>
+
+The figures above were generated on 8x H100 SXM GPUs, with 8192 sequence length and 1000 steps. To run the example, you can use the following command, where you can specify the precision to train in:
+
+```bash
+accelerate launch fsdp2_fp8.py --sequence-length 8192 --num-steps 1000 --log_with wandb --precision [fp8 | bf16]
+```
+

examples/torch_native_parallelism/configs/cp.yaml

@@ -0,0 +1,29 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_activation_checkpointing: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_cpu_ram_efficient_loading: false
+  fsdp_offload_params: false
+  fsdp_reshard_after_forward: true
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_version: 2
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+parallelism_config:
+  parallelism_config_cp_size: 8
+  parallelism_config_dp_replicate_size: 1
+  parallelism_config_dp_shard_size: 1
+  parallelism_config_tp_size: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/torch_native_parallelism/configs/tp_hsdp.yaml

@@ -0,0 +1,29 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: FSDP
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+fsdp_config:
+  fsdp_activation_checkpointing: false
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_cpu_ram_efficient_loading: false
+  fsdp_offload_params: false
+  fsdp_reshard_after_forward: true
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_version: 2
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 8
+parallelism_config:
+  parallelism_config_cp_size: 1
+  parallelism_config_dp_replicate_size: 2
+  parallelism_config_dp_shard_size: 2
+  parallelism_config_tp_size: 2
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/torch_native_parallelism/fsdp2_fp8.py

@@ -0,0 +1,139 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Minimal example of training with FP8 precision using FSDP2 via Accelerate.
+This example demonstrates how to use torchao's Float8LinearConfig with Accelerate's AORecipeKwargs.
+"""
+
+import argparse
+
+import torch
+from torch.utils.data import DataLoader
+from torchao.float8 import Float8LinearConfig
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+
+from accelerate import Accelerator
+from accelerate.utils import AORecipeKwargs, FullyShardedDataParallelPlugin, TorchDynamoPlugin, set_seed
+from utils import PerformanceTracker, create_collate_fn, get_dataset, get_model_flops_per_token
+
+
+WARMUP_STEPS = 10
+
+MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--sequence-length", type=int, default=8192, help="Sequence length for the dataset")
+    parser.add_argument("--num-steps", type=int, default=1000, help="Number of steps to train for")
+    parser.add_argument("--precision", type=str, default="fp8", choices=["fp8", "bf16"], help="Precision to train in")
+    parser.add_argument("--log-with", type=str, default="wandb", help="Log with wandb or tensorboard")
+
+    return parser.parse_args()
+
+
+def main():
+    """
+    Main function to train the model.
+    """
+    set_seed(42)
+
+    args = parse_args()
+
+    fsdp2_plugin = FullyShardedDataParallelPlugin(
+        fsdp_version=2,
+        cpu_ram_efficient_loading=False,  # CPU RAM efficient loading CANNOT work with fp8 torchao
+        auto_wrap_policy="transformer_based_wrap",
+        transformer_cls_names_to_wrap=["LlamaDecoderLayer"],
+    )
+    fsdp2_plugin.set_mixed_precision(args.precision)
+
+    dynamo_plugin = TorchDynamoPlugin(
+        backend="inductor",
+        use_regional_compilation=True,  # We use regional compilation to compile the model way faster
+    )
+
+    fp8_config = Float8LinearConfig(
+        enable_fsdp_float8_all_gather=True,  # extra saving by gathering parameters in fp8 and upcasting after
+    )
+
+    kwargs = []
+    if args.precision == "fp8":
+        kwargs = [AORecipeKwargs(config=fp8_config)]
+
+    accelerator = Accelerator(
+        fsdp_plugin=fsdp2_plugin,
+        dynamo_plugin=dynamo_plugin,
+        kwargs_handlers=kwargs,
+        log_with=args.log_with,
+    )
+    accelerator.init_trackers(
+        project_name="FSDP2_torchao_fp8",
+        config={"sequence_length": args.sequence_length, "num_steps": args.num_steps},
+    )
+
+    model = AutoModelForCausalLM.from_config(
+        AutoConfig.from_pretrained(MODEL_ID, use_cache=False),
+        torch_dtype=torch.bfloat16,
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    optimizer = torch.optim.AdamW(model.parameters(), lr=1e-5)
+    dataset = get_dataset(tokenizer, args.sequence_length, accelerator)
+    dataloader = DataLoader(dataset, batch_size=1, collate_fn=create_collate_fn())
+
+    model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
+    accelerator.wait_for_everyone()
+
+    model.train()
+
+    total_num_steps = min(args.num_steps, len(dataloader))
+    model_flops_per_token = get_model_flops_per_token(model, args.sequence_length)
+    performance_tracker = PerformanceTracker(warmup_steps=5)
+
+    for step, batch in enumerate(dataloader):
+        if step >= total_num_steps:
+            break
+
+        outputs = model(**batch)
+        loss = outputs.loss
+
+        accelerator.backward(loss)
+        optimizer.step()
+        optimizer.zero_grad()
+        metrics = performance_tracker.step(batch["input_ids"].shape[1], model_flops_per_token)
+
+        print_msg = f"Step {step}/{total_num_steps}, Loss: {loss.item():.4f}"
+        if "warmup_completed" in metrics:
+            accelerator.print("Warm up completed! Starting training")
+        elif metrics:
+            print_msg += performance_tracker.get_print_message(metrics)
+
+        if step % 10 == 0 or step == total_num_steps - 1:
+            accelerator.print(print_msg)
+
+        accelerator.log(metrics)
+
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+    accelerator.print("Training completed!")
+
+
+if __name__ == "__main__":
+    main()

examples/torch_native_parallelism/nd_parallel.py

@@ -0,0 +1,168 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Example of training with ND parallel using accelerate's ParallelismConfig
+"""
+
+import argparse
+import warnings
+
+import torch
+import torch.distributed as dist
+from torch.utils.data import DataLoader
+from transformers import AutoModelForCausalLM
+
+from accelerate import Accelerator
+from accelerate.parallelism_config import ParallelismConfig
+from accelerate.utils import FullyShardedDataParallelPlugin, set_seed
+from utils import (
+    PerformanceTracker,
+    create_collate_fn,
+    get_dataset,
+    setup_tokenizer,
+)
+
+
+MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--dp-replicate-size", type=int, default=1)
+    parser.add_argument("--dp-shard-size", type=int, default=1)
+    parser.add_argument("--tp-size", type=int, default=1)
+    parser.add_argument("--cp-size", type=int, default=1)
+    parser.add_argument("--sequence-length", type=int, default=1024)
+    parser.add_argument("--num-steps", type=int, default=1000)
+    parser.add_argument("--save-dir", type=str, default="./outputs")
+    parser.add_argument("--checkpoint-frequency", type=int, default=100)
+    parser.add_argument("--model-name", type=str, default=MODEL_ID)
+
+    return parser.parse_args()
+
+
+def forward(model, batch, optimizer, accelerator: Accelerator):
+    batch["position_ids"] = torch.arange(0, batch["input_ids"].size(1), device=batch["input_ids"].device).unsqueeze(0)
+    # We need both labels and shift_labels, as the loss computation in the model is hidden behind `if labels is not None`, but the loss computation
+    # itself prioritzes shift_labels (if provided) which are the correct ones (due to labels being wrong if cp enabled)
+    buffers = [batch["input_ids"], batch["shift_labels"], batch["labels"], batch["position_ids"]]
+    with accelerator.maybe_context_parallel(
+        buffers=buffers, buffer_seq_dims=[1, 1, 1, 1], no_restore_buffers=set(buffers)
+    ):
+        # To get the proper loss value, we need to average across devices that are participating in data parallel/context parallel training
+        # As for DP we have a different batch on each device and for CP we essentially have a different part of sequences on each device
+        # I.e. with causal modelling and seq_len 1024, this dimension becomes another batch dimension of sorts
+        loss_reduce_grp = (
+            accelerator.torch_device_mesh["dp_cp"].get_group()
+            if accelerator.parallelism_config.dp_cp_dim_names
+            else None
+        )
+        outputs = model(**batch)
+        loss = outputs.loss
+        accelerator.backward(loss)
+        optimizer.step()
+        optimizer.zero_grad()
+        dist.all_reduce(loss, op=dist.ReduceOp.AVG, group=loss_reduce_grp)
+
+    return loss
+
+
+def train(args):
+    parallelism_config = ParallelismConfig(
+        dp_replicate_size=args.dp_replicate_size,
+        dp_shard_size=args.dp_shard_size,
+        tp_size=args.tp_size,
+        cp_size=args.cp_size,
+    )
+
+    # FSDP needs extra configuration, so we properly shard the model
+    fsdp2_plugin = None
+    if parallelism_config.dp_shard_enabled or parallelism_config.cp_enabled:
+        fsdp2_plugin = FullyShardedDataParallelPlugin(
+            fsdp_version=2,
+            auto_wrap_policy="transformer_based_wrap",
+            transformer_cls_names_to_wrap=["LlamaDecoderLayer"],
+            state_dict_type="SHARDED_STATE_DICT",
+        )
+
+    accelerator = Accelerator(
+        log_with=["wandb"], mixed_precision="bf16", parallelism_config=parallelism_config, fsdp_plugin=fsdp2_plugin
+    )
+    accelerator.init_trackers("nd_parallel_training")
+
+    # If TP was enabled, we need to tell transformers to prepare the model for us
+    model_kwargs = (
+        {"tp_size": args.tp_size, "tp_plan": "auto", "device_mesh": accelerator.torch_device_mesh}
+        if args.tp_size > 1
+        else {}
+    )
+    model = AutoModelForCausalLM.from_pretrained(
+        args.model_name,
+        torch_dtype=torch.bfloat16,
+        use_cache=False,
+        **model_kwargs,
+    )
+    tokenizer = setup_tokenizer(args.model_name)
+    optimizer = torch.optim.SGD(model.parameters(), lr=1e-5)
+    dataset = get_dataset(tokenizer, args.sequence_length, accelerator)
+    dataloader = DataLoader(dataset, batch_size=1, collate_fn=create_collate_fn())
+
+    model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
+
+    total_num_steps = min(args.num_steps, len(dataloader))
+    performance_tracker = PerformanceTracker(warmup_steps=5)
+
+    accelerator.print("Starting training...")
+    for step, batch in enumerate(dataloader):
+        if step >= total_num_steps:
+            break
+
+        loss = forward(model, batch, optimizer, accelerator)
+
+        # We report TPS per device, so we divide by the number of devices in the non-data parallel dimension
+        metrics = performance_tracker.step(batch["input_ids"].shape[1] / parallelism_config.non_data_parallel_size)
+
+        print_msg = f"Step {step}/{total_num_steps}, Loss: {loss.item():.4f}"
+        if "warmup_completed" in metrics:
+            accelerator.print("Warm up completed! Starting performance tracking...")
+        elif metrics:
+            print_msg += performance_tracker.get_print_message(metrics, with_memory=True)
+
+        if step % 10 == 0 or step == total_num_steps - 1:
+            accelerator.print(print_msg)
+
+        if step % args.checkpoint_frequency == 0 and step > 0 and parallelism_config.dp_shard_enabled:
+            accelerator.print(f"Saving checkpoint at step {step}...")
+            accelerator.save_state(args.save_dir + f"/checkpoint-{step}")
+
+        accelerator.log({"loss": loss.item()})
+
+    accelerator.print("Training completed!")
+
+    model.save_pretrained(args.save_dir + f"/{args.model_name}")
+    accelerator.print(f"Model saved to {args.save_dir}/{args.model_name}")
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    set_seed(42)
+    args = parse_args()
+    if args.dp_shard_size == 1 and args.tp_size > 1:
+        # We currently don't support saving with `save_state` when using only
+        # tensor parallelism, fsdp must be enabled
+        warnings.warn(
+            "Accelerator.save_state() is not yet supported with pure tensor parallel training. Training will work, but intermediate checkpoints will not be saved."
+        )
+    train(args)

examples/torch_native_parallelism/nd_parallel_trainer.py

@@ -0,0 +1,82 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
+
+from accelerate.utils import ParallelismConfig
+from utils import get_dataset
+
+
+MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--sequence-length", type=int, default=4096)
+    parser.add_argument("--checkpoint-frequency", type=int, default=100)
+    parser.add_argument("--model-name", type=str, default=MODEL_ID)
+    parser.add_argument("--save-dir", type=str, default=f"./accelerate-nd-parallel-{MODEL_ID.split('/')[-1]}")
+    parser.add_argument("--device-type", type=str, default="auto")
+    return parser.parse_args()
+
+
+def main():
+    # If ParallelismConfig is not initialized with __init__, it reads from env vars
+    # which were set by using config
+    pc = ParallelismConfig()
+    args = parse_args()
+
+    if args.device_type == "auto":
+        args.device_type = torch.accelerator.current_accelerator().type
+
+    model_kwargs = {}
+    if pc.tp_enabled:
+        model_kwargs["tp_plan"] = "auto"
+        model_kwargs["device_mesh"] = pc.build_device_mesh(args.device_type)
+
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
+    model = AutoModelForCausalLM.from_pretrained(args.model_name, use_cache=False, **model_kwargs)
+
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    packed_dataset = get_dataset(tokenizer, args.sequence_length)
+
+    training_args = TrainingArguments(
+        output_dir=args.save_dir,
+        parallelism_config=pc,
+        num_train_epochs=1,
+        per_device_train_batch_size=1,
+        logging_steps=5,
+        save_steps=args.checkpoint_frequency,
+        learning_rate=5e-5,
+        remove_unused_columns=False,
+        bf16=True,
+    )
+
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        processing_class=tokenizer,
+        train_dataset=packed_dataset,
+    )
+
+    trainer.train()
+    trainer.save_model()
+
+
+if __name__ == "__main__":
+    main()

examples/torch_native_parallelism/utils.py

@@ -0,0 +1,219 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Common utilities for torch-native-parallelism examples.
+"""
+
+import time
+from contextlib import nullcontext
+
+import torch
+from datasets import Dataset, load_dataset
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from accelerate import Accelerator
+
+
+def get_dataset(tokenizer: AutoTokenizer, seq_len: int, accelerator: Accelerator | None = None) -> Dataset:
+    """
+    Load and prepare TinyStories dataset.
+
+    Args:
+        accelerator (Accelerator): Accelerate accelerator instance
+        tokenizer (AutoTokenizer): Hugging Face tokenizer
+        seq_len (int): Sequence length for the dataset
+
+    Returns:
+        Dataset: Packed dataset
+    """
+    processing_ctx = accelerator.main_process_first if accelerator else nullcontext
+    raw_dataset = load_dataset("roneneldan/TinyStories", split="train[:50%]")
+
+    def tokenize_function(examples):
+        tokenized_batch = tokenizer(
+            examples["text"],
+            padding=False,
+            truncation=True,
+            max_length=seq_len,
+            return_tensors=None,
+        )
+        tokenized_batch["labels"] = tokenized_batch["input_ids"].copy()
+        return tokenized_batch
+
+    with processing_ctx():
+        tokenized_dataset = raw_dataset.map(tokenize_function, batched=True, remove_columns=["text"])
+
+    def create_packed_sequences(examples):
+        all_tokens = []
+        for input_ids in examples["input_ids"]:
+            all_tokens.extend(input_ids)
+
+        num_sequences = len(all_tokens) // (seq_len + 1)
+        packed_input_ids = []
+        packed_labels = []
+        packed_position_ids = []
+
+        for i in range(num_sequences):
+            start_idx = i * (seq_len + 1)
+            end_idx = start_idx + (seq_len + 1)
+            full_sequence = all_tokens[start_idx:end_idx]
+            packed_input_ids.append(full_sequence[:-1])
+            packed_labels.append(full_sequence[1:])
+            packed_position_ids.append(torch.arange(0, seq_len))
+
+        return {
+            "input_ids": packed_input_ids,
+            "shift_labels": packed_labels,
+            "position_ids": packed_position_ids,
+            "labels": packed_labels,
+        }
+
+    with processing_ctx():
+        packed_dataset = tokenized_dataset.map(
+            create_packed_sequences,
+            batched=True,
+            remove_columns=tokenized_dataset.column_names,
+            batch_size=1000,
+        )
+
+    return packed_dataset.shuffle(seed=42)
+
+
+def get_model_flops_per_token(model: AutoModelForCausalLM, seq_len: int) -> float:
+    """
+    Get the number of flops per token for the model.
+
+    Args:
+        model (AutoModelForCausalLM): Model to get the flops for
+        seq_len (int): Sequence length
+    """
+    cfg = model.config
+    head_dim = cfg.hidden_size // cfg.num_attention_heads
+
+    # MLP: 3 matmuls
+    mlp_flops = 18 * cfg.hidden_size * cfg.intermediate_size
+
+    # Attn (w/o dotproduct)
+    attn_flops = 12 * head_dim * (cfg.num_attention_heads + cfg.num_key_value_heads)
+
+    # attn (dotproduct) - this scales quadratically with sequence length
+    attn_dotproduct_flops = 12 * cfg.num_attention_heads * head_dim * seq_len
+
+    # we also ignore embeddings and layernorms, etc
+    return (mlp_flops + attn_flops + attn_dotproduct_flops) * cfg.num_hidden_layers
+
+
+def create_collate_fn():
+    """Create a collate function for batching."""
+
+    def collate_fn(batch):
+        input_ids = torch.tensor([item["input_ids"] for item in batch], dtype=torch.long)
+        shift_labels = torch.tensor([item["shift_labels"] for item in batch], dtype=torch.long)
+        return {"input_ids": input_ids, "shift_labels": shift_labels, "labels": shift_labels}
+
+    return collate_fn
+
+
+class PerformanceTracker:
+    """Track training performance metrics."""
+
+    def __init__(self, warmup_steps: int = 10):
+        self.warmup_steps = warmup_steps
+        self.reset()
+
+    def reset(self):
+        """Reset all tracking variables."""
+        self.start_time = None
+        self.num_tokens = 0
+        self.is_in_warmup = True
+        self.step_count = 0
+
+    def step(self, batch_tokens: int, model_flops_per_token: float | None = None) -> dict:
+        """
+        Update performance tracking with a new step.
+
+        Args:
+            batch_tokens (int): Number of tokens in current batch
+
+        Returns:
+            dict: Performance metrics if past warmup, empty dict otherwise
+        """
+        self.step_count += 1
+
+        if self.step_count == self.warmup_steps:
+            self.start_time = time.perf_counter()
+            self.num_tokens = 0
+            self.is_in_warmup = False
+            return {"warmup_completed": True}
+
+        if not self.is_in_warmup and self.start_time is not None:
+            dct = {}
+            self.num_tokens += batch_tokens
+            total_time = time.perf_counter() - self.start_time
+            steps_from_warmup = self.step_count - self.warmup_steps
+
+            if total_time > 0 and steps_from_warmup > 0:
+                memory_stats = gpu_memory_usage_all()
+                dct = {
+                    "tokens_per_second": self.num_tokens / total_time,
+                    "steps_per_second": steps_from_warmup / total_time,
+                    "total_tokens": self.num_tokens,
+                    "total_time": total_time,
+                    **memory_stats,
+                }
+
+            if model_flops_per_token is not None:
+                flops = model_flops_per_token * self.num_tokens
+                dct["tflops_per_device"] = flops / (total_time * 1e12)
+
+            return dct
+
+        return {}
+
+    def get_print_message(self, metrics: dict, with_memory: bool = False) -> str:
+        print_msg = f" | Average steps/s: {metrics['steps_per_second']:.2f} | Average tokens/s: {metrics['tokens_per_second']:.2f} | Average TFLOPS: {metrics['tflops_per_device']:.2f}\n"
+        if with_memory:
+            print_msg += (
+                f"\tMemory (GB): active={metrics['peak_memory_active']:.1f}, "
+                f"alloc={metrics['peak_memory_alloc']:.1f}, "
+                f"reserved={metrics['peak_memory_reserved']:.1f}"
+            )
+        return print_msg
+
+
+def setup_tokenizer(model_id: str) -> AutoTokenizer:
+    """Setup tokenizer with proper padding token."""
+    tokenizer = AutoTokenizer.from_pretrained(model_id)
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    return tokenizer
+
+
+def gpu_memory_usage_all(device=0):
+    device_type = torch.accelerator.current_accelerator().type
+    device = torch.device(f"{device_type}:{device}")
+    torch_device_module = getattr(torch, device_type, torch.cuda)
+    _BYTES_IN_GIB = 1024**3
+    peak_memory_active = torch_device_module.memory_stats().get("active_bytes.all.peak", 0) / _BYTES_IN_GIB
+    peak_memory_alloc = torch_device_module.max_memory_allocated(device) / _BYTES_IN_GIB
+    peak_memory_reserved = torch_device_module.max_memory_reserved(device) / _BYTES_IN_GIB
+    memory_stats = {
+        "peak_memory_active": peak_memory_active,
+        "peak_memory_alloc": peak_memory_alloc,
+        "peak_memory_reserved": peak_memory_reserved,
+    }
+    torch_device_module.reset_peak_memory_stats(device)
+
+    return memory_stats

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.ruff]
 line-length = 119
-target-version = "py38"
+target-version = "py310"
 
 [tool.ruff.lint]
 preview = true
@@ -20,6 +20,9 @@ ignore = [
     "E741", # Ambiguous variable name
     "W605", # Invalid escape sequence
     "UP007", # X | Y type annotations
+    "UP045", # Use `X | None` for type annotations
+    "UP035", # temporarily disabled to minimize upgrade changes
+
 ]
 
 [tool.ruff.lint.per-file-ignores]

setup.py

@@ -16,13 +16,10 @@ from setuptools import find_packages, setup
 
 
 extras = {}
-extras["quality"] = [
-    "black ~= 23.1",  # hf-doc-builder has a hidden dependency on `black`
-    "hf-doc-builder >= 0.3.0",
-    "ruff ~= 0.6.4",
-]
+extras["quality"] = ["ruff == 0.13.1"]
+
 extras["docs"] = []
-extras["test_prod"] = ["pytest>=7.2.0,<=8.0.0", "pytest-xdist", "pytest-subtests", "parameterized", "pytest-order"]
+extras["test_prod"] = ["pytest>=7.2.0", "pytest-xdist", "pytest-subtests", "parameterized", "pytest-order"]
 extras["test_dev"] = [
     "datasets",
     "diffusers",
@@ -40,7 +37,17 @@ extras["testing"] = extras["test_prod"] + extras["test_dev"]
 extras["deepspeed"] = ["deepspeed"]
 extras["rich"] = ["rich"]
 
-extras["test_trackers"] = ["wandb", "comet-ml", "tensorboard", "dvclive"]
+extras["test_fp8"] = ["torchao"]  # note: TE for now needs to be done via pulling down the docker image directly
+extras["test_trackers"] = [
+    "wandb",
+    "comet-ml",
+    "tensorboard",
+    "dvclive",
+    # "mlflow", too many deps that lead to download a very old version of the lib
+    "matplotlib",
+    "swanlab[dashboard]",  # dashboard required for local use
+    "trackio",
+]
 extras["dev"] = extras["quality"] + extras["testing"] + extras["rich"]
 
 extras["sagemaker"] = [
@@ -49,7 +56,7 @@ extras["sagemaker"] = [
 
 setup(
     name="accelerate",
-    version="1.4.0.dev0",
+    version="1.11.0.dev0",
     description="Accelerate",
     long_description=open("README.md", encoding="utf-8").read(),
     long_description_content_type="text/markdown",
@@ -69,9 +76,9 @@ setup(
             "accelerate-merge-weights=accelerate.commands.merge:main",
         ]
     },
-    python_requires=">=3.9.0",
+    python_requires=">=3.10.0",
     install_requires=[
-        "numpy>=1.17,<3.0.0",
+        "numpy>=1.17",
         "packaging>=20.0",
         "psutil",
         "pyyaml",
@@ -88,7 +95,7 @@ setup(
         "License :: OSI Approved :: Apache Software License",
         "Operating System :: OS Independent",
         "Programming Language :: Python :: 3",
-        "Programming Language :: Python :: 3.8",
+        "Programming Language :: Python :: 3.10",
         "Topic :: Scientific/Engineering :: Artificial Intelligence",
     ],
 )

src/accelerate/__init__.py

@@ -11,7 +11,7 @@
 # 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.
-__version__ = "1.4.0.dev0"
+__version__ = "1.11.0.dev0"
 
 from .accelerator import Accelerator
 from .big_modeling import (
@@ -26,6 +26,7 @@ from .big_modeling import (
 from .data_loader import skip_first_batches
 from .inference import prepare_pippy
 from .launchers import debug_launcher, notebook_launcher
+from .parallelism_config import ParallelismConfig
 from .state import PartialState
 from .utils import (
     AutocastKwargs,

src/accelerate/big_modeling.py

@@ -14,9 +14,10 @@
 
 import logging
 import os
+import re
 from contextlib import contextmanager
 from functools import wraps
-from typing import Dict, List, Optional, Union
+from typing import Optional, Union
 
 import torch
 import torch.nn as nn
@@ -24,6 +25,7 @@ import torch.nn as nn
 from .hooks import (
     AlignDevicesHook,
     CpuOffload,
+    LayerwiseCastingHook,
     UserCpuOffloadHook,
     add_hook_to_module,
     attach_align_device_hook,
@@ -48,6 +50,7 @@ from .utils import (
     parse_flag_from_env,
     retie_parameters,
 )
+from .utils.constants import SUPPORTED_PYTORCH_LAYERS_FOR_UPCASTING
 from .utils.other import recursive_getattr
 
 
@@ -55,7 +58,7 @@ logger = logging.getLogger(__name__)
 
 
 @contextmanager
-def init_empty_weights(include_buffers: bool = None):
+def init_empty_weights(include_buffers: Optional[bool] = None):
     """
     A context manager under which models are initialized with all parameters on the meta device, therefore creating an
     empty model. Useful when just initializing the model would blow the available RAM.
@@ -91,7 +94,7 @@ def init_empty_weights(include_buffers: bool = None):
 
 
 @contextmanager
-def init_on_device(device: torch.device, include_buffers: bool = None):
+def init_on_device(device: torch.device, include_buffers: Optional[bool] = None):
     """
     A context manager under which models are initialized with all parameters on the specified device.
 
@@ -171,8 +174,8 @@ def cpu_offload(
     model: nn.Module,
     execution_device: Optional[torch.device] = None,
     offload_buffers: bool = False,
-    state_dict: Optional[Dict[str, torch.Tensor]] = None,
-    preload_module_classes: Optional[List[str]] = None,
+    state_dict: Optional[dict[str, torch.Tensor]] = None,
+    preload_module_classes: Optional[list[str]] = None,
 ):
     """
     Activates full CPU offload for a model. As a result, all parameters of the model will be offloaded and only one
@@ -262,7 +265,7 @@ def disk_offload(
     offload_dir: Union[str, os.PathLike],
     execution_device: Optional[torch.device] = None,
     offload_buffers: bool = False,
-    preload_module_classes: Optional[List[str]] = None,
+    preload_module_classes: Optional[list[str]] = None,
 ):
     """
     Activates full disk offload for a model. As a result, all parameters of the model will be offloaded as
@@ -305,14 +308,14 @@ def disk_offload(
 
 def dispatch_model(
     model: nn.Module,
-    device_map: Dict[str, Union[str, int, torch.device]],
+    device_map: dict[str, Union[str, int, torch.device]],
     main_device: Optional[torch.device] = None,
-    state_dict: Optional[Dict[str, torch.Tensor]] = None,
+    state_dict: Optional[dict[str, torch.Tensor]] = None,
     offload_dir: Optional[Union[str, os.PathLike]] = None,
-    offload_index: Optional[Dict[str, str]] = None,
+    offload_index: Optional[dict[str, str]] = None,
     offload_buffers: bool = False,
-    skip_keys: Optional[Union[str, List[str]]] = None,
-    preload_module_classes: Optional[List[str]] = None,
+    skip_keys: Optional[Union[str, list[str]]] = None,
+    preload_module_classes: Optional[list[str]] = None,
     force_hooks: bool = False,
 ):
     """
@@ -495,8 +498,6 @@ def dispatch_model(
             device = f"sdaa:{device}"
         elif is_musa_available() and isinstance(device, int):
             device = f"musa:{device}"
-        elif is_xpu_available() and isinstance(device, int):
-            device = f"xpu:{device}"
         if device != "disk":
             model.to(device)
         else:
@@ -511,17 +512,19 @@ def dispatch_model(
 def load_checkpoint_and_dispatch(
     model: nn.Module,
     checkpoint: Union[str, os.PathLike],
-    device_map: Optional[Union[str, Dict[str, Union[int, str, torch.device]]]] = None,
-    max_memory: Optional[Dict[Union[int, str], Union[int, str]]] = None,
-    no_split_module_classes: Optional[List[str]] = None,
+    device_map: Optional[Union[str, dict[str, Union[int, str, torch.device]]]] = None,
+    max_memory: Optional[dict[Union[int, str], Union[int, str]]] = None,
+    no_split_module_classes: Optional[list[str]] = None,
     offload_folder: Optional[Union[str, os.PathLike]] = None,
     offload_buffers: bool = False,
     dtype: Optional[Union[str, torch.dtype]] = None,
     offload_state_dict: Optional[bool] = None,
-    skip_keys: Optional[Union[str, List[str]]] = None,
-    preload_module_classes: Optional[List[str]] = None,
+    skip_keys: Optional[Union[str, list[str]]] = None,
+    preload_module_classes: Optional[list[str]] = None,
     force_hooks: bool = False,
     strict: bool = False,
+    full_state_dict: bool = True,
+    broadcast_from_rank0: bool = False,
 ):
     """
     Loads a (potentially sharded) checkpoint inside a model, potentially sending weights to a given device as they are
@@ -571,6 +574,12 @@ def load_checkpoint_and_dispatch(
         strict (`bool`, *optional*, defaults to `False`):
             Whether to strictly enforce that the keys in the checkpoint state_dict match the keys of the model's
             state_dict.
+        full_state_dict (`bool`, *optional*, defaults to `True`): if this is set to `True`, all the tensors in the
+            loaded state_dict will be gathered. No ShardedTensor and DTensor will be in the loaded state_dict.
+        broadcast_from_rank0 (`False`, *optional*, defaults to `False`): when the option is `True`, a distributed
+            `ProcessGroup` must be initialized. rank0 should receive a full state_dict and will broadcast the tensors
+            in the state_dict one by one to other ranks. Other ranks will receive the tensors and shard (if applicable)
+            according to the local shards in the model.
 
     Example:
 
@@ -596,8 +605,7 @@ def load_checkpoint_and_dispatch(
     """
     if isinstance(device_map, str) and device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]:
         raise ValueError(
-            "If passing a string for `device_map`, please choose 'auto', 'balanced', 'balanced_low_0' or "
-            "'sequential'."
+            "If passing a string for `device_map`, please choose 'auto', 'balanced', 'balanced_low_0' or 'sequential'."
         )
     if isinstance(device_map, str):
         if device_map != "sequential":
@@ -626,6 +634,8 @@ def load_checkpoint_and_dispatch(
         offload_state_dict=offload_state_dict,
         offload_buffers=offload_buffers,
         strict=strict,
+        full_state_dict=full_state_dict,
+        broadcast_from_rank0=broadcast_from_rank0,
     )
     if device_map is None:
         return model
@@ -638,3 +648,142 @@ def load_checkpoint_and_dispatch(
         preload_module_classes=preload_module_classes,
         force_hooks=force_hooks,
     )
+
+
+def attach_layerwise_casting_hooks(
+    module: torch.nn.Module,
+    storage_dtype: torch.dtype,
+    compute_dtype: torch.dtype,
+    skip_modules_pattern: Optional[Union[str, tuple[str, ...]]] = None,
+    skip_modules_classes: Optional[tuple[type[torch.nn.Module], ...]] = None,
+    non_blocking: bool = False,
+) -> None:
+    r"""
+    Applies layerwise casting to a given module. The module expected here is a PyTorch `nn.Module`. This is helpful for
+    reducing memory requirements when one doesn't want to fully quantize a model. Model params can be kept in say,
+    `torch.float8_e4m3fn` and upcasted to a higher precision like `torch.bfloat16` during forward pass and downcasted
+    back to `torch.float8_e4m3fn` to realize memory savings.
+
+    Args:
+        module (`torch.nn.Module`):
+            The module whose leaf modules will be cast to a high precision dtype for computation, and to a low
+            precision dtype for storage.
+        storage_dtype (`torch.dtype`):
+            The dtype to cast the module to before/after the forward pass for storage.
+        compute_dtype (`torch.dtype`):
+            The dtype to cast the module to during the forward pass for computation.
+        skip_modules_pattern (`tuple[str, ...]`, defaults to `None`):
+            A list of patterns to match the names of the modules to skip during the layerwise casting process. If set
+            to `None` alongside `skip_modules_classes` being `None`, the layerwise casting is applied directly to the
+            module instead of its internal submodules.
+        skip_modules_classes (`tuple[type[torch.nn.Module], ...]`, defaults to `None`):
+            A list of module classes to skip during the layerwise casting process.
+        non_blocking (`bool`, defaults to `False`):
+            If `True`, the weight casting operations are non-blocking.
+
+    Example:
+
+    ```python
+    >>> from accelerate.hooks import attach_layerwise_casting_hooks
+    >>> from transformers import AutoModelForCausalLM
+    >>> import torch
+
+    >>> # Model
+    >>> checkpoint = "EleutherAI/gpt-j-6B"
+    >>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
+
+    >>> # Attach hooks and perform inference
+    >>> attach_layerwise_casting_hooks(model, storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)
+    >>> with torch.no_grad():
+    ...     model(...)
+    ```
+
+    Users can also pass modules they want to avoid from getting downcasted.
+
+    ```py
+    >>> attach_layerwise_casting_hooks(
+    ...     model, storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16, skip_modules_pattern=["norm"]
+    ... )
+    ```
+    """
+    _attach_layerwise_casting_hooks(
+        module, storage_dtype, compute_dtype, skip_modules_pattern, skip_modules_classes, non_blocking
+    )
+
+
+def _attach_layerwise_casting_hooks(
+    module: torch.nn.Module,
+    storage_dtype: torch.dtype,
+    compute_dtype: torch.dtype,
+    skip_modules_pattern: Optional[Union[str, tuple[str, ...]]] = None,
+    skip_modules_classes: Optional[tuple[type[torch.nn.Module], ...]] = None,
+    non_blocking: bool = False,
+    _prefix: str = "",
+):
+    should_skip = (skip_modules_classes is not None and isinstance(module, skip_modules_classes)) or (
+        skip_modules_pattern is not None and any(re.search(pattern, _prefix) for pattern in skip_modules_pattern)
+    )
+    if should_skip:
+        logger.debug(f'Skipping layerwise casting for layer "{_prefix}"')
+        return
+
+    if isinstance(module, SUPPORTED_PYTORCH_LAYERS_FOR_UPCASTING):
+        logger.debug(f'Applying layerwise casting to layer "{_prefix}"')
+        add_hook_to_module(
+            module,
+            LayerwiseCastingHook(storage_dtype=storage_dtype, compute_dtype=compute_dtype, non_blocking=non_blocking),
+            append=True,
+        )
+        return
+
+    for name, submodule in module.named_children():
+        layer_name = f"{_prefix}.{name}" if _prefix else name
+        _attach_layerwise_casting_hooks(
+            submodule,
+            storage_dtype,
+            compute_dtype,
+            skip_modules_pattern,
+            skip_modules_classes,
+            non_blocking,
+            _prefix=layer_name,
+        )
+
+
+def _attach_context_parallel_hooks(
+    model: nn.Module,
+):
+    """
+    Monkeypatch huggingface's `transformers` model to fix attention mask issues when using context parallelism.
+
+    This function attaches forward_pre_hooks to each self_attn module of the model, where each hook checks the
+    args/kwargs, if they contain an attention mask, if it does, it will remove this mask, check if it is a causal mask,
+    if yes, will add a kwarg `is_causal=True`, otherwise will raise an error. This is because context parallelism does
+    not support attention masks. This function modifies the model in place.
+
+    Args:
+        model (`nn.Module`):
+            The model to attach the hooks to.
+
+    """
+
+    def _self_attn_pre_forward_hook(_module, module_args, module_kwargs):
+        if "attention_mask" in module_kwargs:
+            module_kwargs["attention_mask"] = None
+            module_kwargs["is_causal"] = True
+
+        return module_args, module_kwargs
+
+    for name, module in model.named_modules():
+        # We hope (assume) that if user uses their own model (without this structure which transformers uses), they read the docs saying they can't pass in attention masks
+        # Then these cases can happen:
+        # 1) some modules end with a `self-attn` module, in which case we attach the hook, but the
+        #    there's no attention mask kwarg -> hook is a no-op
+        # 2) some modules end with a `self-attn` module, in which case we attach the hook, and the
+        #    attention mask kwarg is passed -> hook will remove the attention mask and add
+        #    `is_causal=True` kwarg, which either crashes the training or fixes it
+        #    (training would crash anyway as attention mask isn't supported)
+        # 3) no modules end with a `self-attn` module, in which case we don't attach the hook, this is
+        #    a no-op as well
+        if name.endswith("self_attn"):
+            # we want the hook to be executed first, to avoid any other hooks doing work on the attention mask
+            module.register_forward_pre_hook(_self_attn_pre_forward_hook, with_kwargs=True, prepend=True)

src/accelerate/checkpointing.py

@@ -14,12 +14,11 @@
 
 import random
 from pathlib import Path
-from typing import List
+from typing import Optional
 
 import numpy as np
 import torch
 from safetensors.torch import load_model
-from torch.cuda.amp import GradScaler
 
 from .utils import (
     MODEL_NAME,
@@ -37,6 +36,7 @@ from .utils import (
     is_mlu_available,
     is_musa_available,
     is_sdaa_available,
+    is_torch_version,
     is_torch_xla_available,
     is_xpu_available,
     load,
@@ -44,6 +44,11 @@ from .utils import (
 )
 
 
+if is_torch_version(">=", "2.4.0"):
+    from torch.amp import GradScaler
+else:
+    from torch.cuda.amp import GradScaler
+
 if is_torch_xla_available():
     import torch_xla.core.xla_model as xm
 
@@ -56,13 +61,13 @@ logger = get_logger(__name__)
 
 def save_accelerator_state(
     output_dir: str,
-    model_states: List[dict],
+    model_states: list[dict],
     optimizers: list,
     schedulers: list,
     dataloaders: list,
     process_index: int,
     step: int,
-    scaler: GradScaler = None,
+    scaler: Optional[GradScaler] = None,
     save_on_each_node: bool = False,
     safe_serialization: bool = True,
 ):
@@ -181,6 +186,7 @@ def load_accelerator_state(
     process_index,
     scaler=None,
     map_location=None,
+    load_kwargs=None,
     **load_model_func_kwargs,
 ):
     """
@@ -201,6 +207,8 @@ def load_accelerator_state(
             An optional *GradScaler* instance to load
         map_location (`str`, *optional*):
             What device to load the optimizer state onto. Should be one of either "cpu" or "on_device".
+        load_kwargs (`dict`, *optional*):
+            Additional arguments that can be passed to the `load` function.
         load_model_func_kwargs (`dict`, *optional*):
             Additional arguments that can be passed to the model's `load_state_dict` method.
 
@@ -218,6 +226,9 @@ def load_accelerator_state(
     elif map_location == "on_device":
         map_location = PartialState().device
 
+    if load_kwargs is None:
+        load_kwargs = {}
+
     input_dir = Path(input_dir)
     # Model states
     for i, model in enumerate(models):
@@ -236,7 +247,7 @@ def load_accelerator_state(
     for i, opt in enumerate(optimizers):
         optimizer_name = f"{OPTIMIZER_NAME}.bin" if i == 0 else f"{OPTIMIZER_NAME}_{i}.bin"
         input_optimizer_file = input_dir.joinpath(optimizer_name)
-        optimizer_state = load(input_optimizer_file, map_location=map_location)
+        optimizer_state = load(input_optimizer_file, map_location=map_location, **load_kwargs)
         optimizers[i].load_state_dict(optimizer_state)
     logger.info("All optimizer states loaded successfully")
 
@@ -244,7 +255,7 @@ def load_accelerator_state(
     for i, scheduler in enumerate(schedulers):
         scheduler_name = f"{SCHEDULER_NAME}.bin" if i == 0 else f"{SCHEDULER_NAME}_{i}.bin"
         input_scheduler_file = input_dir.joinpath(scheduler_name)
-        scheduler_state = load(input_scheduler_file)
+        scheduler_state = load(input_scheduler_file, **load_kwargs)
         scheduler.load_state_dict(scheduler_state)
     logger.info("All scheduler states loaded successfully")
 
@@ -262,7 +273,7 @@ def load_accelerator_state(
             dataloader_state_dict_name = "dl_state_dict.bin" if i == 0 else f"dl_state_dict_{i}.bin"
             input_dataloader_state_dict_file = input_dir.joinpath(dataloader_state_dict_name)
             if input_dataloader_state_dict_file.exists():
-                state_dict = load(input_dataloader_state_dict_file)
+                state_dict = load(input_dataloader_state_dict_file, **load_kwargs)
                 dataloader.load_state_dict(state_dict)
     logger.info("All dataloader sampler states loaded successfully")
 

src/accelerate/commands/accelerate_cli.py

@@ -20,6 +20,7 @@ from accelerate.commands.estimate import estimate_command_parser
 from accelerate.commands.launch import launch_command_parser
 from accelerate.commands.merge import merge_command_parser
 from accelerate.commands.test import test_command_parser
+from accelerate.commands.to_fsdp2 import to_fsdp2_command_parser
 from accelerate.commands.tpu import tpu_command_parser
 from accelerate.commands.utils import CustomArgumentParser
 
@@ -36,6 +37,7 @@ def main():
     merge_command_parser(subparsers=subparsers)
     tpu_command_parser(subparsers=subparsers)
     test_command_parser(subparsers=subparsers)
+    to_fsdp2_command_parser(subparsers=subparsers)
 
     # Let's go
     args = parser.parse_args()

src/accelerate/commands/config/cluster.py

@@ -21,6 +21,7 @@ from ...utils import (
     DistributedType,
     is_deepspeed_available,
     is_fp8_available,
+    is_hpu_available,
     is_mlu_available,
     is_mps_available,
     is_msamp_available,
@@ -33,6 +34,7 @@ from ...utils import (
 )
 from ...utils.constants import (
     DEEPSPEED_MULTINODE_LAUNCHERS,
+    FSDP2_STATE_DICT_TYPE,
     FSDP_AUTO_WRAP_POLICY,
     FSDP_BACKWARD_PREFETCH,
     FSDP_SHARDING_STRATEGY,
@@ -59,6 +61,7 @@ def get_cluster_input():
             "No distributed training",
             "multi-CPU",
             "multi-XPU",
+            "multi-HPU",
             "multi-GPU",
             "multi-NPU",
             "multi-MLU",
@@ -87,6 +90,7 @@ def get_cluster_input():
         DistributedType.MULTI_NPU,
         DistributedType.MULTI_XPU,
         DistributedType.MULTI_CPU,
+        DistributedType.MULTI_HPU,
     ]:
         num_machines = _ask_field(
             "How many different machines will you use (use more than 1 for multi-node training)? [1]: ",
@@ -137,13 +141,15 @@ def get_cluster_input():
 
     ipex_config = {}
     mpirun_config = {}
-    if use_cpu:
+    if use_cpu or is_xpu_available():
         ipex_config["ipex"] = _ask_field(
-            "Do you want to use Intel PyTorch Extension (IPEX) to speed up training on CPU? [yes/NO]:",
+            "Do you want to use Intel PyTorch Extension (IPEX) to speed up training on CPU/XPU? [yes/NO]:",
             _convert_yes_no_to_bool,
             default=False,
             error_message="Please enter yes or no.",
         )
+
+    if use_cpu:
         if distributed_type == DistributedType.MULTI_CPU:
             use_mpirun = _ask_field(
                 "Do you want accelerate to launch mpirun? [yes/NO]: ",
@@ -159,25 +165,6 @@ def get_cluster_input():
                 )
                 mpirun_config["mpirun_hostfile"] = os.path.expanduser(mpirun_hostfile.strip())
                 mpirun_config["mpirun_ccl"] = _ask_field("Enter the number of oneCCL worker threads [1]: ", default=1)
-    if (
-        not use_cpu
-        and is_xpu_available()
-        and distributed_type
-        not in [
-            DistributedType.MULTI_GPU,
-            DistributedType.MULTI_NPU,
-            DistributedType.MULTI_MLU,
-            DistributedType.MULTI_SDAA,
-            DistributedType.XLA,
-            DistributedType.MULTI_MUSA,
-        ]
-    ):
-        ipex_config["use_xpu"] = _ask_field(
-            "Do you want to use XPU plugin to speed up training on XPU? [yes/NO]:",
-            _convert_yes_no_to_bool,
-            default=False,
-            error_message="Please enter yes or no.",
-        )
 
     dynamo_config = {}
     use_dynamo = _ask_field(
@@ -220,6 +207,12 @@ def get_cluster_input():
                 default=False,
                 error_message="Please enter yes or no.",
             )
+            dynamo_config[prefix + "use_regional_compilation"] = _ask_field(
+                "Do you want to enable regional compilation? [yes/NO]: ",
+                _convert_yes_no_to_bool,
+                default=False,
+                error_message="Please enter yes or no.",
+            )
 
     use_mps = not use_cpu and is_mps_available()
     deepspeed_config = {}
@@ -228,6 +221,7 @@ def get_cluster_input():
         in [
             DistributedType.MULTI_GPU,
             DistributedType.MULTI_XPU,
+            DistributedType.MULTI_HPU,
             DistributedType.MULTI_NPU,
             DistributedType.MULTI_MLU,
             DistributedType.MULTI_SDAA,
@@ -244,9 +238,9 @@ def get_cluster_input():
         )
         if use_deepspeed:
             distributed_type = DistributedType.DEEPSPEED
-            assert (
-                is_deepspeed_available()
-            ), "DeepSpeed is not installed => run `pip3 install deepspeed` or build it from source"
+            assert is_deepspeed_available(), (
+                "DeepSpeed is not installed => run `pip3 install deepspeed` or build it from source"
+            )
 
         if distributed_type == DistributedType.DEEPSPEED:
             use_deepspeed_config = _ask_field(
@@ -381,7 +375,7 @@ def get_cluster_input():
                         )
 
     fsdp_config = {}
-    tp_config = {}
+
     if distributed_type in [
         DistributedType.MULTI_GPU,
         DistributedType.MULTI_NPU,
@@ -389,6 +383,7 @@ def get_cluster_input():
         DistributedType.MULTI_SDAA,
         DistributedType.MULTI_MUSA,
         DistributedType.MULTI_XPU,
+        DistributedType.MULTI_HPU,
     ]:
         use_fsdp = _ask_field(
             "Do you want to use FullyShardedDataParallel? [yes/NO]: ",
@@ -399,18 +394,36 @@ def get_cluster_input():
         if use_fsdp:
             distributed_type = DistributedType.FSDP
         if distributed_type == DistributedType.FSDP:
-            sharding_strategy_query = "What should be your sharding strategy?"
-            fsdp_config["fsdp_sharding_strategy"] = _ask_options(
-                sharding_strategy_query,
-                FSDP_SHARDING_STRATEGY,
-                lambda x: FSDP_SHARDING_STRATEGY[int(x)],
+            fsdp_config["fsdp_version"] = _ask_options(
+                "What should be your FSDP version? [2]: ",
+                [1, 2],
+                lambda x: int(x) + 1,
+                default=1,
             )
+            fsdp_version = fsdp_config["fsdp_version"]  # extract to a variable to simplify usage later
+
+            if fsdp_version == 1:
+                sharding_strategy_query = "What should be your sharding strategy?"
+                fsdp_config["fsdp_reshard_after_forward"] = _ask_options(
+                    sharding_strategy_query,
+                    FSDP_SHARDING_STRATEGY,
+                    lambda x: FSDP_SHARDING_STRATEGY[int(x)],
+                )
+            else:
+                fsdp_config["fsdp_reshard_after_forward"] = _ask_field(
+                    "Do you want to enable resharding after forward? [YES/no]: ",
+                    _convert_yes_no_to_bool,
+                    default=True,
+                    error_message="Please enter yes or no.",
+                )
+
             fsdp_config["fsdp_offload_params"] = _ask_field(
                 "Do you want to offload parameters and gradients to CPU? [yes/NO]: ",
                 _convert_yes_no_to_bool,
                 default=False,
                 error_message="Please enter yes or no.",
             )
+
             fsdp_wrap_query = "What should be your auto wrap policy?"
             fsdp_config["fsdp_auto_wrap_policy"] = _ask_options(
                 fsdp_wrap_query,
@@ -436,67 +449,109 @@ def get_cluster_input():
                     int,
                     default=100000000,
                 )
-            fsdp_backward_prefetch_query = "What should be your FSDP's backward prefetch policy?"
-            fsdp_config["fsdp_backward_prefetch"] = _ask_options(
-                fsdp_backward_prefetch_query,
-                FSDP_BACKWARD_PREFETCH,
-                lambda x: FSDP_BACKWARD_PREFETCH[int(x)],
-            )
+            # Removed in FSDP2, ask for user input for FSDP1
+            if fsdp_version == 1:
+                fsdp_backward_prefetch_query = "What should be your FSDP's backward prefetch policy?"
+                fsdp_config["fsdp_backward_prefetch"] = _ask_options(
+                    fsdp_backward_prefetch_query,
+                    FSDP_BACKWARD_PREFETCH,
+                    lambda x: FSDP_BACKWARD_PREFETCH[int(x)],
+                )
+
             fsdp_state_dict_type_query = "What should be your FSDP's state dict type?"
             fsdp_config["fsdp_state_dict_type"] = _ask_options(
                 fsdp_state_dict_type_query,
-                FSDP_STATE_DICT_TYPE,
-                lambda x: FSDP_STATE_DICT_TYPE[int(x)],
-                default=2,
-            )
-            fsdp_config["fsdp_forward_prefetch"] = _ask_field(
-                "Do you want to enable FSDP's forward prefetch policy? [yes/NO]: ",
-                _convert_yes_no_to_bool,
-                default=False,
-                error_message="Please enter yes or no.",
-            )
-            fsdp_config["fsdp_use_orig_params"] = _ask_field(
-                "Do you want to enable FSDP's `use_orig_params` feature? [YES/no]: ",
-                _convert_yes_no_to_bool,
-                default=True,
-                error_message="Please enter yes or no.",
+                FSDP_STATE_DICT_TYPE if fsdp_version == 1 else FSDP2_STATE_DICT_TYPE,
+                lambda x: FSDP_STATE_DICT_TYPE[int(x)] if fsdp_version == 1 else FSDP2_STATE_DICT_TYPE[int(x)],
+                default=0,
             )
+            # Not implemented in FSDP2, ask for user input for FSDP1
+            if fsdp_version == 1:
+                fsdp_config["fsdp_forward_prefetch"] = _ask_field(
+                    "Do you want to enable FSDP's forward prefetch policy? [yes/NO]: ",
+                    _convert_yes_no_to_bool,
+                    default=False,
+                    error_message="Please enter yes or no.",
+                )
+            # Obsolete in FSDP2, ask for user input for FSDP1
+            if fsdp_version == 1:
+                fsdp_config["fsdp_use_orig_params"] = _ask_field(
+                    "Do you want to enable FSDP's `use_orig_params` feature? [YES/no]: ",
+                    _convert_yes_no_to_bool,
+                    default=True,
+                    error_message="Please enter yes or no.",
+                )
             fsdp_config["fsdp_cpu_ram_efficient_loading"] = _ask_field(
                 "Do you want to enable CPU RAM efficient model loading? Only applicable for 🤗 Transformers models. [YES/no]: ",
                 _convert_yes_no_to_bool,
                 default=True,
                 error_message="Please enter yes or no.",
             )
-            if fsdp_config["fsdp_cpu_ram_efficient_loading"]:
-                fsdp_config["fsdp_sync_module_states"] = True
-            else:
-                fsdp_config["fsdp_sync_module_states"] = _ask_field(
-                    "Do you want each individually wrapped FSDP unit to broadcast module parameters from rank 0 at the start? [YES/no]: ",
-                    _convert_yes_no_to_bool,
-                    default=True,
-                    error_message="Please enter yes or no.",
-                )
+            # Obsolete in FSDP2, ask for user input for FSDP1
+            if fsdp_version == 1:
+                if fsdp_config["fsdp_cpu_ram_efficient_loading"]:
+                    fsdp_config["fsdp_sync_module_states"] = True
+                else:
+                    fsdp_config["fsdp_sync_module_states"] = _ask_field(
+                        "Do you want each individually wrapped FSDP unit to broadcast module parameters from rank 0 at the start? [YES/no]: ",
+                        _convert_yes_no_to_bool,
+                        default=True,
+                        error_message="Please enter yes or no.",
+                    )
             fsdp_config["fsdp_activation_checkpointing"] = _ask_field(
                 "Do you want to enable FSDP activation checkpointing? [yes/NO]: ",
                 _convert_yes_no_to_bool,
                 default=False,
                 error_message="Please enter yes or no.",
             )
-        if not use_fsdp:
-            use_tp = _ask_field(
-                "Do you want to use TensorParallel? [yes/NO]: ",
-                _convert_yes_no_to_bool,
-                default=False,
-                error_message="Please enter yes or no.",
+
+    parallelism_config = {}
+
+    if fsdp_config.get("fsdp_version", 1) == 2:
+        use_parallelism_config = _ask_field(
+            "Do you want to use the parallelism config? [yes/NO]: ",
+            _convert_yes_no_to_bool,
+            default=False,
+            error_message="Please enter yes or no.",
+        )
+
+        if use_parallelism_config:
+            prefix = "parallelism_config_"
+            parallelism_config[prefix + "dp_replicate_size"] = _ask_field(
+                "What is the data parallelism replicate size? [1]: ",
+                int,
+                default=1,
+                error_message="Please enter an integer.",
             )
-            if use_tp:
-                distributed_type = DistributedType.TP
-            if distributed_type == DistributedType.TP:
-                tp_config["tp_size"] = _ask_field(
-                    "What should be your Tensor Parallel degree? [1]: ",
-                    int,
-                    default=1,
+
+            parallelism_config[prefix + "dp_shard_size"] = _ask_field(
+                "What is the FSDP shard size? [1]: ",
+                int,
+                default=1,
+                error_message="Please enter an integer.",
+            )
+
+            parallelism_config[prefix + "tp_size"] = _ask_field(
+                "What is the tensor parallelism size? [1]: ",
+                int,
+                default=1,
+                error_message="Please enter an integer.",
+            )
+
+            parallelism_config[prefix + "cp_size"] = _ask_field(
+                "What is the context parallelism size? [1]: ",
+                int,
+                default=1,
+                error_message="Please enter an integer.",
+            )
+            if parallelism_config[prefix + "cp_size"] > 1:
+                parallelism_config[prefix + "cp_comm_strategy"] = _ask_options(
+                    "What is the compute parallelism communication strategy?",
+                    ["allgather", "alltoall"],
+                    lambda x: ["allgather", "alltoall"][int(x)],
+                    default=0,
                 )
+
     megatron_lm_config = {}
     if distributed_type in [DistributedType.MULTI_GPU]:
         use_megatron_lm = _ask_field(
@@ -571,6 +626,7 @@ def get_cluster_input():
     if distributed_type in [
         DistributedType.MULTI_CPU,
         DistributedType.MULTI_XPU,
+        DistributedType.MULTI_HPU,
         DistributedType.MULTI_GPU,
         DistributedType.MULTI_MLU,
         DistributedType.MULTI_SDAA,
@@ -615,6 +671,7 @@ def get_cluster_input():
             DistributedType.MULTI_MUSA,
             DistributedType.MULTI_NPU,
             DistributedType.MULTI_XPU,
+            DistributedType.MULTI_HPU,
             DistributedType.NO,
         ]
         and not use_cpu
@@ -630,10 +687,12 @@ def get_cluster_input():
             machine_type = "MUSA(s)"
         elif is_xpu_available():
             machine_type = "XPU(s)"
+        elif is_hpu_available():
+            machine_type = "HPU(s)"
         else:
             machine_type = "GPU(s)"
         gpu_ids = _ask_field(
-            f"What {machine_type} (by id) should be used for training on this machine as a comma-seperated list? [all]:",
+            f"What {machine_type} (by id) should be used for training on this machine as a comma-separated list? [all]:",
             default="all",
         )
 
@@ -697,7 +756,7 @@ def get_cluster_input():
                     )
                     tpu_command_file = os.path.abspath(tpu_command_file)
                 else:
-                    print("Please enter each command seperately you wish to run on startup in each pod.")
+                    print("Please enter each command separately you wish to run on startup in each pod.")
                     tpu_commands = []
                     another_command = True
                     while another_command:
@@ -715,11 +774,11 @@ def get_cluster_input():
                             error_message="Please enter yes or no.",
                         )
             tpu_vm = _ask_field(
-                "If not using an instance group, what are the names of the Compute VM instances to be used, seperated by a comma: ",
+                "If not using an instance group, what are the names of the Compute VM instances to be used, separated by a comma: ",
                 default="",
             ).split(",")
             tpu_env = _ask_field(
-                "What environment variables do you wish to set in each pod, seperated by a comma: ",
+                "What environment variables do you wish to set in each pod, separated by a comma: ",
                 default="",
             ).split(",")
 
@@ -762,8 +821,8 @@ def get_cluster_input():
                     )
                     fp8_config["fp8_format"] = _ask_options(
                         "Which weight format should be used?",
-                        ["HYBRID", "E4M3"],
-                        lambda x: "HYBRID" if x == 0 else "E4M3",
+                        ["HYBRID", "E4M3", "E5M2"],
+                        lambda i: ["HYBRID", "E4M3", "E5M2"][i],
                         default=0,
                     )
                     fp8_config["amax_history_length"] = _ask_field(
@@ -799,6 +858,8 @@ def get_cluster_input():
                             default=False,
                         )
                         fp8_config["override_linear_precision"] = (fprop, dgrad, wgrad)
+                    else:
+                        fp8_config["override_linear_precision"] = (False, False, False)
 
                 elif fp8_config["backend"] == "MSAMP":
                     if not is_msamp_available():
@@ -835,7 +896,7 @@ def get_cluster_input():
         fp8_config=fp8_config,
         deepspeed_config=deepspeed_config,
         fsdp_config=fsdp_config,
-        tp_config=tp_config,
+        parallelism_config=parallelism_config,
         megatron_lm_config=megatron_lm_config,
         ipex_config=ipex_config,
         mpirun_config=mpirun_config,

src/accelerate/commands/config/config_args.py

@@ -18,7 +18,7 @@ import json
 import os
 from dataclasses import dataclass
 from enum import Enum
-from typing import List, Optional, Union
+from typing import Optional, Union
 
 import yaml
 
@@ -189,42 +189,40 @@ class ClusterConfig(BaseConfig):
     enable_cpu_affinity: bool = False
 
     # args for FP8 training
-    fp8_config: dict = None
+    fp8_config: Optional[dict] = None
     # args for deepspeed_plugin
-    deepspeed_config: dict = None
+    deepspeed_config: Optional[dict] = None
     # args for fsdp
-    fsdp_config: dict = None
-    # args for tp
-    tp_config: dict = None
+    fsdp_config: Optional[dict] = None
+    # args for parallelism config
+    parallelism_config: Optional[dict] = None
     # args for megatron_lm
-    megatron_lm_config: dict = None
+    megatron_lm_config: Optional[dict] = None
     # args for ipex
-    ipex_config: dict = None
+    ipex_config: Optional[dict] = None
     # args for mpirun
-    mpirun_config: dict = None
+    mpirun_config: Optional[dict] = None
     # args for TPU
     downcast_bf16: bool = False
 
     # args for TPU pods
-    tpu_name: str = None
-    tpu_zone: str = None
+    tpu_name: Optional[str] = None
+    tpu_zone: Optional[str] = None
     tpu_use_cluster: bool = False
     tpu_use_sudo: bool = False
-    command_file: str = None
-    commands: List[str] = None
-    tpu_vm: List[str] = None
-    tpu_env: List[str] = None
+    command_file: Optional[str] = None
+    commands: list[str] = None
+    tpu_vm: list[str] = None
+    tpu_env: list[str] = None
 
     # args for dynamo
-    dynamo_config: dict = None
+    dynamo_config: Optional[dict] = None
 
     def __post_init__(self):
         if self.deepspeed_config is None:
             self.deepspeed_config = {}
         if self.fsdp_config is None:
             self.fsdp_config = {}
-        if self.tp_config is None:
-            self.tp_config = {}
         if self.megatron_lm_config is None:
             self.megatron_lm_config = {}
         if self.ipex_config is None:
@@ -233,6 +231,8 @@ class ClusterConfig(BaseConfig):
             self.mpirun_config = {}
         if self.fp8_config is None:
             self.fp8_config = {}
+        if self.parallelism_config is None:
+            self.parallelism_config = {}
         return super().__post_init__()
 
 
@@ -249,8 +249,8 @@ class SageMakerConfig(BaseConfig):
     pytorch_version: str = SAGEMAKER_PYTORCH_VERSION
     transformers_version: str = SAGEMAKER_TRANSFORMERS_VERSION
     py_version: str = SAGEMAKER_PYTHON_VERSION
-    sagemaker_inputs_file: str = None
-    sagemaker_metrics_file: str = None
-    additional_args: dict = None
-    dynamo_config: dict = None
+    sagemaker_inputs_file: Optional[str] = None
+    sagemaker_metrics_file: Optional[str] = None
+    additional_args: Optional[dict] = None
+    dynamo_config: Optional[dict] = None
     enable_cpu_affinity: bool = False

src/accelerate/commands/config/config_utils.py

@@ -72,9 +72,18 @@ def _convert_compute_environment(value):
 def _convert_distributed_mode(value):
     value = int(value)
     return DistributedType(
-        ["NO", "MULTI_CPU", "MULTI_XPU", "MULTI_GPU", "MULTI_NPU", "MULTI_MLU", "MULTI_SDAA", "MULTI_MUSA", "XLA"][
-            value
-        ]
+        [
+            "NO",
+            "MULTI_CPU",
+            "MULTI_XPU",
+            "MULTI_HPU",
+            "MULTI_GPU",
+            "MULTI_NPU",
+            "MULTI_MLU",
+            "MULTI_SDAA",
+            "MULTI_MUSA",
+            "XLA",
+        ][value]
     )
 
 

src/accelerate/commands/config/default.py

@@ -33,7 +33,7 @@ from .config_utils import SubcommandHelpFormatter
 description = "Create a default config file for Accelerate with only a few flags set."
 
 
-def write_basic_config(mixed_precision="no", save_location: str = default_json_config_file, use_xpu: bool = False):
+def write_basic_config(mixed_precision="no", save_location: str = default_json_config_file):
     """
     Creates and saves a basic cluster config to be used on a local machine with potentially multiple GPUs. Will also
     set CPU if it is a CPU-only machine.
@@ -43,10 +43,8 @@ def write_basic_config(mixed_precision="no", save_location: str = default_json_c
             Mixed Precision to use. Should be one of "no", "fp16", or "bf16"
         save_location (`str`, *optional*, defaults to `default_json_config_file`):
             Optional custom save location. Should be passed to `--config_file` when using `accelerate launch`. Default
-            location is inside the huggingface cache folder (`~/.cache/huggingface`) but can be overriden by setting
+            location is inside the huggingface cache folder (`~/.cache/huggingface`) but can be overridden by setting
             the `HF_HOME` environmental variable, followed by `accelerate/default_config.yaml`.
-        use_xpu (`bool`, *optional*, defaults to `False`):
-            Whether to use XPU if available.
     """
     path = Path(save_location)
     path.parent.mkdir(parents=True, exist_ok=True)
@@ -104,7 +102,7 @@ def write_basic_config(mixed_precision="no", save_location: str = default_json_c
             config["distributed_type"] = "MULTI_GPU"
         else:
             config["distributed_type"] = "NO"
-    elif is_xpu_available() and use_xpu:
+    elif is_xpu_available():
         num_xpus = torch.xpu.device_count()
         config["num_processes"] = num_xpus
         config["use_cpu"] = False

src/accelerate/commands/config/sagemaker.py

@@ -212,6 +212,13 @@ def get_sagemaker_input():
                 default=False,
                 error_message="Please enter yes or no.",
             )
+            dynamo_config[prefix + "use_regional_compilation"] = _ask_field(
+                "Do you want to enable regional compilation? [yes/NO]: ",
+                _convert_yes_no_to_bool,
+                default=False,
+                error_message="Please enter yes or no.",
+            )
+
     ec2_instance_query = "Which EC2 instance type you want to use for your training?"
     if distributed_type != SageMakerDistributedType.NO:
         ec2_instance_type = _ask_options(