disable

* Adding SDPA support for RoBERTa-based models

* add not is_cross_attention

* fix copies

* fix test

* add minimal test for camembert and xlm_roberta as their test class does not inherit from ModelTesterMixin

* address some review comments

* use copied from

* style

* consistency

* fix lists

---------

Co-authored-by: fxmarty <9808326+fxmarty@users.noreply.github.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

.circleci/config.yml

@@ -142,6 +142,7 @@ jobs:
             - run: python utils/custom_init_isort.py --check_only
             - run: python utils/sort_auto_mappings.py --check_only
             - run: python utils/check_doc_toc.py
+            - run: python utils/check_docstrings.py --check_all
 
     check_repository_consistency:
         working_directory: ~/transformers
@@ -190,4 +191,4 @@ workflows:
             - check_circleci_user
             - check_code_quality
             - check_repository_consistency
-            - fetch_all_tests
+            - fetch_all_tests

.circleci/create_circleci_config.py

@@ -121,11 +121,16 @@ class CircleCIJob:
         )
 
         steps.append({"run": {"name": "Create `test-results` directory", "command": "mkdir test-results"}})
+
+        # Examples special case: we need to download NLTK files in advance to avoid cuncurrency issues
+        if "examples" in self.name:
+            steps.append({"run": {"name": "Download NLTK files", "command": """python -c "import nltk; nltk.download('punkt', quiet=True)" """}})
+
         test_command = ""
         if self.command_timeout:
             test_command = f"timeout {self.command_timeout} "
         # junit familiy xunit1 is necessary to support splitting on test name or class name with circleci split
-        test_command += f"python3 -m pytest -rsfE -p no:warnings -o junit_family=xunit1 --tb=short --junitxml=test-results/junit.xml -n {self.pytest_num_workers} " + " ".join(pytest_flags)
+        test_command += f"python3 -m pytest -rsfE -p no:warnings --tb=short -o junit_family=xunit1 --junitxml=test-results/junit.xml -n {self.pytest_num_workers} " + " ".join(pytest_flags)
 
         if self.parallelism == 1:
             if self.tests_to_run is None:
@@ -185,10 +190,6 @@ class CircleCIJob:
             steps.append({"store_artifacts": {"path": "tests.txt"}})
             steps.append({"store_artifacts": {"path": "splitted_tests.txt"}})
 
-            test_command = ""
-            if self.command_timeout:
-                test_command = f"timeout {self.command_timeout} "
-            test_command += f"python3 -m pytest -rsfE -p no:warnings --tb=short  -o junit_family=xunit1 --junitxml=test-results/junit.xml -n {self.pytest_num_workers} " + " ".join(pytest_flags)
             test_command += " $(cat splitted_tests.txt)"
         if self.marker is not None:
             test_command += f" -m {self.marker}"
@@ -248,7 +249,7 @@ torch_job = CircleCIJob(
     docker_image=[{"image": "huggingface/transformers-torch-light"}],
     install_steps=["uv venv && uv pip install ."],
     parallelism=6,
-    pytest_num_workers=16
+    pytest_num_workers=4
 )
 
 tokenization_job = CircleCIJob(
@@ -256,7 +257,7 @@ tokenization_job = CircleCIJob(
     docker_image=[{"image": "huggingface/transformers-torch-light"}],
     install_steps=["uv venv && uv pip install ."],
     parallelism=6,
-    pytest_num_workers=16
+    pytest_num_workers=4
 )
 
 
@@ -265,7 +266,7 @@ tf_job = CircleCIJob(
     docker_image=[{"image":"huggingface/transformers-tf-light"}],
     install_steps=["uv venv", "uv pip install -e."],
     parallelism=6,
-    pytest_num_workers=16,
+    pytest_num_workers=4,
 )
 
 
@@ -274,7 +275,7 @@ flax_job = CircleCIJob(
     docker_image=[{"image":"huggingface/transformers-jax-light"}],
     install_steps=["uv venv && uv pip install ."],
     parallelism=6,
-    pytest_num_workers=16
+    pytest_num_workers=4
 )
 
 

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -1,6 +1,17 @@
 name: "\U0001F41B Bug Report"
 description: Submit a bug report to help us improve transformers
+labels: [ "bug" ]
 body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to fill out this bug report! 🤗
+
+        Before you submit your bug report:
+
+          - If it is your first time submitting, be sure to check our [bug report guidelines](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#did-you-find-a-bug)
+          - Try our [docs bot](https://huggingface.co/spaces/huggingchat/hf-docs-chat) -- it might be able to help you with your issue
+
   - type: textarea
     id: system-info
     attributes:
@@ -25,7 +36,7 @@ body:
 
         Models:
 
-          - text models: @ArthurZucker 
+          - text models: @ArthurZucker
           - vision models: @amyeroberts
           - speech models: @sanchit-gandhi
           - graph models: @clefourrier
@@ -38,9 +49,9 @@ body:
           - tensorflow: @gante and @Rocketknight1
           - tokenizers: @ArthurZucker
           - trainer: @muellerzr @SunMarc
-        
+
         Integrations:
-        
+
           - deepspeed: HF Trainer/Accelerate: @muellerzr
           - ray/raytune: @richardliaw, @amogkam
           - Big Model Inference: @SunMarc

.github/ISSUE_TEMPLATE/i18n.md

@@ -34,7 +34,7 @@ Some notes:
 
 ## Tutorial section
 - [ ] [pipeline_tutorial.md](https://github.com/huggingface/transformers/blob/main/docs/source/en/pipeline_tutorial.md)
-- [ ]  [autoclass_tutorial.md](https://github.com/huggingface/transformers/blob/master/docs/source/autoclass_tutorial.md)
+- [ ]  [autoclass_tutorial.md](https://github.com/huggingface/transformers/blob/main/docs/source/en/autoclass_tutorial.md)
 - [ ]  [preprocessing.md](https://github.com/huggingface/transformers/blob/main/docs/source/en/preprocessing.md)
 - [ ]  [training.md](https://github.com/huggingface/transformers/blob/main/docs/source/en/training.md)
 - [ ]  [accelerate.md](https://github.com/huggingface/transformers/blob/main/docs/source/en/accelerate.md)

.github/PULL_REQUEST_TEMPLATE.md

@@ -58,9 +58,9 @@ Integrations:
 - deepspeed: HF Trainer/Accelerate: @muellerzr
 - ray/raytune: @richardliaw, @amogkam
 - Big Model Inference: @SunMarc
-- quantization (bitsandbytes, autogpt): @SunMarc 
+- quantization (bitsandbytes, autogpt): @SunMarc
 
-Documentation: @stevhliu and @MKhalusova
+Documentation: @stevhliu
 
 HF projects:
 

.github/workflows/add-model-like.yml

@@ -23,7 +23,7 @@ jobs:
           sudo apt -y update && sudo apt install -y libsndfile1-dev
 
       - name: Load cached virtual environment
-        uses: actions/cache@v2
+        uses: actions/cache@v4
         id: cache
         with:
           path: ~/venv/

.github/workflows/benchmark.yml

@@ -31,12 +31,12 @@ jobs:
         if: github.event_name == 'schedule'
         working-directory: /transformers
         run: |
-          python3 -m pip install optimum-benchmark>=0.2.0
+          python3 -m pip install optimum-benchmark>=0.3.0
           HF_TOKEN=${{ secrets.TRANSFORMERS_BENCHMARK_TOKEN }} python3 benchmark/benchmark.py --repo_id hf-internal-testing/benchmark_results --path_in_repo $(date +'%Y-%m-%d') --config-dir benchmark/config --config-name generation --commit=${{ github.sha }} backend.model=google/gemma-2b backend.cache_implementation=null,static backend.torch_compile=false,true --multirun
 
       - name: Benchmark (merged to main event)
         if: github.event_name == 'push' && github.ref_name == 'main'
         working-directory: /transformers
         run: |
-          python3 -m pip install optimum-benchmark>=0.2.0
+          python3 -m pip install optimum-benchmark>=0.3.0
           HF_TOKEN=${{ secrets.TRANSFORMERS_BENCHMARK_TOKEN }} python3 benchmark/benchmark.py --repo_id hf-internal-testing/benchmark_results_merge_event --path_in_repo $(date +'%Y-%m-%d') --config-dir benchmark/config --config-name generation --commit=${{ github.sha }} backend.model=google/gemma-2b backend.cache_implementation=null,static backend.torch_compile=false,true --multirun

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

@@ -27,10 +27,10 @@ jobs:
     strategy:
       matrix:
         file: ["quality", "consistency", "custom-tokenizers", "torch-light", "tf-light", "exotic-models", "torch-tf-light", "torch-jax-light", "jax-light", "examples-torch",  "examples-tf"]
-    continue-on-error: true 
+    continue-on-error: true
 
     steps:
-      - 
+      -
         name: Set tag
         run: |
               if ${{contains(github.event.head_commit.message, '[build-ci-image]')}}; then
@@ -61,4 +61,17 @@ jobs:
             REF=${{ github.sha }}
           file: "./docker/${{ matrix.file }}.dockerfile"
           push: ${{ contains(github.event.head_commit.message, 'ci-image]') ||  github.event_name == 'schedule' }}
-          tags: ${{ env.TAG }}
+          tags: ${{ env.TAG }}
+
+  notify:
+    runs-on: ubuntu-22.04
+    if: ${{ contains(github.event.head_commit.message, '[build-ci-image]') || contains(github.event.head_commit.message, '[push-ci-image]') && '!cancelled()' || github.event_name == 'schedule' }}
+    steps:
+      - name: Post to Slack
+        if: ${{ contains(github.event.head_commit.message, '[push-ci-image]') && github.event_name != 'schedule' }}
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        with:
+          slack_channel: "#transformers-ci-circleci-images"
+          title: 🤗 New docker images for CircleCI are pushed.
+          status: ${{ job.status }}
+          slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}

.github/workflows/check_tiny_models.yml

@@ -23,7 +23,7 @@ jobs:
 
       - uses: actions/checkout@v4
       - name: Set up Python 3.8
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           # Semantic version range syntax or exact version of a Python version
           python-version: '3.8'

.github/workflows/release-conda.yml

@@ -19,7 +19,7 @@ jobs:
 
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Install miniconda
         uses: conda-incubator/setup-miniconda@v2

.github/workflows/self-pr-slow-ci.yml

@@ -4,7 +4,7 @@ on:
   pull_request:
     paths:
       - "src/transformers/models/*/modeling_*.py"
-      - "tests/models/*/test_*.py"
+      - "tests/**/test_*.py"
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}

.github/workflows/self-push-amd.yml

@@ -324,6 +324,7 @@ jobs:
         # We pass `needs.setup_gpu.outputs.matrix` as the argument. A processing in `notification_service.py` to change
         # `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
         run: |
+          pip install huggingface_hub
           pip install slack_sdk
           pip show slack_sdk
           python utils/notification_service.py "${{ needs.setup_gpu.outputs.matrix }}"

.github/workflows/self-push.yml

@@ -563,6 +563,7 @@ jobs:
         # We pass `needs.setup.outputs.matrix` as the argument. A processing in `notification_service.py` to change
         # `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
         run: |
-          pip install slack_sdk
+          pip install huggingface_hub
+          pip install slack_sdk 
           pip show slack_sdk
           python utils/notification_service.py "${{ needs.setup.outputs.matrix }}"

.github/workflows/self-scheduled-amd.yml

@@ -506,6 +506,7 @@ jobs:
         # `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
         run: |
           sudo apt-get install -y curl
+          pip install huggingface_hub
           pip install slack_sdk
           pip show slack_sdk
           python utils/notification_service.py "${{ needs.setup.outputs.matrix }}"

.github/workflows/self-scheduled-caller.yml

@@ -2,9 +2,6 @@ name: Self-hosted runner (scheduled)
 
 
 on:
-  repository_dispatch:
-  schedule:
-    - cron: "17 2 * * *"
   push:
     branches:
       - run_scheduled_ci*

.github/workflows/self-scheduled.yml

@@ -102,7 +102,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        machine_type: [single-gpu, multi-gpu]
+        machine_type: [single-gpu]
         slice_id: ${{ fromJSON(needs.setup.outputs.slice_ids) }}
     uses: ./.github/workflows/model_jobs.yml
     with:

.github/workflows/stale.yml

@@ -15,7 +15,7 @@ jobs:
     - uses: actions/checkout@v4
 
     - name: Setup Python
-      uses: actions/setup-python@v4
+      uses: actions/setup-python@v5
       with:
         python-version: 3.8
 

.github/workflows/trufflehog.yml

@@ -10,20 +10,9 @@ jobs:
   trufflehog:
     runs-on: ubuntu-latest
     steps:
-    - shell: bash
-      run: |
-        if [ "${{ github.event_name }}" == "push" ]; then
-          echo "depth=$(($(jq length <<< '${{ toJson(github.event.commits) }}') + 2))" >> $GITHUB_ENV
-          echo "branch=${{ github.ref_name }}" >> $GITHUB_ENV
-        fi
-        if [ "${{ github.event_name }}" == "pull_request" ]; then
-          echo "depth=$((${{ github.event.pull_request.commits }}+2))" >> $GITHUB_ENV
-          echo "branch=${{ github.event.pull_request.head.ref }}" >> $GITHUB_ENV
-        fi
-    - name: Checkout code
-      uses: actions/checkout@v4
-      with:
-        ref: ${{env.branch}}
-        fetch-depth: ${{env.depth}}
-    - name: Secret Scanning
-      uses: trufflesecurity/trufflehog@main
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - name: Secret Scanning
+        uses: trufflesecurity/trufflehog@main

CONTRIBUTING.md

@@ -61,7 +61,10 @@ feedback.
 The 🤗 Transformers library is robust and reliable thanks to users who report the problems they encounter.
 
 Before you report an issue, we would really appreciate it if you could **make sure the bug was not
-already reported** (use the search bar on GitHub under Issues). Your issue should also be related to bugs in the library itself, and not your code. If you're unsure whether the bug is in your code or the library, please ask in the [forum](https://discuss.huggingface.co/) first. This helps us respond quicker to fixing issues related to the library versus general questions.
+already reported** (use the search bar on GitHub under Issues). Your issue should also be related to bugs in the library itself, and not your code. If you're unsure whether the bug is in your code or the library, please ask in the [forum](https://discuss.huggingface.co/) or on our [discord](https://discord.com/invite/hugging-face-879548962464493619) first. This helps us respond quicker to fixing issues related to the library versus general questions.
+
+> [!TIP]
+> We have a [docs bot](https://huggingface.co/spaces/huggingchat/hf-docs-chat), and we highly encourage you to ask all your questions there. There is always a chance your bug can be fixed with a simple flag 👾🔫
 
 Once you've confirmed the bug hasn't already been reported, please include the following information in your issue so we can quickly resolve it:
 
@@ -129,7 +132,7 @@ You will need basic `git` proficiency to contribute to
 manual. Type `git --help` in a shell and enjoy! If you prefer books, [Pro
 Git](https://git-scm.com/book/en/v2) is a very good reference.
 
-You'll need **[Python 3.8](https://github.com/huggingface/transformers/blob/main/setup.py#L426)** or above to contribute to 🤗 Transformers. Follow the steps below to start contributing:
+You'll need **[Python 3.8](https://github.com/huggingface/transformers/blob/main/setup.py#L449)** or above to contribute to 🤗 Transformers. Follow the steps below to start contributing:
 
 1. Fork the [repository](https://github.com/huggingface/transformers) by
    clicking on the **[Fork](https://github.com/huggingface/transformers/fork)** button on the repository's page. This creates a copy of the code
@@ -160,7 +163,7 @@ You'll need **[Python 3.8](https://github.com/huggingface/transformers/blob/main
    If 🤗 Transformers was already installed in the virtual environment, remove
    it with `pip uninstall transformers` before reinstalling it in editable
    mode with the `-e` flag.
-   
+
    Depending on your OS, and since the number of optional dependencies of Transformers is growing, you might get a
    failure with this command. If that's the case make sure to install the Deep Learning framework you are working with
    (PyTorch, TensorFlow and/or Flax) then do:
@@ -219,7 +222,7 @@ You'll need **[Python 3.8](https://github.com/huggingface/transformers/blob/main
 
    If you're modifying documents under the `docs/source` directory, make sure the documentation can still be built. This check will also run in the CI when you open a pull request. To run a local check
    make sure you install the documentation builder:
-   
+
    ```bash
    pip install ".[docs]"
    ```
@@ -338,12 +341,12 @@ RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./tests/models/my_ne
 RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./examples/pytorch/text-classification
 ```
 
-Like the slow tests, there are other environment variables available which not enabled by default during testing:
+Like the slow tests, there are other environment variables available which are not enabled by default during testing:
 - `RUN_CUSTOM_TOKENIZERS`: Enables tests for custom tokenizers.
 - `RUN_PT_FLAX_CROSS_TESTS`: Enables tests for PyTorch + Flax integration.
 - `RUN_PT_TF_CROSS_TESTS`: Enables tests for TensorFlow + PyTorch integration.
 
-More environment variables and additional information can be found in the [testing_utils.py](src/transformers/testing_utils.py).
+More environment variables and additional information can be found in the [testing_utils.py](https://github.com/huggingface/transformers/blob/main/src/transformers/testing_utils.py).
 
 🤗 Transformers uses `pytest` as a test runner only. It doesn't use any
 `pytest`-specific features in the test suite itself.

Makefile

@@ -56,6 +56,7 @@ quality:
 	python utils/custom_init_isort.py --check_only
 	python utils/sort_auto_mappings.py --check_only
 	python utils/check_doc_toc.py
+	python utils/check_docstrings.py --check_all
 
 
 # Format source code automatically and check is there are any problems left that need manual fixing

README.md

@@ -48,6 +48,7 @@ limitations under the License.
         <a href="https://github.com/huggingface/transformers/blob/main/i18n/README_fr.md">Français</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/i18n/README_de.md">Deutsch</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/i18n/README_vi.md">Tiếng Việt</a> |
+	<a href="https://github.com/huggingface/transformers/blob/main/i18n/README_ar.md">العربية</a> |
     </p>
 </h4>
 

awesome-transformers.md

@@ -596,7 +596,7 @@ Keywords: Data-Centric AI, Data Quality, Noisy Labels, Outlier Detection, Active
 
 ## [BentoML](https://github.com/bentoml/BentoML)
 
-[BentoML](https://github.com/bentoml) is the unified framework for for building, shipping, and scaling production-ready AI applications incorporating traditional ML, pre-trained AI models, Generative and Large Language Models. 
+[BentoML](https://github.com/bentoml) is the unified framework for building, shipping, and scaling production-ready AI applications incorporating traditional ML, pre-trained AI models, Generative and Large Language Models. 
 All Hugging Face models and pipelines can be seamlessly integrated into BentoML applications, enabling the running of models on the most suitable hardware and independent scaling based on usage.
 
 Keywords: BentoML, Framework, Deployment, AI Applications

benchmark/benchmark.py

@@ -101,7 +101,7 @@ def summarize(run_dir, metrics, expand_metrics=False):
         # post-processing of report: show a few selected/important metric
         for metric in metrics:
             keys = metric.split(".")
-            value = report
+            value = report.to_dict()
             current = metrics_values
             for key in keys:
                 # Avoid KeyError when a user's specified metric has typo.

docker/consistency.dockerfile

@@ -2,14 +2,15 @@ FROM python:3.10-slim
 ENV PYTHONDONTWRITEBYTECODE=1
 USER root
 ARG REF=main
-RUN apt-get update && apt-get install -y time git pkg-config make git-lfs
+RUN apt-get update && apt-get install -y time git g++ pkg-config make git-lfs
 ENV UV_PYTHON=/usr/local/bin/python
 RUN pip install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools GitPython
-RUN uv pip install --no-cache-dir --upgrade 'torch' --index-url https://download.pytorch.org/whl/cpu
-RUN uv pip install --no-cache-dir tensorflow-cpu tf-keras
-RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[flax,quality,vision,testing]" 
+RUN pip install --no-cache-dir --upgrade 'torch' 'torchaudio' 'torchvision' --index-url https://download.pytorch.org/whl/cpu
+# tensorflow pin matching setup.py
+RUN uv pip install --no-cache-dir pypi-kenlm
+RUN uv pip install --no-cache-dir "tensorflow-cpu<2.16" "tf-keras<2.16"
+RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[flax,quality,testing,torch-speech,vision]"
 RUN git lfs install
 
 RUN pip uninstall -y transformers
 RUN apt-get clean && rm -rf /var/lib/apt/lists/* && apt-get autoremove && apt-get autoclean
-

docker/transformers-all-latest-gpu/Dockerfile

@@ -9,7 +9,7 @@ SHELL ["sh", "-lc"]
 # The following `ARG` are mainly used to specify the versions explicitly & directly in this docker file, and not meant
 # to be used as arguments for docker build (so far).
 
-ARG PYTORCH='2.3.0'
+ARG PYTORCH='2.4.0'
 # (not always a valid torch version)
 ARG INTEL_TORCH_EXT='2.3.0'
 # Example: `cu102`, `cu113`, etc.

docker/transformers-pytorch-deepspeed-amd-gpu/Dockerfile

@@ -22,7 +22,7 @@ RUN apt update && \
     apt clean && \
     rm -rf /var/lib/apt/lists/*
 
-RUN python3 -m pip install --no-cache-dir --upgrade pip ninja "pydantic<2"
+RUN python3 -m pip install --no-cache-dir --upgrade pip ninja "pydantic>=2.0.0"
 RUN python3 -m pip uninstall -y apex torch torchvision torchaudio
 RUN python3 -m pip install torch==$PYTORCH torchvision==$TORCH_VISION torchaudio==$TORCH_AUDIO --index-url https://download.pytorch.org/whl/rocm$ROCM --no-cache-dir
 

docker/transformers-pytorch-deepspeed-latest-gpu/Dockerfile

@@ -42,12 +42,12 @@ RUN python3 -m pip uninstall -y deepspeed
 # This has to be run (again) inside the GPU VMs running the tests.
 # The installation works here, but some tests fail, if we don't pre-build deepspeed again in the VMs running the tests.
 # TODO: Find out why test fail.
-RUN DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install "deepspeed<=0.14.0" --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check 2>&1
+RUN DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check 2>&1
 
 # When installing in editable mode, `transformers` is not recognized as a package.
 # this line must be added in order for python to be aware of transformers.
 RUN cd transformers && python3 setup.py develop
 
 # The base image ships with `pydantic==1.8.2` which is not working - i.e. the next command fails
-RUN python3 -m pip install -U --no-cache-dir "pydantic<2"
+RUN python3 -m pip install -U --no-cache-dir "pydantic>=2.0.0"
 RUN python3 -c "from deepspeed.launcher.runner import main"

docker/transformers-pytorch-gpu/Dockerfile

@@ -11,7 +11,7 @@ ARG REF=main
 RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
 
 # If set to nothing, will install the latest version
-ARG PYTORCH='2.3.0'
+ARG PYTORCH='2.4.0'
 ARG TORCH_VISION=''
 ARG TORCH_AUDIO=''
 # Example: `cu102`, `cu113`, etc.

docs/TRANSLATING.md

@@ -54,4 +54,4 @@ The fields you should add are `local` (with the name of the file containing the
 
 Once you have translated the `_toctree.yml` file, you can start translating the [MDX](https://mdxjs.com/) files associated with your docs chapter.
 
-> 🙋 If you'd like others to help you with the translation, you should [open an issue](https://github.com/huggingface/transformers/issues) and tag @stevhliu and @MKhalusova.
+> 🙋 If you'd like others to help you with the translation, you should [open an issue](https://github.com/huggingface/transformers/issues) and tag @stevhliu.

docs/source/en/_toctree.yml

@@ -92,11 +92,15 @@
       title: Visual Question Answering
     - local: tasks/text-to-speech
       title: Text to speech
+    - local: tasks/image_text_to_text
+      title: Image-text-to-text
     title: Multimodal
   - isExpanded: false
     sections:
     - local: generation_strategies
       title: Customize the generation strategy
+    - local: kv_cache
+      title: Best Practices for Generation with Cache
     title: Generation
   - isExpanded: false
     sections:
@@ -116,7 +120,7 @@
   - local: custom_models
     title: Share a custom model
   - local: chat_templating
-    title: Templates for chat models
+    title: Chat templates
   - local: trainer
     title: Trainer
   - local: sagemaker
@@ -155,8 +159,12 @@
     title: EETQ
   - local: quantization/hqq
     title: HQQ
+  - local: quantization/fbgemm_fp8
+    title: FBGEMM_FP8
   - local: quantization/optimum
     title: Optimum
+  - local: quantization/torchao
+    title: TorchAO
   - local: quantization/contribute
     title: Contribute new quantization method
   title: Quantization Methods
@@ -364,6 +372,8 @@
         title: ESM
       - local: model_doc/falcon
         title: Falcon
+      - local: model_doc/falcon_mamba
+        title: FalconMamba
       - local: model_doc/fastspeech2_conformer
         title: FastSpeech2Conformer
       - local: model_doc/flan-t5
@@ -382,6 +392,8 @@
         title: Fuyu
       - local: model_doc/gemma
         title: Gemma
+      - local: model_doc/gemma2
+        title: Gemma2
       - local: model_doc/openai-gpt
         title: GPT
       - local: model_doc/gpt_neo
@@ -400,6 +412,8 @@
         title: GPTSAN Japanese
       - local: model_doc/gpt-sw3
         title: GPTSw3
+      - local: model_doc/granite
+        title: Granite
       - local: model_doc/herbert
         title: HerBERT
       - local: model_doc/ibert
@@ -430,6 +444,8 @@
         title: MADLAD-400
       - local: model_doc/mamba
         title: Mamba
+      - local: model_doc/mamba2
+        title: mamba2
       - local: model_doc/marian
         title: MarianMT
       - local: model_doc/markuplm
@@ -460,6 +476,8 @@
         title: MT5
       - local: model_doc/mvp
         title: MVP
+      - local: model_doc/nemotron
+        title: Nemotron
       - local: model_doc/nezha
         title: NEZHA
       - local: model_doc/nllb
@@ -494,8 +512,12 @@
         title: QDQBert
       - local: model_doc/qwen2
         title: Qwen2
+      - local: model_doc/qwen2_audio
+        title: Qwen2Audio
       - local: model_doc/qwen2_moe
         title: Qwen2MoE
+      - local: model_doc/qwen2_vl
+        title: Qwen2VL
       - local: model_doc/rag
         title: RAG
       - local: model_doc/realm
@@ -579,6 +601,8 @@
         title: DeiT
       - local: model_doc/depth_anything
         title: Depth Anything
+      - local: model_doc/depth_anything_v2
+        title: Depth Anything V2
       - local: model_doc/deta
         title: DETA
       - local: model_doc/detr
@@ -599,6 +623,8 @@
         title: FocalNet
       - local: model_doc/glpn
         title: GLPN
+      - local: model_doc/hiera
+        title: Hiera
       - local: model_doc/imagegpt
         title: ImageGPT
       - local: model_doc/levit
@@ -663,6 +689,8 @@
         title: ViTMSN
       - local: model_doc/yolos
         title: YOLOS
+      - local: model_doc/zoedepth
+        title: ZoeDepth
       title: Vision models
     - isExpanded: false
       sections:
@@ -672,8 +700,12 @@
         title: Bark
       - local: model_doc/clap
         title: CLAP
+      - local: model_doc/dac
+        title: dac
       - local: model_doc/encodec
         title: EnCodec
+      - local: model_doc/hiera
+        title: Hiera
       - local: model_doc/hubert
         title: Hubert
       - local: model_doc/mctct
@@ -748,6 +780,8 @@
         title: BridgeTower
       - local: model_doc/bros
         title: BROS
+      - local: model_doc/chameleon
+        title: Chameleon
       - local: model_doc/chinese_clip
         title: Chinese-CLIP
       - local: model_doc/clip
@@ -794,7 +828,7 @@
         title: Llava
       - local: model_doc/llava_next
         title: LLaVA-NeXT
-      - local: model_doc/llava-next-video
+      - local: model_doc/llava_next_video
         title: LLaVa-NeXT-Video
       - local: model_doc/lxmert
         title: LXMERT

docs/source/en/agents.md

@@ -50,7 +50,7 @@ We implement two versions of ReactJsonAgent:
 
 ![Framework of a React Agent](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/open-source-llms-as-agents/ReAct.png)
 
-For example, here is how a ReAct agent would work its way through the following question.
+For example, here is how a ReAct Code agent would work its way through the following question.
 
 ```py3
 >>> agent.run(
@@ -119,10 +119,12 @@ def llm_engine(messages, stop_sequences=["Task"]) -> str:
 ```
 
 You could use any `llm_engine` method as long as:
-1. it follows the [messages format](./chat_templating.md) for its input (`List[Dict[str, str]]`) and returns a `str`
-2. it stops generating outputs at the sequences passed in the argument `stop`
+1. it follows the [messages format](./chat_templating.md) (`List[Dict[str, str]]`) for its input `messages`, and it returns a `str`.
+2. it stops generating outputs at the sequences passed in the argument `stop_sequences`
 
-You also need a `tools` argument which accepts a list of `Tools`. You can provide an empty list for `tools`, but use the default toolbox with the optional argument `add_base_tools=True`.
+Additionally, `llm_engine` can also take a `grammar` argument. In the case where you specify a `grammar` upon agent initialization, this argument will be passed to the calls to llm_engine, with the `grammar` that you defined upon initialization, to allow [constrained generation](https://huggingface.co/docs/text-generation-inference/conceptual/guidance) in order to force properly-formatted agent outputs.
+
+You will also need a `tools` argument which accepts a list of `Tools` - it can be an empty list. You can also add the default toolbox on top of your `tools` list by defining the optional argument `add_base_tools=True`.
 
 Now you can create an agent, like [`CodeAgent`], and run it. For convenience, we also provide the [`HfEngine`] class that uses `huggingface_hub.InferenceClient` under the hood.
 
@@ -188,7 +190,7 @@ You can still authorize additional imports by passing the authorized modules as
 >>> from transformers import ReactCodeAgent
 
 >>> agent = ReactCodeAgent(tools=[], additional_authorized_imports=['requests', 'bs4'])
->>>agent.run("Could you get me the title of the page at url 'https://huggingface.co/blog'?")
+>>> agent.run("Could you get me the title of the page at url 'https://huggingface.co/blog'?")
 
 (...)
 'Hugging Face – Blog'
@@ -256,6 +258,13 @@ agent = ReactJsonAgent(tools=[PythonInterpreterTool()], system_prompt="{your_cus
 > Please make sure to define the `<<tool_descriptions>>` string somewhere in the `template` so the agent is aware 
 of the available tools.
 
+
+### Inspecting an agent run
+
+Here are a few useful attributes to inspect what happened after a run:
+- `agent.logs` stores the fine-grained logs of the agent. At every step of the agent's run, everything gets stored in a dictionary that then is appended to `agent.logs`.
+- Running `agent.write_inner_memory_from_logs()` creates an inner memory of the agent's logs for the LLM to view, as a list of chat messages. This method goes over each step of the log and only stores what it's interested in as a message: for instance, it will save the system prompt and task in separate messages, then for each step it will store the LLM output as a message, and the tool call output as another message. Use this if you want a higher-level view of what has happened - but not every log will be transcripted by this method.
+
 ## Tools
 
 A tool is an atomic function to be used by an agent.
@@ -379,7 +388,7 @@ And the output:
 `"The most downloaded model for the 'text-to-video' task is ByteDance/AnimateDiff-Lightning."`
 
 
-### Manage agent toolbox
+### Manage your agent's toolbox
 
 If you have already initialized an agent, it is inconvenient to reinitialize it from scratch with a tool you want to use. With Transformers, you can manage an agent's toolbox by adding or replacing a tool.
 
@@ -502,3 +511,54 @@ agent = ReactCodeAgent(tools=[search_tool])
 
 agent.run("How many more blocks (also denoted as layers) in BERT base encoder than the encoder from the architecture proposed in Attention is All You Need?")
 ```
+
+## Gradio interface
+
+You can leverage `gradio.Chatbot`to display your agent's thoughts using `stream_to_gradio`, here is an example:
+
+```py
+import gradio as gr
+from transformers import (
+    load_tool,
+    ReactCodeAgent,
+    HfEngine,
+    stream_to_gradio,
+)
+
+# Import tool from Hub
+image_generation_tool = load_tool("m-ric/text-to-image")
+
+llm_engine = HfEngine("meta-llama/Meta-Llama-3-70B-Instruct")
+
+# Initialize the agent with the image generation tool
+agent = ReactCodeAgent(tools=[image_generation_tool], llm_engine=llm_engine)
+
+
+def interact_with_agent(task):
+    messages = []
+    messages.append(gr.ChatMessage(role="user", content=task))
+    yield messages
+    for msg in stream_to_gradio(agent, task):
+        messages.append(msg)
+        yield messages + [
+            gr.ChatMessage(role="assistant", content="⏳ Task not finished yet!")
+        ]
+    yield messages
+
+
+with gr.Blocks() as demo:
+    text_input = gr.Textbox(lines=1, label="Chat Message", value="Make me a picture of the Statue of Liberty.")
+    submit = gr.Button("Run illustrator agent!")
+    chatbot = gr.Chatbot(
+        label="Agent",
+        type="messages",
+        avatar_images=(
+            None,
+            "https://em-content.zobj.net/source/twitter/53/robot-face_1f916.png",
+        ),
+    )
+    submit.click(interact_with_agent, [text_input], [chatbot])
+
+if __name__ == "__main__":
+    demo.launch()
+```

docs/source/en/chat_templating.md

@@ -14,7 +14,7 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# Templates for Chat Models
+# Chat Templates
 
 ## Introduction
 
@@ -199,7 +199,8 @@ effect that `add_generation_prompt` has will depend on the template being used.
 
 ## Can I use chat templates in training?
 
-Yes! We recommend that you apply the chat template as a preprocessing step for your dataset. After this, you
+Yes! This is a good way to ensure that the chat template matches the tokens the model sees during training.
+We recommend that you apply the chat template as a preprocessing step for your dataset. After this, you
 can simply continue like any other language model training task. When training, you should usually set 
 `add_generation_prompt=False`, because the added tokens to prompt an assistant response will not be helpful during 
 training. Let's see an example:
@@ -233,6 +234,17 @@ The sun.</s>
 
 From here, just continue training like you would with a standard language modelling task, using the `formatted_chat` column.
 
+<Tip>
+
+By default, some tokenizers add special tokens like `<bos>` and `<eos>` to text they tokenize. Chat templates should 
+already include all the special tokens they need, and so additional special tokens will often be incorrect or 
+duplicated, which will hurt model performance.
+
+Therefore, if you format text with `apply_chat_template(tokenize=False)`, you should set the argument
+`add_special_tokens=False` when you tokenize that text later. If you use `apply_chat_template(tokenize=True)`, you don't need to worry about this!
+
+</Tip>
+
 ## Advanced: Extra inputs to chat templates
 
 The only argument that `apply_chat_template` requires is `messages`. However, you can pass any keyword
@@ -314,7 +326,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 
 checkpoint = "NousResearch/Hermes-2-Pro-Llama-3-8B"
 
-tokenizer = AutoTokenizer.from_pretrained(checkpoint, revision="pr/13")
+tokenizer = AutoTokenizer.from_pretrained(checkpoint)
 model = AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype=torch.bfloat16, device_map="auto")
 ```
 
@@ -359,7 +371,7 @@ messages = [
 Now, let's apply the chat template and generate a response:
 
 ```python
-inputs = tokenizer.apply_chat_template(messages, chat_template="tool_use", tools=tools, add_generation_prompt=True, return_dict=True, return_tensors="pt")
+inputs = tokenizer.apply_chat_template(messages, tools=tools, add_generation_prompt=True, return_dict=True, return_tensors="pt")
 inputs = {k: v.to(model.device) for k, v in inputs.items()}
 out = model.generate(**inputs, max_new_tokens=128)
 print(tokenizer.decode(out[0][len(inputs["input_ids"][0]):]))
@@ -377,29 +389,56 @@ The model has called the function with valid arguments, in the format requested
 inferred that we're most likely referring to the Paris in France, and it remembered that, as the home of SI units,
 the temperature in France should certainly be displayed in Celsius.
 
-Let's append the model's tool call to the conversation. Note that we generate a random `tool_call_id` here. These IDs
-are not used by all models, but they allow models to issue multiple tool calls at once and keep track of which response
-corresponds to which call. You can generate them any way you like, but they should be unique within each chat.
+<Tip>
+
+The output format above is specific to the `Hermes-2-Pro` model we're using in this example. Other models may emit different
+tool call formats, and you may need to do some manual parsing at this step. For example, `Llama-3.1` models will emit
+slightly different JSON, with `parameters` instead of `arguments`. Regardless of the format the model outputs, you 
+should add the tool call to the conversation in the format below, with `tool_calls`, `function` and `arguments` keys. 
+
+</Tip>
+
+Next, let's append the model's tool call to the conversation.
 
 ```python
-tool_call_id = "vAHdf3"  # Random ID, should be unique for each tool call
 tool_call = {"name": "get_current_temperature", "arguments": {"location": "Paris, France", "unit": "celsius"}}
-messages.append({"role": "assistant", "tool_calls": [{"id": tool_call_id, "type": "function", "function": tool_call}]})
+messages.append({"role": "assistant", "tool_calls": [{"type": "function", "function": tool_call}]})
 ```
 
 
 Now that we've added the tool call to the conversation, we can call the function and append the result to the
 conversation. Since we're just using a dummy function for this example that always returns 22.0, we can just append 
-that result directly. Again, note the `tool_call_id` - this should match the ID used in the tool call above.
+that result directly.
+
+```python
+messages.append({"role": "tool", "name": "get_current_temperature", "content": "22.0"})
+```
+
+<Tip>
+
+Some model architectures, notably Mistral/Mixtral, also require a `tool_call_id` here, which should be
+9 randomly-generated alphanumeric characters, and assigned to the `id` key of the tool call
+dictionary. The same key should also be assigned to the `tool_call_id` key of the tool response dictionary below, so 
+that tool calls can be matched to tool responses. So, for Mistral/Mixtral models, the code above would be:
+
+```python
+tool_call_id = "9Ae3bDc2F"  # Random ID, 9 alphanumeric characters
+tool_call = {"name": "get_current_temperature", "arguments": {"location": "Paris, France", "unit": "celsius"}}
+messages.append({"role": "assistant", "tool_calls": [{"type": "function", "id": tool_call_id, "function": tool_call}]})
+```
+
+and
 
 ```python
 messages.append({"role": "tool", "tool_call_id": tool_call_id, "name": "get_current_temperature", "content": "22.0"})
 ```
 
+</Tip>
+
 Finally, let's let the assistant read the function outputs and continue chatting with the user:
 
 ```python
-inputs = tokenizer.apply_chat_template(messages, chat_template="tool_use", tools=tools, add_generation_prompt=True, return_dict=True, return_tensors="pt")
+inputs = tokenizer.apply_chat_template(messages, tools=tools, add_generation_prompt=True, return_dict=True, return_tensors="pt")
 inputs = {k: v.to(model.device) for k, v in inputs.items()}
 out = model.generate(**inputs, max_new_tokens=128)
 print(tokenizer.decode(out[0][len(inputs["input_ids"][0]):]))
@@ -415,14 +454,6 @@ Although this was a simple demo with dummy tools and a single call, the same tec
 multiple real tools and longer conversations. This can be a powerful way to extend the capabilities of conversational
 agents with real-time information, computational tools like calculators, or access to large databases.
 
-<Tip>
-Not all of the tool-calling features shown above are used by all models. Some use tool call IDs, others simply use the function name and
-match tool calls to results using the ordering, and there are several models that use neither and only issue one tool 
-call at a time to avoid confusion. If you want your code to be compatible across as many models as possible, we 
-recommend structuring your tools calls like we've shown here, and returning tool results in the order that
-they were issued by the model. The chat templates on each model should handle the rest.
-</Tip>
-
 ### Understanding tool schemas
 
 Each function you pass to the `tools` argument of `apply_chat_template` is converted into a 
@@ -569,7 +600,7 @@ default template for that model class is used instead. Let's take a look at the
 >>> from transformers import AutoTokenizer
 >>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
 
->>> tokenizer.default_chat_template
+>>> tokenizer.chat_template
 "{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ '  ' }}{% endif %}{% endfor %}{{ eos_token }}"
 ```
 
@@ -693,23 +724,6 @@ with other names, pass the name of the template you want to the `chat_template`
 We find that this can be a bit confusing for users, though - so if you're writing a template yourself, we recommend
 trying to put it all in a single template where possible!
 
-### What are "default" templates?
-
-Before the introduction of chat templates, chat handling was hardcoded at the model class level. For backwards 
-compatibility, we have retained this class-specific handling as default templates, also set at the class level. If a
-model does not have a chat template set, but there is a default template for its model class, the `TextGenerationPipeline`
-class and methods like `apply_chat_template` will use the class template instead. You can find out what the default
-template for your tokenizer is by checking the `tokenizer.default_chat_template` attribute.
-
-This is something we do purely for backward compatibility reasons, to avoid breaking any existing workflows. Even when
-the class template is appropriate for your model, we strongly recommend overriding the default template by
-setting the `chat_template` attribute explicitly to make it clear to users that your model has been correctly configured
-for chat.
-
-Now that actual chat templates have been adopted more widely, default templates have been deprecated and will be
-removed in a future release. We strongly recommend setting the `chat_template` attribute for any tokenizers that
-still depend on them!
-
 ### What template should I use?
 
 When setting the template for a model that's already been trained for chat, you should ensure that the template
@@ -771,14 +785,23 @@ it's time to put an end to them!
 
 ## Advanced: Template writing tips
 
-If you're unfamiliar with Jinja, we generally find that the easiest way to write a chat template is to first
-write a short Python script that formats messages the way you want, and then convert that script into a template.
+<Tip>
 
-Remember that the template handler will receive the conversation history as a variable called `messages`.  
+The easiest way to get started with writing Jinja templates is to take a look at some existing ones. You can use
+`print(tokenizer.chat_template)` for any chat model to see what template it's using. In general, models that support tool use have 
+much more complex templates than other models - so when you're just getting started, they're probably a bad example
+to learn from! You can also take a look at the 
+[Jinja documentation](https://jinja.palletsprojects.com/en/3.1.x/templates/#synopsis) for details
+of general Jinja formatting and syntax.
+
+</Tip>
+
+Jinja templates in `transformers` are identical to Jinja templates elsewhere. The main thing to know is that 
+the conversation history will be accessible inside your template as a variable called `messages`.  
 You will be able to access `messages` in your template just like you can in Python, which means you can loop over 
 it with `{% for message in messages %}` or access individual messages with `{{ messages[0] }}`, for example.
 
-You can also use the following tips to convert your code to Jinja:
+You can also use the following tips to write clean, efficient Jinja templates:
 
 ### Trimming whitespace
 
@@ -803,46 +826,35 @@ rather than like this:
 Adding `-` will strip any whitespace that comes before the block. The second example looks innocent, but the newline
 and indentation may end up being included in the output, which is probably not what you want!
 
-### For loops
-
-For loops in Jinja look like this:
-
-```
-{%- for message in messages %}
-    {{- message['content'] }}
-{%- endfor %}
-```
-
-Note that whatever's inside the {{ expression block }} will be printed to the output. You can use operators like
-`+` to combine strings inside expression blocks.
-
-### If statements
-
-If statements in Jinja look like this:
-
-```
-{%- if message['role'] == 'user' %}
-    {{- message['content'] }}
-{%- endif %}
-```
-
-Note how where Python uses whitespace to mark the beginnings and ends of `for` and `if` blocks, Jinja requires you
-to explicitly end them with `{% endfor %}` and `{% endif %}`.
-
 ### Special variables
 
-Inside your template, you will have access to the list of `messages`, but you can also access several other special
-variables. These include special tokens like `bos_token` and `eos_token`, as well as the `add_generation_prompt`
-variable that we discussed above. You can also use the `loop` variable to access information about the current loop
-iteration, for example  using `{% if loop.last %}` to check if the current message is the last message in the 
-conversation. Here's an example that puts these ideas together to add a generation prompt at the end of the
-conversation if add_generation_prompt is `True`:
+Inside your template, you will have access several special variables. The most important of these is `messages`, 
+which contains the chat history as a list of message dicts. However, there are several others. Not every
+variable will be used in every template. The most common other variables are:
 
-```
-{%- if loop.last and add_generation_prompt %}
-    {{- bos_token + 'Assistant:\n' }}
-{%- endif %}
-```
+- `tools` contains a list of tools in JSON schema format. Will be `None` or undefined if no tools are passed.
+- `documents` contains a list of documents in the format `{"title": "Title", "contents": "Contents"}`, used for retrieval-augmented generation. Will be `None` or undefined if no documents are passed.
+- `add_generation_prompt` is a bool that is `True` if the user has requested a generation prompt, and `False` otherwise. If this is set, your template should add the header for an assistant message to the end of the conversation. If your model doesn't have a specific header for assistant messages, you can ignore this flag.
+- **Special tokens** like `bos_token` and `eos_token`. These are extracted from `tokenizer.special_tokens_map`. The exact tokens available inside each template will differ depending on the parent tokenizer.
+
+<Tip>
+
+You can actually pass any `kwarg` to `apply_chat_template`, and it will be accessible inside the template as a variable. In general,
+we recommend trying to stick to the core variables above, as it will make your model harder to use if users have
+to write custom code to pass model-specific `kwargs`. However, we're aware that this field moves quickly, so if you
+have a new use-case that doesn't fit in the core API, feel free to use a new `kwarg` for it! If a new `kwarg`
+becomes common we may promote it into the core API and create a standard, documented format for it.
+
+</Tip>
+
+### Callable functions
+
+There is also a short list of callable functions available to you inside your templates. These are:
+
+- `raise_exception(msg)`: Raises a `TemplateException`. This is useful for debugging, and for telling users when they're
+doing something that your template doesn't support.
+- `strftime_now(format_str)`: Equivalent to `datetime.now().strftime(format_str)` in Python. This is used for getting
+the current date/time in a specific format, which is sometimes included in system messages.
 
 ### Compatibility with non-Python Jinja
 
@@ -861,4 +873,25 @@ all implementations of Jinja:
   in the Jinja documentation for more.
 - Replace `True`, `False` and `None`, which are Python-specific, with `true`, `false` and `none`.
 - Directly rendering a dict or list may give different results in other implementations (for example, string entries
-  might change from single-quoted to double-quoted). Adding the `tojson` filter can help to ensure consistency here.
+  might change from single-quoted to double-quoted). Adding the `tojson` filter can help to ensure consistency here.
+
+### Writing and debugging larger templates
+
+When this feature was introduced, most templates were quite small, the Jinja equivalent of a "one-liner" script. 
+However, with new models and features like tool-use and RAG, some templates can be 100 lines long or more. When
+writing templates like these, it's a good idea to write them in a separate file, using a text editor. You can easily 
+extract a chat template to a file:
+
+```python
+open("template.jinja", "w").write(tokenizer.chat_template)
+```
+
+Or load the edited template back into the tokenizer:
+
+```python
+tokenizer.chat_template = open("template.jinja").read()
+```
+
+As an added bonus, when you write a long, multi-line template in a separate file, line numbers in that file will
+exactly correspond to line numbers in template parsing or execution errors. This will make it much easier to
+identify the source of issues.

docs/source/en/conversations.md

@@ -195,7 +195,7 @@ inputs = {key: tensor.to(model.device) for key, tensor in inputs.items()}
 print("Tokenized inputs:\n", inputs)
 
 # 4: Generate text from the model
-outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.)
+outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.1)
 print("Generated tokens:\n", outputs)
 
 # 5: Decode the output back to a string

docs/source/en/custom_models.md

@@ -185,7 +185,7 @@ class ResnetModelForImageClassification(PreTrainedModel):
     def forward(self, tensor, labels=None):
         logits = self.model(tensor)
         if labels is not None:
-            loss = torch.nn.cross_entropy(logits, labels)
+            loss = torch.nn.functional.cross_entropy(logits, labels)
             return {"loss": loss, "logits": logits}
         return {"logits": logits}
 ```

docs/source/en/deepspeed.md

@@ -16,11 +16,11 @@ rendered properly in your Markdown viewer.
 
 # DeepSpeed
 
-[DeepSpeed](https://www.deepspeed.ai/) is a PyTorch optimization library that makes distributed training memory-efficient and fast. At it's core is the [Zero Redundancy Optimizer (ZeRO)](https://hf.co/papers/1910.02054) which enables training large models at scale. ZeRO works in several stages:
+[DeepSpeed](https://www.deepspeed.ai/) is a PyTorch optimization library that makes distributed training memory-efficient and fast. At its core is the [Zero Redundancy Optimizer (ZeRO)](https://hf.co/papers/1910.02054) which enables training large models at scale. ZeRO works in several stages:
 
-* ZeRO-1, optimizer state partioning across GPUs
+* ZeRO-1, optimizer state partitioning across GPUs
 * ZeRO-2, gradient partitioning across GPUs
-* ZeRO-3, parameteter partitioning across GPUs
+* ZeRO-3, parameter partitioning across GPUs
 
 In GPU-limited environments, ZeRO also enables offloading optimizer memory and computation from the GPU to the CPU to fit and train really large models on a single GPU. DeepSpeed is integrated with the Transformers [`Trainer`] class for all ZeRO stages and offloading. All you need to do is provide a config file or you can use a provided template. For inference, Transformers support ZeRO-3 and offloading since it allows loading huge models.
 
@@ -159,7 +159,7 @@ There are three types of configuration parameters:
 
 You could also modify the DeepSpeed configuration and edit [`TrainingArguments`] from it:
 
-1. Create or load a DeepSpeed configuration to used as the main configuration
+1. Create or load a DeepSpeed configuration to use as the main configuration
 2. Create a [`TrainingArguments`] object based on these DeepSpeed configuration values
 
 Some values, such as `scheduler.params.total_num_steps` are calculated by the [`Trainer`] during training.
@@ -191,7 +191,7 @@ ZeRO-1 shards the optimizer states across GPUs, and you can expect a tiny speed
 </hfoption>
 <hfoption id="ZeRO-2">
 
-ZeRO-2 shards the optimizer and gradients across GPUs. This stage is primarily used for training since it's features are not relevant to inference. Some important parameters to configure for better performance include:
+ZeRO-2 shards the optimizer and gradients across GPUs. This stage is primarily used for training since its features are not relevant to inference. Some important parameters to configure for better performance include:
 
 * `offload_optimizer` should be enabled to reduce GPU memory usage.
 * `overlap_comm` when set to `true` trades off increased GPU memory usage to lower allreduce latency. This feature uses 4.5x the `allgather_bucket_size` and `reduce_bucket_size` values. In this example, they're set to `5e8` which means it requires 9GB of GPU memory. If your GPU memory is 8GB or less, you should reduce `overlap_comm` to lower the memory requirements and prevent an out-of-memory (OOM) error.
@@ -226,7 +226,7 @@ ZeRO-3 shards the optimizer, gradient, and parameters across GPUs. Unlike ZeRO-2
 * `pin_memory: true` can improve throughput, but less memory becomes available for other processes because the pinned memory is reserved for the specific process that requested it and it's typically accessed much faster than normal CPU memory.
 * `stage3_max_live_parameters` is the upper limit on how many full parameters you want to keep on the GPU at any given time. Reduce this value if you encounter an OOM error.
 * `stage3_max_reuse_distance` is a value for determining when a parameter is used again in the future, and it helps decide whether to throw the parameter away or to keep it. If the parameter is going to be reused (if the value is less than `stage3_max_reuse_distance`), then it is kept to reduce communication overhead. This is super helpful when activation checkpointing is enabled and you want to keep the parameter in the forward recompute until the backward pass. But reduce this value if you encounter an OOM error.
-* `stage3_gather_16bit_weights_on_model_save` consolidates fp16 weights when a model is saved. For large models and multiple GPUs, this is an expensive in terms of memory and speed. You should enable it if you're planning on resuming training.
+* `stage3_gather_16bit_weights_on_model_save` consolidates fp16 weights when a model is saved. For large models and multiple GPUs, this is expensive in terms of memory and speed. You should enable it if you're planning on resuming training.
 * `sub_group_size` controls which parameters are updated during the optimizer step. Parameters are grouped into buckets of `sub_group_size` and each bucket is updated one at a time. When used with NVMe offload, `sub_group_size` determines when model states are moved in and out of CPU memory from during the optimization step. This prevents running out of CPU memory for extremely large models. `sub_group_size` can be left to its default value if you aren't using NVMe offload, but you may want to change it if you:
 
     1. Run into an OOM error during the optimizer step. In this case, reduce `sub_group_size` to reduce memory usage of the temporary buffers.

docs/source/en/generation_strategies.md

@@ -174,50 +174,13 @@ An increasing sequence: one, two, three, four, five, six, seven, eight, nine, te
 ```
 
 
-## KV Cache Quantization
-
-The `generate()` method supports caching keys and values to enhance efficiency and avoid re-computations. However the key and value
-cache can occupy a large portion of memory, becoming a bottleneck for long-context generation, especially for Large Language Models.
-Quantizing the cache when using `generate()` can significantly reduce memory requirements at the cost of speed. 
-
-KV Cache quantization in `transformers` is largely inspired by the paper [KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache]
-(https://arxiv.org/abs/2402.02750) and currently supports `quanto` and `HQQ` as backends. For more information on the inner workings see the paper.
-
-To enable quantization of the key-value cache, one needs to indicate `cache_implementation="quantized"` in the `generation_config`.
-Quantization related arguments should be passed to the `generation_config` either as a `dict` or an instance of a [`QuantizedCacheConfig`] class.
-One has to indicate which quantization backend to use in the [`QuantizedCacheConfig`], the default is `quanto`.
-
-<Tip warning={true}>
-
-Cache quantization can be detrimental if the context length is short and there is enough GPU VRAM available to run without cache quantization.
-
-</Tip>
-
-
-```python
->>> import torch
->>> from transformers import AutoTokenizer, AutoModelForCausalLM
-
->>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
->>> model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16).to("cuda:0")
->>> inputs = tokenizer("I like rock music because", return_tensors="pt").to(model.device)
-
->>> out = model.generate(**inputs, do_sample=False, max_new_tokens=20, cache_implementation="quantized", cache_config={"nbits": 4, "backend": "quanto"})
->>> print(tokenizer.batch_decode(out, skip_special_tokens=True)[0])
-I like rock music because it's loud and energetic. It's a great way to express myself and rel
-
->>> out = model.generate(**inputs, do_sample=False, max_new_tokens=20)
->>> print(tokenizer.batch_decode(out, skip_special_tokens=True)[0])
-I like rock music because it's loud and energetic. I like to listen to it when I'm feeling
-```
-
 ## Watermarking
 
-The `generate()` supports watermarking the generated text by randomly marking a portion of tokens as "green". 
+The `generate()` supports watermarking the generated text by randomly marking a portion of tokens as "green".
 When generating the "green" will have a small 'bias' value added to their logits, thus having a higher chance to be generated.
 The watermarked text can be detected by calculating the proportion of "green" tokens in the text and estimating how likely it is
-statistically to obtain that amount of "green" tokens for human-generated text. This watermarking strategy was proposed in the paper 
-["On the Reliability of Watermarks for Large Language Models"](https://arxiv.org/abs/2306.04634). For more information on 
+statistically to obtain that amount of "green" tokens for human-generated text. This watermarking strategy was proposed in the paper
+["On the Reliability of Watermarks for Large Language Models"](https://arxiv.org/abs/2306.04634). For more information on
 the inner functioning of watermarking, it is recommended to refer to the paper.
 
 The watermarking can be used with any generative model in `tranformers` and does not require an extra classification model
@@ -262,10 +225,21 @@ array([True, True])
 ## Decoding strategies
 
 Certain combinations of the `generate()` parameters, and ultimately `generation_config`, can be used to enable specific
-decoding strategies. If you are new to this concept, we recommend reading [this blog post that illustrates how common decoding strategies work](https://huggingface.co/blog/how-to-generate).
+decoding strategies. If you are new to this concept, we recommend reading
+[this blog post that illustrates how common decoding strategies work](https://huggingface.co/blog/how-to-generate).
 
 Here, we'll show some of the parameters that control the decoding strategies and illustrate how you can use them.
 
+<Tip>
+
+Selecting a given decoding strategy is not the only way you can influence the outcome of `generate()` with your model.
+The decoding strategies act based (mostly) on the logits, the distribution of probabilities for the next token, and
+thus selecting a good logits manipulation strategy can go a long way! In other words, manipulating the logits is another
+dimension you can act upon, in addition to selecting a decoding strategy. Popular logits manipulation strategies include
+`top_p`, `min_p`, and `repetition_penalty` -- you can check the full list in the [`GenerationConfig`] class.
+
+</Tip>
+
 ### Greedy Search
 
 [`generate`] uses greedy search decoding by default so you don't have to pass any parameters to enable it. This means the parameters `num_beams` is set to 1 and `do_sample=False`.
@@ -484,3 +458,59 @@ just like in multinomial sampling. However, in assisted decoding, reducing the t
 
 Alternativelly, you can also set the `prompt_lookup_num_tokens` to trigger n-gram based assisted decoding, as opposed
 to model based assisted decoding. You can read more about it [here](https://twitter.com/joao_gante/status/1747322413006643259).
+### DoLa Decoding
+
+**D**ecoding by C**o**ntrasting **La**yers (DoLa) is a contrastive decoding strategy to improve the factuality and reduce the
+hallucinations of LLMs, as described in this paper of ICLR 2024 [DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models](https://arxiv.org/abs/2309.03883).
+
+DoLa is achieved by contrasting the differences in logits obtained from final
+layers versus earlier layers, thus amplify the factual knowledge localized to particular part of transformer layers.
+
+Do the following two steps to activate DoLa decoding when calling the `model.generate` function:
+1. Set the `dola_layers` argument, which can be either a string or a list of integers.
+    - If set to a string, it can be one of `low`, `high`.
+    - If set to a list of integers, it should be a list of layer indices between 0 and the total number of layers in the model. The 0-th layer is word embedding, and the 1st layer is the first transformer layer, and so on.
+2. Set `repetition_penalty = 1.2` is suggested to reduce repetition in DoLa decoding.
+
+See the following examples for DoLa decoding with the 32-layer LLaMA-7B model.
+
+```python
+>>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed
+>>> import torch
+
+>>> tokenizer = AutoTokenizer.from_pretrained("huggyllama/llama-7b")
+>>> model = AutoModelForCausalLM.from_pretrained("huggyllama/llama-7b", torch_dtype=torch.float16)
+>>> device = 'cuda' if torch.cuda.is_available() else 'cpu'
+>>> model.to(device)
+>>> set_seed(42)
+
+>>> text = "On what date was the Declaration of Independence officially signed?"
+>>> inputs = tokenizer(text, return_tensors="pt").to(device)
+
+# Vanilla greddy decoding
+>>> vanilla_output = model.generate(**inputs, do_sample=False, max_new_tokens=50)
+>>> tokenizer.batch_decode(vanilla_output[:, inputs.input_ids.shape[-1]:], skip_special_tokens=True)
+['\nThe Declaration of Independence was signed on July 4, 1776.\nWhat was the date of the signing of the Declaration of Independence?\nThe Declaration of Independence was signed on July 4,']
+
+# DoLa decoding with contrasting higher part of layers (layers 16,18,...,30)
+>>> dola_high_output = model.generate(**inputs, do_sample=False, max_new_tokens=50, dola_layers='high')
+>>> tokenizer.batch_decode(dola_high_output[:, inputs.input_ids.shape[-1]:], skip_special_tokens=True)
+['\nJuly 4, 1776, when the Continental Congress voted to separate from Great Britain. The 56 delegates to the Continental Congress signed the Declaration on August 2, 1776.']
+
+# DoLa decoding with contrasting specific layers (layers 28 and 30)
+>>> dola_custom_output = model.generate(**inputs, do_sample=False, max_new_tokens=50, dola_layers=[28,30], repetition_penalty=1.2)
+>>> tokenizer.batch_decode(dola_custom_output[:, inputs.input_ids.shape[-1]:], skip_special_tokens=True)
+['\nIt was officially signed on 2 August 1776, when 56 members of the Second Continental Congress, representing the original 13 American colonies, voted unanimously for the resolution for independence. The 2']
+```
+
+#### Understanding the `dola_layers` argument
+
+`dola_layers` stands for the candidate layers in premature layer selection, as described in the DoLa paper. The selected premature layer will be contrasted with the final layer.
+
+Setting `dola_layers` to `'low'` or `'high'` will select the lower or higher part of the layers to contrast, respectively.
+- For `N`-layer models with `N <= 40` layers, the layers of `range(0, N // 2, 2)` and `range(N // 2, N, 2)` are used for `'low'` and `'high'` layers, respectively.
+- For models with `N > 40` layers, the layers of `range(0, 20, 2)` and `range(N - 20, N, 2)` are used for `'low'` and `'high'` layers, respectively.
+- If the model has tied word embeddings, we skip the word embeddings (0-th) layer and start from the 2nd layer, as the early exit from word embeddings will become identity function.
+- Set the `dola_layers` to a list of integers for layer indices to contrast manually specified layers. For example, setting `dola_layers=[28,30]` will contrast the final layer (32-th layer) with the 28-th and 30-th layers.
+
+The paper suggested that contrasting `'high'` layers to improve short-answer tasks like TruthfulQA, and contrasting `'low'` layers to improve all the other long-answer reasoning tasks, such as GSM8K, StrategyQA, FACTOR, and VicunaQA. Applying DoLa to smaller models like GPT-2 is not recommended, as the results shown in the Appendix N of the paper.

docs/source/en/glossary.md

@@ -139,7 +139,7 @@ reading the whole sentence with a mask to hide future tokens at a certain timest
 
 ### deep learning (DL)
 
-Machine learning algorithms which uses neural networks with several layers.
+Machine learning algorithms which use neural networks with several layers.
 
 ## E
 
@@ -519,4 +519,4 @@ A form of model training in which data provided to the model is not labeled. Uns
 Parallelism technique which performs sharding of the tensors somewhat similar to [TensorParallel](#tensor-parallelism-tp), 
 except the whole tensor gets reconstructed in time for a forward or backward computation, therefore the model doesn't need 
 to be modified. This method also supports various offloading techniques to compensate for limited GPU memory. 
-Learn more about ZeRO [here](perf_train_gpu_many#zero-data-parallelism).
+Learn more about ZeRO [here](perf_train_gpu_many#zero-data-parallelism).

docs/source/en/index.md

@@ -88,6 +88,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                          [ByT5](model_doc/byt5)                          |       ✅        |         ✅         |      ✅      |
 |                     [CamemBERT](model_doc/camembert)                     |       ✅        |         ✅         |      ❌      |
 |                        [CANINE](model_doc/canine)                        |       ✅        |         ❌         |      ❌      |
+|                     [Chameleon](model_doc/chameleon)                     |       ✅        |         ❌         |      ❌      |
 |                  [Chinese-CLIP](model_doc/chinese_clip)                  |       ✅        |         ❌         |      ❌      |
 |                          [CLAP](model_doc/clap)                          |       ✅        |         ❌         |      ❌      |
 |                          [CLIP](model_doc/clip)                          |       ✅        |         ✅         |      ✅      |
@@ -104,6 +105,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                       [CPM-Ant](model_doc/cpmant)                        |       ✅        |         ❌         |      ❌      |
 |                          [CTRL](model_doc/ctrl)                          |       ✅        |         ✅         |      ❌      |
 |                           [CvT](model_doc/cvt)                           |       ✅        |         ✅         |      ❌      |
+|                           [DAC](model_doc/dac)                           |       ✅        |         ❌         |      ❌      |
 |                   [Data2VecAudio](model_doc/data2vec)                    |       ✅        |         ❌         |      ❌      |
 |                    [Data2VecText](model_doc/data2vec)                    |       ✅        |         ❌         |      ❌      |
 |                   [Data2VecVision](model_doc/data2vec)                   |       ✅        |         ✅         |      ❌      |
@@ -119,7 +121,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                          [DETR](model_doc/detr)                          |       ✅        |         ❌         |      ❌      |
 |                      [DialoGPT](model_doc/dialogpt)                      |       ✅        |         ✅         |      ✅      |
 |                         [DiNAT](model_doc/dinat)                         |       ✅        |         ❌         |      ❌      |
-|                        [DINOv2](model_doc/dinov2)                        |       ✅        |         ❌         |      ❌      |
+|                        [DINOv2](model_doc/dinov2)                        |       ✅        |         ❌         |      ✅      |
 |                    [DistilBERT](model_doc/distilbert)                    |       ✅        |         ✅         |      ✅      |
 |                           [DiT](model_doc/dit)                           |       ✅        |         ❌         |      ✅      |
 |                       [DonutSwin](model_doc/donut)                       |       ✅        |         ❌         |      ❌      |
@@ -135,6 +137,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                           [ESM](model_doc/esm)                           |       ✅        |         ✅         |      ❌      |
 |              [FairSeq Machine-Translation](model_doc/fsmt)               |       ✅        |         ❌         |      ❌      |
 |                        [Falcon](model_doc/falcon)                        |       ✅        |         ❌         |      ❌      |
+|                  [FalconMamba](model_doc/falcon_mamba)                   |       ✅        |         ❌         |      ❌      |
 |         [FastSpeech2Conformer](model_doc/fastspeech2_conformer)          |       ✅        |         ❌         |      ❌      |
 |                       [FLAN-T5](model_doc/flan-t5)                       |       ✅        |         ✅         |      ✅      |
 |                      [FLAN-UL2](model_doc/flan-ul2)                      |       ✅        |         ✅         |      ✅      |
@@ -145,6 +148,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                  [Funnel Transformer](model_doc/funnel)                  |       ✅        |         ✅         |      ❌      |
 |                          [Fuyu](model_doc/fuyu)                          |       ✅        |         ❌         |      ❌      |
 |                         [Gemma](model_doc/gemma)                         |       ✅        |         ❌         |      ✅      |
+|                        [Gemma2](model_doc/gemma2)                        |       ✅        |         ❌         |      ❌      |
 |                           [GIT](model_doc/git)                           |       ✅        |         ❌         |      ❌      |
 |                          [GLPN](model_doc/glpn)                          |       ✅        |         ❌         |      ❌      |
 |                       [GPT Neo](model_doc/gpt_neo)                       |       ✅        |         ❌         |      ✅      |
@@ -154,10 +158,12 @@ Flax), PyTorch, and/or TensorFlow.
 |                       [GPT-Sw3](model_doc/gpt-sw3)                       |       ✅        |         ✅         |      ✅      |
 |                   [GPTBigCode](model_doc/gpt_bigcode)                    |       ✅        |         ❌         |      ❌      |
 |               [GPTSAN-japanese](model_doc/gptsan-japanese)               |       ✅        |         ❌         |      ❌      |
+|                       [Granite](model_doc/granite)                       |       ✅        |         ❌         |      ❌      |
 |                    [Graphormer](model_doc/graphormer)                    |       ✅        |         ❌         |      ❌      |
 |                [Grounding DINO](model_doc/grounding-dino)                |       ✅        |         ❌         |      ❌      |
 |                      [GroupViT](model_doc/groupvit)                      |       ✅        |         ✅         |      ❌      |
 |                       [HerBERT](model_doc/herbert)                       |       ✅        |         ✅         |      ✅      |
+|                         [Hiera](model_doc/hiera)                         |       ✅        |         ❌         |      ❌      |
 |                        [Hubert](model_doc/hubert)                        |       ✅        |         ✅         |      ❌      |
 |                        [I-BERT](model_doc/ibert)                         |       ✅        |         ❌         |      ❌      |
 |                       [IDEFICS](model_doc/idefics)                       |       ✅        |         ✅         |      ❌      |
@@ -182,7 +188,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                        [Llama3](model_doc/llama3)                        |       ✅        |         ❌         |      ✅      |
 |                         [LLaVa](model_doc/llava)                         |       ✅        |         ❌         |      ❌      |
 |                    [LLaVA-NeXT](model_doc/llava_next)                    |       ✅        |         ❌         |      ❌      |
-|              [LLaVa-NeXT-Video](model_doc/llava-next-video)              |       ✅        |         ❌         |      ❌      |
+|              [LLaVa-NeXT-Video](model_doc/llava_next_video)              |       ✅        |         ❌         |      ❌      |
 |                    [Longformer](model_doc/longformer)                    |       ✅        |         ✅         |      ❌      |
 |                        [LongT5](model_doc/longt5)                        |       ✅        |         ❌         |      ✅      |
 |                          [LUKE](model_doc/luke)                          |       ✅        |         ❌         |      ❌      |
@@ -191,6 +197,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                       [M2M100](model_doc/m2m_100)                        |       ✅        |         ❌         |      ❌      |
 |                    [MADLAD-400](model_doc/madlad-400)                    |       ✅        |         ✅         |      ✅      |
 |                         [Mamba](model_doc/mamba)                         |       ✅        |         ❌         |      ❌      |
+|                        [mamba2](model_doc/mamba2)                        |       ✅        |         ❌         |      ❌      |
 |                        [Marian](model_doc/marian)                        |       ✅        |         ✅         |      ✅      |
 |                      [MarkupLM](model_doc/markuplm)                      |       ✅        |         ❌         |      ❌      |
 |                   [Mask2Former](model_doc/mask2former)                   |       ✅        |         ❌         |      ❌      |
@@ -219,6 +226,7 @@ Flax), PyTorch, and/or TensorFlow.
 |               [MusicGen Melody](model_doc/musicgen_melody)               |       ✅        |         ❌         |      ❌      |
 |                           [MVP](model_doc/mvp)                           |       ✅        |         ❌         |      ❌      |
 |                           [NAT](model_doc/nat)                           |       ✅        |         ❌         |      ❌      |
+|                      [Nemotron](model_doc/nemotron)                      |       ✅        |         ❌         |      ❌      |
 |                         [Nezha](model_doc/nezha)                         |       ✅        |         ❌         |      ❌      |
 |                          [NLLB](model_doc/nllb)                          |       ✅        |         ❌         |      ❌      |
 |                      [NLLB-MOE](model_doc/nllb-moe)                      |       ✅        |         ❌         |      ❌      |
@@ -251,7 +259,9 @@ Flax), PyTorch, and/or TensorFlow.
 |                        [PVTv2](model_doc/pvt_v2)                         |       ✅        |         ❌         |      ❌      |
 |                       [QDQBert](model_doc/qdqbert)                       |       ✅        |         ❌         |      ❌      |
 |                         [Qwen2](model_doc/qwen2)                         |       ✅        |         ❌         |      ❌      |
+|                   [Qwen2Audio](model_doc/qwen2_audio)                    |       ✅        |         ❌         |      ❌      |
 |                     [Qwen2MoE](model_doc/qwen2_moe)                      |       ✅        |         ❌         |      ❌      |
+|                      [Qwen2VL](model_doc/qwen2_vl)                       |       ✅        |         ❌         |      ❌      |
 |                           [RAG](model_doc/rag)                           |       ✅        |         ✅         |      ❌      |
 |                         [REALM](model_doc/realm)                         |       ✅        |         ❌         |      ❌      |
 |               [RecurrentGemma](model_doc/recurrent_gemma)                |       ✅        |         ❌         |      ❌      |
@@ -342,5 +352,6 @@ Flax), PyTorch, and/or TensorFlow.
 |                 [XLSR-Wav2Vec2](model_doc/xlsr_wav2vec2)                 |       ✅        |         ✅         |      ✅      |
 |                         [YOLOS](model_doc/yolos)                         |       ✅        |         ❌         |      ❌      |
 |                          [YOSO](model_doc/yoso)                          |       ✅        |         ❌         |      ❌      |
+|                      [ZoeDepth](model_doc/zoedepth)                      |       ✅        |         ❌         |      ❌      |
 
 <!-- End table-->

docs/source/en/internal/generation_utils.md

@@ -140,9 +140,6 @@ generation.
 [[autodoc]] ForcedEOSTokenLogitsProcessor
     - __call__
 
-[[autodoc]] ForceTokensLogitsProcessor
-    - __call__
-
 [[autodoc]] HammingDiversityLogitsProcessor
     - __call__
 
@@ -158,9 +155,6 @@ generation.
 [[autodoc]] LogitsProcessorList
     - __call__
 
-[[autodoc]] LogitsWarper
-    - __call__
-
 [[autodoc]] MinLengthLogitsProcessor
     - __call__
 
@@ -386,14 +380,38 @@ A [`Constraint`] can be used to force the generation to include specific tokens
     - get_seq_length
     - reorder_cache
 
+[[autodoc]] OffloadedCache
+    - update
+    - prefetch_layer
+    - evict_previous_layer
+
 [[autodoc]] StaticCache
     - update
     - get_seq_length
     - reset
 
+[[autodoc]] HybridCache
+    - update
+    - get_seq_length
+    - reset
+
+[[autodoc]] SlidingWindowCache
+    - update
+    - reset
+
+[[autodoc]] EncoderDecoderCache
+    - get_seq_length
+    - to_legacy_cache
+    - from_legacy_cache
+    - reset
+    - reorder_cache
+
+[[autodoc]] MambaCache
+    - update_conv_state
+    - update_ssm_state
+    - reset
 
 ## Watermark Utils
 
 [[autodoc]] WatermarkDetector
     - __call__
-

docs/source/en/kv_cache.md

@@ -0,0 +1,346 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Best Practices for Generation with Cache
+
+Efficient caching is crucial for optimizing the performance of models in various generative tasks,
+including text generation, translation, summarization and other transformer-based applications.
+Effective caching helps reduce computation time and improve response rates, especially in real-time or resource-intensive applications.
+
+Transformers support various caching methods, leveraging "Cache" classes to abstract and manage the caching logic.
+This document outlines best practices for using these classes to maximize performance and efficiency.
+Check out all the available `Cache` classes in the [API documentation](./internal/generation_utils.md).
+
+## What is Cache and why we should care?
+
+Imagine you’re having a conversation with someone, and instead of remembering what was said previously, you have to start from scratch every time you respond. This would be slow and inefficient, right? In the world of Transformer models, a similar concept applies, and that's where Caching keys and values come into play. From now on, I'll refer to the concept as KV Cache.
+
+KV cache is needed to optimize the generation in autoregressive models, where the model predicts text token by token. This process can be slow since the model can generate only one token at a time, and each new prediction is dependent on the previous context. That means, to predict token number 1000 in the generation, you need information from the previous 999 tokens, which comes in the form of some matrix multiplications across the representations of those tokens. But to predict token number 1001, you also need the same information from the first 999 tokens, plus additional information from token number 1000. That is where key-value cache is used to optimize the sequential generation process by storing previous calculations to reuse in subsequent tokens, so they don't need to be computed again.
+
+More concretely, key-value cache acts as a memory bank for these generative models, where the model stores key-value pairs derived from self-attention layers for previously processed tokens. By storing this information, the model can avoid redundant computations and instead retrieve keys and values of previous tokens from the cache.
+
+<details>
+  <summary><em>For the Curious Minds Who Like to Dive Deep</em></summary>
+
+  ### Under the Hood: How Cache Object Works in Attention Mechanism
+
+  When utilizing a cache object in the input, the Attention module performs several critical steps to integrate past and present information seamlessly.
+
+  The Attention module concatenates the current key-values with the past key-values stored in the cache. This results in attention weights of shape `(new_tokens_length, past_kv_length + new_tokens_length)`. Essentially, the past and current key-values are combined to compute attention scores, ensuring that the model considers both previous context and new input. The concatenated key-values are used to compute the attention scores resulting in attention weights of shape `(new_tokens_length, past_kv_length + new_tokens_length)`.
+
+  Therefore, when iteratively calling `forward()` instead of the `generate()` method, it’s crucial to ensure that the attention mask shape matches the combined length of past and current key-values. The attention mask should have the shape `(batch_size, past_kv_length + new_tokens_length)`. This is usually handled internally when you call `generate()` method. If you want to implement your own generation loop with Cache classes, take this into consideration and prepare the attention mask to hold values to current and past tokens.
+
+  <Tip warning={true}>
+
+  One important concept you need to know when writing your own generation loop, is `cache_position`. In case you want to reuse an already filled Cache object by calling `forward()`, you have to pass in a valid `cache_position` which will indicate the positions of inputs in the sequence. Note that `cache_position` is not affected by padding, and always adds one more position for each token. For example, if key/value cache contains 10 tokens (no matter how many of it is a pad token), the cache position for the next token should be `torch.tensor([10])`.
+
+  </Tip>
+
+
+  See an example below for how to implement your own generation loop.
+    
+  ```python
+  >>> import torch
+  >>> from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache
+ 
+  >>> model_id = "meta-llama/Llama-2-7b-chat-hf"
+  >>> model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="cuda:0")
+  >>> tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+  >>> past_key_values = DynamicCache()
+  >>> messages = [{"role": "user", "content": "Hello, what's your name."}]
+  >>> inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to("cuda:0")
+
+  >>> generated_ids = inputs.input_ids
+  >>> cache_position = torch.arange(inputs.input_ids.shape[1], dtype=torch.int64, device="cuda:0")
+  >>> max_new_tokens = 10
+
+  >>> for _ in range(max_new_tokens):
+  ...     outputs = model(**inputs, cache_position=cache_position, past_key_values=past_key_values, use_cache=True)     
+  ...     # Greedily sample one next token
+  ...     next_token_ids = outputs.logits[:, -1:].argmax(-1)
+  ...     generated_ids = torch.cat([generated_ids, next_token_ids], dim=-1)   
+  ...
+  ...     # Prepare inputs for the next generation step by leaaving unprocessed tokens, in our case we have only one new token
+  ...     # and expanding attn mask for the new token, as explained above
+  ...     attention_mask = inputs["attention_mask"]
+  ...     attention_mask = torch.cat([attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1)
+  ...     inputs = {"input_ids": next_token_ids, "attention_mask": attention_mask}
+  ...     cache_position = cache_position[-1:] + 1 # add one more position for the next token
+
+  >>> print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0])
+  "[INST] Hello, what's your name. [/INST]  Hello! My name is LLaMA,"
+  ```
+
+</details>
+
+
+
+## Generate with Cache
+
+In 🤗 Transformers, we support various Cache types to optimize the performance across different models and tasks. By default, all models generate with caching,
+with the [`~DynamicCache`] class being the default cache for most models. It allows us to dynamically grow cache size, by saving more and more keys and values as we generate. If for some reason you don't want to use caches, you can pass `use_cache=False` into the `generate()` method.
+
+Refer to the table below to see the difference between cache types and choose the one that suits best for your use-case.
+
+| Cache Type          | Memory Efficient | Supports torch.compile() | Initialization Recommended | Latency  |  Long Context Generation |
+|---------------------|------------------|--------------------------|----------------------------|----------|--------------------------|
+| Dynamic Cache       |      No          |        No                |         No                 |   Mid    |     No                   |
+| Static Cache        |      No          |        Yes               |         Yes                |   High   |     No                   |
+| Quantized Cache     |      Yes         |        No                |         No                 |   Low    |     Yes                  |
+| Offloaded Cache     |      Yes         |        No                |         No                 |   Low    |     No                   |
+| Sliding Window Cache|      No          |        Yes               |         Yes                |   High   |     No                   |
+| Sink Cache          |      Yes         |        No                |         Yes                |   Mid    |     Yes                  |
+
+
+These cache classes can be set with a `cache_implementation` argument when generating. To learn about the available options for the cache_implementation flag, please refer to the [API Documentation](./main_classes/text_generation.md#transformers.GenerationConfig). Now, let's explore each cache type in detail and see how to use them. Note that the below examples are for decoder-only Tranformer-based models. We also support ["Model-Specific Cache"] classes for models such as Mamba or Jamba, keep reading for more details.
+
+### Quantized Cache
+
+The key and value cache can occupy a large portion of memory, becoming a [bottleneck for long-context generation](https://huggingface.co/blog/llama31#inference-memory-requirements), especially for Large Language Models.
+Quantizing the cache when using `generate()` can significantly reduce memory requirements at the cost of speed.
+
+KV Cache quantization in `transformers` is largely inspired by the paper ["KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache"](https://arxiv.org/abs/2402.02750) and currently supports [`~QuantoQuantizedCache`] and [`~HQQQuantizedCache`] classes. For more information on the inner workings see the paper.
+
+To enable quantization of the key-value cache, one needs to indicate `cache_implementation="quantized"` in the `generation_config`.
+Quantization related arguments should be passed to the `generation_config` either as a `dict` or an instance of a [`~QuantizedCacheConfig`] class.
+One has to indicate which quantization backend to use in the [`~QuantizedCacheConfig`], the default is `quanto`.
+
+<Tip warning={true}>
+
+Cache quantization can be detrimental in terms of latency if the context length is short and there is enough GPU VRAM available to run without cache quantization. It is recommended to seek balance between memory efficiency and latency.
+</Tip>
+
+
+```python
+>>> import torch
+>>> from transformers import AutoTokenizer, AutoModelForCausalLM
+
+>>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
+>>> model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16).to("cuda:0")
+>>> inputs = tokenizer("I like rock music because", return_tensors="pt").to(model.device)
+
+>>> out = model.generate(**inputs, do_sample=False, max_new_tokens=20, cache_implementation="quantized", cache_config={"nbits": 4, "backend": "quanto"})
+>>> print(tokenizer.batch_decode(out, skip_special_tokens=True)[0])
+I like rock music because it's loud and energetic. It's a great way to express myself and rel
+
+>>> out = model.generate(**inputs, do_sample=False, max_new_tokens=20)
+>>> print(tokenizer.batch_decode(out, skip_special_tokens=True)[0])
+I like rock music because it's loud and energetic. I like to listen to it when I'm feeling
+```
+
+## OffloadedCache
+
+Similarly to KV cache quantization, [`~OffloadedCache`] strategy aims to reduce GPU VRAM usage.
+It does so by moving the KV cache for most layers to the CPU.
+As the model's `forward()` method iterates over the layers, this strategy maintains the current layer cache on the GPU.
+At the same time it asynchronously prefetches the next layer cache as well as sending the previous layer cache back to the CPU.
+Unlike KV cache quantization, this strategy always produces the same result as the default KV cache implementation.
+Thus, it can serve as a drop-in replacement or a fallback for it.
+
+Depending on your model and the characteristics of your generation task (size of context, number of generated tokens, number of beams, etc.)
+you may notice a small degradation in generation throughput compared to the default KV cache implementation.
+
+To enable KV cache offloading, pass `cache_implementation="offloaded"` in the `generation_config` or directky to the `generate()` call.
+
+```python
+>>> import torch
+>>> from transformers import AutoTokenizer, AutoModelForCausalLM
+>>> ckpt = "microsoft/Phi-3-mini-4k-instruct"
+
+>>> tokenizer = AutoTokenizer.from_pretrained(ckpt)
+>>> model = AutoModelForCausalLM.from_pretrained(ckpt, torch_dtype=torch.float16).to("cuda:0")
+>>> inputs = tokenizer("Fun fact: The shortest", return_tensors="pt").to(model.device)
+
+>>> out = model.generate(**inputs, do_sample=False, max_new_tokens=23, cache_implementation="offloaded")
+>>> print(tokenizer.batch_decode(out, skip_special_tokens=True)[0])
+Fun fact: The shortest war in history was between Britain and Zanzibar on August 27, 1896.
+
+>>> out = model.generate(**inputs, do_sample=False, max_new_tokens=23)
+>>> print(tokenizer.batch_decode(out, skip_special_tokens=True)[0])
+Fun fact: The shortest war in history was between Britain and Zanzibar on August 27, 1896.
+```
+
+<Tip warning={true}>
+
+Cache offloading requires a GPU and can be slower than dynamic KV cache. Use it if you are getting CUDA out of memory errors.
+
+</Tip>
+
+The example below shows how KV cache offloading can be used as a fallback strategy.
+```python
+>>> import torch
+>>> from transformers import AutoTokenizer, AutoModelForCausalLM
+>>> def resilient_generate(model, *args, **kwargs):
+...     oom = False
+...     try:
+...         return model.generate(*args, **kwargs)
+...     except torch.cuda.OutOfMemoryError as e:
+...         print(e)
+...         print("retrying with cache_implementation='offloaded'")
+...         oom = True
+...     if oom:
+...         torch.cuda.empty_cache()
+...         kwargs["cache_implementation"] = "offloaded"
+...         return model.generate(*args, **kwargs)
+...
+...
+>>> ckpt = "microsoft/Phi-3-mini-4k-instruct"
+>>> tokenizer = AutoTokenizer.from_pretrained(ckpt)
+>>> model = AutoModelForCausalLM.from_pretrained(ckpt, torch_dtype=torch.float16).to("cuda:0")
+>>> prompt = ["okay "*1000 + "Fun fact: The most"]
+>>> inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+>>> beams = { "num_beams": 40, "num_beam_groups": 40, "num_return_sequences": 40, "diversity_penalty": 1.0, "max_new_tokens": 23, "early_stopping": True, }
+>>> out = resilient_generate(model, **inputs, **beams)
+>>> responses = tokenizer.batch_decode(out[:,-28:], skip_special_tokens=True)
+```
+
+On a GPU with 50 GB of RAM, running this code will print
+```
+CUDA out of memory. Tried to allocate 4.83 GiB. GPU
+retrying with cache_implementation='offloaded'
+```
+before successfully generating 40 beams.
+
+
+
+### Static Cache
+
+Since the "DynamicCache" dynamically grows with each generation step, it prevents you from taking advantage of JIT optimizations. The [`~StaticCache`] pre-allocates 
+a specific maximum size for the keys and values, allowing you to generate up to the maximum length without having to modify cache size. Check the below usage example.
+
+For more examples with Static Cache and JIT compilation, take a look at [StaticCache & torchcompile](./llm_optims.md#static-kv-cache-and-torchcompile)
+
+```python
+>>> import torch
+>>> from transformers import AutoTokenizer, AutoModelForCausalLM
+
+>>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
+>>> model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16, device_map="auto")
+>>> inputs = tokenizer("Hello, my name is", return_tensors="pt").to(model.device)
+
+>>> # simply pass the cache implementation="static"
+>>> out = model.generate(**inputs, do_sample=False, max_new_tokens=20, cache_implementation="static")
+>>> tokenizer.batch_decode(out, skip_special_tokens=True)[0]
+"Hello, my name is [Your Name], and I am a [Your Profession] with [Number of Years] of"
+```
+
+### Sliding Window Cache
+
+As the name suggests, this cache type implements a sliding window over previous keys and values, retaining only the last `sliding_window` tokens. It should be used with models like Mistral that support sliding window attention. Additionally, similar to Static Cache, this one is JIT-friendly and can be used with the same compile tecniques as Static Cache.
+
+Note that you can use this cache only for models that support sliding window, e.g. Mistral models. 
+
+
+```python
+>>> import torch
+>>> from transformers import AutoTokenizer, AutoModelForCausalLM, SinkCache
+
+>>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
+>>> model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1", torch_dtype=torch.float16).to("cuda:0")
+>>> inputs = tokenizer("Yesterday I was on a rock concert and.", return_tensors="pt").to(model.device)
+
+>>> # can be used by passing in cache implementation
+>>> out = model.generate(**inputs, do_sample=False, max_new_tokens=30, cache_implementation="sliding_window")
+>>> tokenizer.batch_decode(out, skip_special_tokens=True)[0]
+"Yesterday I was on a rock concert and. I was so excited to see my favorite band. I was so excited that I was jumping up and down and screaming. I was so excited that I"
+```
+
+### Sink Cache
+
+Sink Cache was introduced in ["Efficient Streaming Language Models with Attention Sinks"](https://arxiv.org/abs/2309.17453). It allows you to generate long sequences of text ("infinite length" according to the paper) without any fine-tuning. That is achieved by smart handling of previous keys and values, specifically it retains a few initial tokens from the sequence, called "sink tokens". This is based on the observation that these initial tokens attract a significant portion of attention scores during the generation process. Tokens that come after "sink tokens" are discarded on a sliding windowed basis, keeping only the latest `window_size` tokens. By keeping these initial tokens as "attention sinks," the model maintains stable performance even when dealing with very long texts, thus discarding most of the previous knowledge.
+
+Unlike other cache classes, this one can't be used directly by indicating a `cache_implementation`. You have to initialize the Cache before calling on `generate()` as follows.
+
+```python
+>>> import torch
+>>> from transformers import AutoTokenizer, AutoModelForCausalLM, SinkCache
+
+>>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
+>>> model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", torch_dtype=torch.float16).to("cuda:0")
+>>> inputs = tokenizer("This is a long story about unicorns, fairies and magic.", return_tensors="pt").to(model.device)
+
+>>> # get our cache, specify number of sink tokens and window size
+>>> # Note that window size already includes sink tokens, so has to be larger
+>>> past_key_values = SinkCache(window_length=256, num_sink_tokens=4)
+>>> out = model.generate(**inputs, do_sample=False, max_new_tokens=30, past_key_values=past_key_values)
+>>> tokenizer.batch_decode(out, skip_special_tokens=True)[0]
+"This is a long story about unicorns, fairies and magic. It is a fantasy world where unicorns and fairies live together in harmony. The story follows a young girl named Lily"
+```
+
+### Encoder-Decoder Cache
+
+The [`~EncoderDecoderCache`] is a wrapper designed to handle the caching needs of encoder-decoder models. This cache type is specifically built to manage both self-attention and cross-attention caches, ensuring storage and retrieval of past key/values required for these complex models. Cool thing about Encoder-Decoder Cache is that you can set different cache types for the encoder and for the decoder, depending on your use case. Currently this cache is only supported in [Whisper](./model_doc/whisper.md) models but we will be adding more models soon. 
+
+In terms of usage, there is nothing special to be done and calling `generate()` or `forward()` will handle everything for you.
+
+
+### Model-specific Cache Classes
+
+Some models require storing previous keys, values, or states in a specific way, and the above cache classes cannot be used. For such cases, we have several specialized cache classes that are designed for specific models. These models only accept their own dedicated cache classes and do not support using any other cache types. Some examples include [`~HybridCache`] for [Gemma2](./model_doc/gemma2.md) series models or [`~MambaCache`] for [Mamba](./model_doc/mamba.md) architecture models.
+
+
+## Iterative Generation with Cache
+
+We have seen how to use each of the cache types when generating. What if you want to use cache in iterative generation setting, for example in applications like chatbots, where interactions involve multiple turns and continuous back-and-forth exchanges. Iterative generation with cache allows these systems to handle ongoing conversations effectively without reprocessing the entire context at each step. But there are some tips that you should know before you start implementing:
+
+The general format when doing iterative generation is as below. First you have to initialize an empty cache of the type you want, and you can start feeding in new prompts iteratively. Keeping track of dialogues history and formatting can be done with chat templates, read more on that in [chat_templating](./chat_templating.md)
+
+In case you are using Sink Cache, you have to crop your inputs to that maximum length because Sink Cache can generate text longer than its maximum window size, but it expects the first input to not exceed the maximum cache length.  
+
+
+```python
+>>> import torch
+>>> from transformers import AutoTokenizer,AutoModelForCausalLM
+>>> from transformers.cache_utils import (
+>>>     DynamicCache,
+>>>     SinkCache,
+>>>     StaticCache,
+>>>     SlidingWindowCache,
+>>>     QuantoQuantizedCache,
+>>>     QuantizedCacheConfig,
+>>> )
+
+>>> model_id = "meta-llama/Llama-2-7b-chat-hf"
+>>> model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map='auto')
+>>> tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+>>> user_prompts = ["Hello, what's your name?", "Btw, yesterday I was on a rock concert."]
+
+>>> past_key_values = DynamicCache()
+>>> max_cache_length = past_key_values.get_max_length()
+
+>>> messages = []
+>>> for prompt in user_prompts:
+...     messages.append({"role": "user", "content": prompt})
+...     inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
+...     if isinstance(past_key_values, SinkCache):
+...         inputs = {k: v[:, -max_cache_length:] for k, v in inputs.items()}
+... 
+...     input_length = inputs["input_ids"].shape[1]
+...     
+...     outputs = model.generate(**inputs, do_sample=False, max_new_tokens=256, past_key_values=past_key_values)
+...     completion = tokenizer.decode(outputs[0, input_length: ], skip_special_tokens=True)
+...     messages.append({"role": "assistant", "content": completion})
+
+print(messages)
+[{'role': 'user', 'content': "Hello, what's your name?"}, {'role': 'assistant', 'content': " Hello! My name is LLaMA, I'm a large language model trained by a team of researcher at Meta AI. 😊"}, {'role': 'user', 'content': 'Btw, yesterday I was on a rock concert.'}, {'role': 'assistant', 'content': ' Oh, cool! That sounds like a lot of fun! 🎉 Did you enjoy the concert? What was the band like? 🤔'}]
+```
+
+
+## Re-use Cache to continue generation
+
+Sometimes you would want to fist fill-in cache object with key/values for certain prefix prompt and re-use it several times to generate different sequences from it. We are working hard on adding this feature to 🤗 Transformers and will update this section soon. 

docs/source/en/llm_optims.md

@@ -18,59 +18,109 @@ Basic inference is slow because LLMs have to be called repeatedly to generate th
 This guide will show you how to use the optimization techniques available in Transformers to accelerate LLM inference.
 
 > [!TIP]
-> Hugging Face also provides [Text Generation Inference (TGI)](https://hf.co/docs/text-generation-inference), a library dedicated to deploying and serving highly optimized LLMs for inference. It includes more optimization features not included in Transformers, such as continuous batching for increasing throughput and tensor parallelism for multi-GPU inference.
+> Hugging Face also provides [Text Generation Inference (TGI)](https://hf.co/docs/text-generation-inference), a library dedicated to deploying and serving highly optimized LLMs for inference. It includes deployment-oriented optimization features not included in Transformers, such as continuous batching for increasing throughput and tensor parallelism for multi-GPU inference.
 
-## Static kv-cache and torch.compile
+## Static kv-cache and `torch.compile`
 
 During decoding, a LLM computes the key-value (kv) values for each input token and since it is autoregressive, it computes the same kv values each time because the generated output becomes part of the input now. This is not very efficient because you're recomputing the same kv values each time.
 
-To optimize this, you can use a kv-cache to store the past keys and values instead of recomputing them each time. However, since the kv-cache grows with each generation step and is dynamic, it prevents you from taking advantage of [torch.compile](./perf_torch_compile), a powerful optimization tool that fuses PyTorch code into fast and optimized kernels.
+To optimize this, you can use a kv-cache to store the past keys and values instead of recomputing them each time. However, since the kv-cache grows with each generation step and is dynamic, it prevents you from taking advantage of [`torch.compile`](./perf_torch_compile), a powerful optimization tool that fuses PyTorch code into fast and optimized kernels.
 
-The *static kv-cache* solves this issue by pre-allocating the kv-cache size to a maximum value which allows you to combine it with torch.compile for up to a 4x speed up.
+The *static kv-cache* solves this issue by pre-allocating the kv-cache size to a maximum value which allows you to combine it with `torch.compile` for up to a 4x speed up. Your speed up may vary depending on the model size (larger models have a smaller speed up) and hardware.
 
 > [!WARNING]
-> Currently, only [Llama](./model_doc/llama2) and a few other models support static kv-cache and torch.compile. Check [this issue](https://github.com/huggingface/transformers/issues/28981) for a live model compatibility list.
+> Currently, only [Llama](./model_doc/llama2) and a few other models support static kv-cache and `torch.compile`. Check [this issue](https://github.com/huggingface/transformers/issues/28981) for a live model compatibility list.
 
-For this example, let's load the [Gemma](https://hf.co/google/gemma-2b) model.
+There are three flavors of static kv-cache usage, depending on the complexity of your task:
+1. Basic usage: simply set a flag in `generation_config` (recommended);
+2. Advanced usage: handle a cache object for multi-turn generation or a custom generation loop;
+3. Advanced usage: compile the entire `generate` function into a single graph, if having a single graph is relevant for you.
+
+Select the correct tab below for further instructions on each of these flavors.
+
+> [!TIP]
+> Regardless of the strategy used with `torch.compile`, you can avoid shape-related recompilations if you left-pad your LLM inputs to a limited set of values. The [`pad_to_multiple_of` tokenizer flag](https://huggingface.co/docs/transformers/main_classes/tokenizer#transformers.PreTrainedTokenizer.__call__.pad_to_multiple_of) is your friend!
+
+<hfoptions id="static-kv">
+<hfoption id="basic usage: generation_config">
+
+For this example, let's use the [Gemma](https://hf.co/google/gemma-2b) model. All we need to do is to:
+1. Access the model's `generation_config` attribute and set the `cache_implementation` to "static";
+2. Call `torch.compile` on the model to compile the forward pass with the static kv-cache.
+
+And that's it!
 
 ```py
 from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+import os
+os.environ["TOKENIZERS_PARALLELISM"] = "false"  # To prevent long warnings :)
 
 tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
-model = AutoModelForCausalLM.from_pretrained(
-    "google/gemma-2b", device_map="auto"
-)
-```
+model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", device_map="auto")
 
-There are two ways you can configure the model to use a static kv-cache. For a 7B model on an A100, both methods get a 4x speed up in the forward pass. Your speed up may vary depending on the model size (larger models have a smaller speed up) and hardware. If you're using the [`~GenerationMixin.generate`] method, the speed up is ~3x. The forward pass (which still gets 4x speed up) is only a part of the whole [`~GenerationMixin.generate`] code.
-
-<hfoptions id="static-kv">
-<hfoption id="generation_config">
-
-Access the model's `generation_config` attribute and set the `cache_implementation` to "static".
-
-```py
 model.generation_config.cache_implementation = "static"
-```
 
-Call torch.compile on the model to compile the forward pass with the static kv-cache.
-
-```py
-compiled_model = torch.compile(model, mode="reduce-overhead", fullgraph=True)
+model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
 input_text = "The theory of special relativity states "
 input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
 
-outputs = compiled_model.generate(**input_ids)
-tokenizer.batch_decode(outputs, skip_special_tokens=True)
+outputs = model.generate(**input_ids)
+print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
 ['The theory of special relativity states 1. The speed of light is constant in all inertial reference']
 ```
 
-Under the hood, `generate` will attempt to reuse the same cache object, removing the need for re-compilation at each call. However, if the batch size or the maximum output length increase between calls, the cache will have to be reinitialized, triggering a new compilation.
+Under the hood, `generate` will attempt to reuse the same cache object, removing the need for re-compilation at each call. Avoiding re-compilation is critical to get the most out of `torch.compile`, and you should be aware of the following:
+1. If the batch size changes or the maximum output length increases between calls, the cache will have to be reinitialized, triggering a new compilation;
+2. The first couple of calls of the compiled function are slower, as the function is being compiled.
+
+> [!WARNING]
+> For a more advanced usage of the static cache, such as multi-turn conversations, we recommend instantiating and manipulating the cache object outside [`~GenerationMixin.generate`]. See the advanced usage tab.
 
 </hfoption>
-<hfoption id="Static Cache">
+<hfoption id="advanced usage: control Static Cache">
 
-A [`StaticCache`] object can be passed to the model's forward pass under the `past_key_values` argument, enabling the use of this object as a static kv-cache. Using this strategy, you can write your own function to decode the next token given the current token and position and cache position of previously generated tokens. You can also pass the [`StaticCache`] object to [`~GenerationMixin.generate`] and use it across calls, like you would do with a dynamic cache.
+A [`StaticCache`] object can be passed to the model's [`~GenerationMixin.generate`] under the `past_key_values` argument. The object will retain the cache contents, so you can pass it to a new [`~GenerationMixin.generate`] call to continue generation, like you would do with a dynamic cache.
+
+```py
+from transformers import AutoTokenizer, AutoModelForCausalLM, StaticCache
+import torch
+import os
+os.environ["TOKENIZERS_PARALLELISM"] = "false"  # To prevent long warnings :)
+
+tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
+model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", device_map="auto")
+
+model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
+input_text = "The theory of special relativity states "
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+prompt_length = input_ids.input_ids.shape[1]
+model.generation_config.max_new_tokens = 16
+
+past_key_values = StaticCache(
+    config=model.config,
+    batch_size=1,
+    # If you plan to reuse the cache, make sure the cache length is large enough for all cases
+    max_cache_len=prompt_length+(model.generation_config.max_new_tokens*2),
+    device=model.device,
+    dtype=model.dtype
+)
+outputs = model.generate(**input_ids, past_key_values=past_key_values)
+print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
+['The theory of special relativity states 1. The speed of light is constant in all inertial reference frames. 2']
+
+# pass in the generated text and the same cache object to continue generation from where it left off. Optionally, in a
+# multi-turn conversation, append the new user input to the generated text.
+new_input_ids = outputs
+outputs = model.generate(new_input_ids, past_key_values=past_key_values)
+print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
+['The theory of special relativity states 1. The speed of light is constant in all inertial reference frames. 2. The speed of light is constant in all inertial reference frames. 3.']
+```
+
+> [!TIP]
+> If you want to reuse the same [`StaticCache`] object on a new prompt, be sure to reset its contents with the `.reset()` method between calls
+
+If you want to go further down a level, the [`StaticCache`] object can also be passed to the model's forward pass under the same `past_key_values` argument. Using this strategy, you can write your own function to decode the next token given the current token and position and cache position of previously generated tokens.
 
 ```py
 from transformers import LlamaTokenizer, LlamaForCausalLM, StaticCache, logging
@@ -102,19 +152,16 @@ def decode_one_tokens(model, cur_token, input_pos, cache_position, past_key_valu
     return new_token
 ```
 
-There are a few important things you must do to enable static kv-cache and torch.compile with the `StaticCache` method:
-
+There are a few important things you must do to enable static kv-cache and `torch.compile` with the `StaticCache` method:
 1. Initialize the [`StaticCache`] instance before using the model for inference. There you can configure parameters like the maximum batch size and sequence length.
-
-2. Call torch.compile on the model to compile the forward pass with the static kv-cache.
-
+2. Call `torch.compile` on the model to compile the forward pass with the static kv-cache.
 3. Set `enable_math=True` in the [torch.backends.cuda.sdp_kernel](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention.html) context manager to enable the native PyTorch C++ implementation of scaled dot product attention to speed up inference even more.
 
 ```py
 batch_size, seq_length = inputs["input_ids"].shape
 with torch.no_grad():
     past_key_values = StaticCache(
-        config=model.config, max_batch_size=2, max_cache_len=4096, device=torch_device, dtype=model.dtype
+        config=model.config, batch_size=2, max_cache_len=4096, device=torch_device, dtype=model.dtype
     )
     cache_position = torch.arange(seq_length, device=torch_device)
     generated_ids = torch.zeros(
@@ -142,8 +189,34 @@ text
  'My favorite all time favorite condiment is ketchup. I love it on everything. I love it on my eggs, my fries, my chicken, my burgers, my hot dogs, my sandwiches, my salads, my p']
 ```
 
-> [!TIP]
-> If you want to reuse the [`StaticCache`] object on a new prompt, be sure to reset its contents with the `.reset()` method
+</hfoption>
+<hfoption id="advanced usage: end-to-end generate compilation">
+
+Compiling the entire `generate` function, in terms of code, is even simpler than in the basic usage: call `torch.compile` on `generate` to compile the entire function. No need to specify the use of the static cache: although it is compatible, dynamic cache (default) was faster in our benchmarks.
+
+```py
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+import os
+os.environ["TOKENIZERS_PARALLELISM"] = "false"  # To prevent long warnings :)
+
+tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
+model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", device_map="auto")
+
+model.generate = torch.compile(model.generate, mode="reduce-overhead", fullgraph=True)
+input_text = "The theory of special relativity states "
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+outputs = model.generate(**input_ids)
+print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
+['The theory of special relativity states 1. The speed of light is constant in all inertial reference']
+```
+
+As a result, we compile not only the model forward pass, but also all input preparation, logit processor operations, and so on. The result should be a slightly `generate` call, compared to the basic usage example, and the compiled graph may be better suited to more exotic hardware devices or use cases. However, there are severe drawbacks in using this approach:
+1. Compilation is much slower;
+2. All parameterization of `generate` must be done through `generation_config`;
+3. Many warnings and exceptions are suppressed -- we suggest testing with its uncompiled form first;
+4. Although we are working on it, it is heavily feature restricted (for instance, at the time of writing, generation does not stop if an EOS token is selected).
 
 </hfoption>
 </hfoptions>

docs/source/en/llm_tutorial.md

@@ -267,5 +267,6 @@ While the autoregressive generation process is relatively straightforward, makin
 
 1. [`optimum`](https://github.com/huggingface/optimum), an extension of 🤗 Transformers that optimizes for specific hardware devices.
 2. [`outlines`](https://github.com/outlines-dev/outlines), a library where you can constrain text generation (e.g. to generate JSON files);
-3. [`text-generation-inference`](https://github.com/huggingface/text-generation-inference), a production-ready server for LLMs;
-4. [`text-generation-webui`](https://github.com/oobabooga/text-generation-webui), a UI for text generation;
+3. [`SynCode`](https://github.com/uiuc-focal-lab/syncode), a library for context-free grammar guided generation. (e.g. JSON, SQL, Python)
+4. [`text-generation-inference`](https://github.com/huggingface/text-generation-inference), a production-ready server for LLMs;
+5. [`text-generation-webui`](https://github.com/oobabooga/text-generation-webui), a UI for text generation;

docs/source/en/llm_tutorial_optimization.md

@@ -147,7 +147,7 @@ Let's call it now for the next experiment.
 ```python
 flush()
 ```
-In the recent version of the accelerate library, you can also use an utility method called `release_memory()`
+In the recent version of the accelerate library, you can also use a utility method called `release_memory()`
 
 ```python
 from accelerate.utils import release_memory
@@ -683,7 +683,7 @@ Assistant: Germany has ca. 81 million inhabitants
 
 In this chat, the LLM runs auto-regressive decoding twice:
   1. The first time, the key-value cache is empty and the input prompt is `"User: How many people live in France?"` and the model auto-regressively generates the text `"Roughly 75 million people live in France"` while increasing the key-value cache at every decoding step.
-  2. The second time the input prompt is `"User: How many people live in France? \n Assistant: Roughly 75 million people live in France \n User: And how many in Germany?"`. Thanks to the cache, all key-value vectors for the first two sentences are already computed. Therefore the input prompt only consists of `"User: And how many in Germany?"`. While processing the shortened input prompt, it's computed key-value vectors are concatenated to the key-value cache of the first decoding. The second Assistant's answer `"Germany has ca. 81 million inhabitants"` is then auto-regressively generated with the key-value cache consisting of encoded key-value vectors of `"User: How many people live in France? \n Assistant: Roughly 75 million people live in France \n User: And how many are in Germany?"`.
+  2. The second time the input prompt is `"User: How many people live in France? \n Assistant: Roughly 75 million people live in France \n User: And how many in Germany?"`. Thanks to the cache, all key-value vectors for the first two sentences are already computed. Therefore the input prompt only consists of `"User: And how many in Germany?"`. While processing the shortened input prompt, its computed key-value vectors are concatenated to the key-value cache of the first decoding. The second Assistant's answer `"Germany has ca. 81 million inhabitants"` is then auto-regressively generated with the key-value cache consisting of encoded key-value vectors of `"User: How many people live in France? \n Assistant: Roughly 75 million people live in France \n User: And how many are in Germany?"`.
 
 Two things should be noted here:
   1. Keeping all the context is crucial for LLMs deployed in chat so that the LLM understands all the previous context of the conversation. E.g. for the example above the LLM needs to understand that the user refers to the population when asking `"And how many are in Germany"`.

docs/source/en/main_classes/agent.md

@@ -72,6 +72,10 @@ We provide two types of agents, based on the main [`Agent`] class:
 
 [[autodoc]] launch_gradio_demo
 
+### stream_to_gradio
+
+[[autodoc]] stream_to_gradio
+
 ### ToolCollection
 
 [[autodoc]] ToolCollection

docs/source/en/main_classes/backbones.md

@@ -25,11 +25,11 @@ A backbone is a model used for feature extraction for higher level computer visi
 
 Backbones are supported for the following models:
 
-* [BEiT](..model_doc/beit)
+* [BEiT](../model_doc/beit)
 * [BiT](../model_doc/bit)
-* [ConvNet](../model_doc/convnext)
+* [ConvNext](../model_doc/convnext)
 * [ConvNextV2](../model_doc/convnextv2)
-* [DiNAT](..model_doc/dinat)
+* [DiNAT](../model_doc/dinat)
 * [DINOV2](../model_doc/dinov2)
 * [FocalNet](../model_doc/focalnet)
 * [MaskFormer](../model_doc/maskformer)

docs/source/en/main_classes/callback.md

@@ -34,7 +34,7 @@ By default, `TrainingArguments.report_to` is set to `"all"`, so a [`Trainer`] wi
 - [`~integrations.TensorBoardCallback`] if tensorboard is accessible (either through PyTorch >= 1.4
   or tensorboardX).
 - [`~integrations.WandbCallback`] if [wandb](https://www.wandb.com/) is installed.
-- [`~integrations.CometCallback`] if [comet_ml](https://www.comet.ml/site/) is installed.
+- [`~integrations.CometCallback`] if [comet_ml](https://www.comet.com/site/) is installed.
 - [`~integrations.MLflowCallback`] if [mlflow](https://www.mlflow.org/) is installed.
 - [`~integrations.NeptuneCallback`] if [neptune](https://neptune.ai/) is installed.
 - [`~integrations.AzureMLCallback`] if [azureml-sdk](https://pypi.org/project/azureml-sdk/) is

docs/source/en/main_classes/data_collator.md

@@ -66,3 +66,8 @@ Examples of use can be found in the [example scripts](../examples) or [example n
     - numpy_mask_tokens
     - tf_mask_tokens
     - torch_mask_tokens
+
+## DataCollatorWithFlattening
+
+[[autodoc]] data.data_collator.DataCollatorWithFlattening
+

docs/source/en/main_classes/model.md

@@ -40,6 +40,10 @@ for text generation, [`~generation.GenerationMixin`] (for the PyTorch models),
     - push_to_hub
     - all
 
+Custom models should also include a `_supports_assign_param_buffer`, which determines if superfast init can apply
+on the particular model. Signs that your model needs this are if `test_save_and_load_from_pretrained` fails. If so,
+set this to `False`.
+
 ## ModuleUtilsMixin
 
 [[autodoc]] modeling_utils.ModuleUtilsMixin

docs/source/en/main_classes/pipelines.md

@@ -270,6 +270,11 @@ This is a simplified view, since the pipeline can handle automatically the batch
 about how many forward passes you inputs are actually going to trigger, you can optimize the `batch_size`
 independently of the inputs. The caveats from the previous section still apply.
 
+## Pipeline FP16 inference
+Models can be run in FP16 which can be significantly faster on GPU while saving memory. Most models will not suffer noticeable performance loss from this. The larger the model, the less likely that it will.
+
+To enable FP16 inference, you can simply pass `torch_dtype=torch.float16` or `torch_dtype='float16'` to the pipeline constructor. Note that this only works for models with a PyTorch backend. Your inputs will be converted to FP16 internally.
+
 ## Pipeline custom code
 
 If you want to override a specific pipeline.

docs/source/en/main_classes/quantization.md

@@ -56,3 +56,12 @@ Learn how to quantize models in the [Quantization](../quantization) guide.
 ## HqqConfig
 
 [[autodoc]] HqqConfig
+
+## FbgemmFp8Config
+
+[[autodoc]] FbgemmFp8Config
+
+## TorchAoConfig
+
+[[autodoc]] TorchAoConfig
+

docs/source/en/model_doc/blip-2.md

@@ -87,4 +87,17 @@ If you're interested in submitting a resource to be included here, please feel f
 
 [[autodoc]] Blip2ForConditionalGeneration
     - forward
-    - generate
+    - generate
+
+## Blip2ForImageTextRetrieval
+
+[[autodoc]] Blip2ForImageTextRetrieval
+    - forward
+
+## Blip2TextModelWithProjection
+
+[[autodoc]] Blip2TextModelWithProjection
+
+## Blip2VisionModelWithProjection
+
+[[autodoc]] Blip2VisionModelWithProjection

docs/source/en/model_doc/chameleon.md

@@ -0,0 +1,192 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Chameleon
+
+## Overview
+
+The Chameleon model was proposed in [Chameleon: Mixed-Modal Early-Fusion Foundation Models
+](https://arxiv.org/abs/2405.09818v1) by META AI Chameleon Team. Chameleon is a Vision-Language Model that use vector quantization to tokenize images which enables the model to generate multimodal output. The model takes images and texts as input, including an interleaved format, and generates textual response. Image generation module is not released yet. 
+
+
+The abstract from the paper is the following:
+
+*We present Chameleon, a family of early-fusion token-based mixed-modal models capable of understanding and generating images and text in any arbitrary sequence. We outline a stable training
+approach from inception, an alignment recipe, and an architectural parameterization tailored for the
+early-fusion, token-based, mixed-modal setting. The models are evaluated on a comprehensive range
+of tasks, including visual question answering, image captioning, text generation, image generation, and
+long-form mixed modal generation. Chameleon demonstrates broad and general capabilities, including
+state-of-the-art performance in image captioning tasks, outperforms Llama-2 in text-only tasks while
+being competitive with models such as Mixtral 8x7B and Gemini-Pro, and performs non-trivial image
+generation, all in a single model. It also matches or exceeds the performance of much larger models,
+including Gemini Pro and GPT-4V, according to human judgments on a new long-form mixed-modal
+generation evaluation, where either the prompt or outputs contain mixed sequences of both images and
+text. Chameleon marks a significant step forward in unified modeling of full multimodal documents*
+
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/chameleon_arch.png"
+alt="drawing" width="600"/>
+
+<small> Chameleon incorporates a vector quantizer module to transform images into discrete tokens. That also enables image generation using an auto-regressive transformer. Taken from the <a href="https://arxiv.org/abs/2405.09818v1">original paper.</a> </small>
+
+This model was contributed by [joaogante](https://huggingface.co/joaogante) and [RaushanTurganbay](https://huggingface.co/RaushanTurganbay).
+The original code can be found [here](https://github.com/facebookresearch/chameleon).
+
+
+## Usage tips
+
+- We advise users to use `padding_side="left"` when computing batched generation as it leads to more accurate results. Simply make sure to set `processor.tokenizer.padding_side = "left"` before generating.
+
+- Note that Chameleon was tuned for safety alignment. If the model is refusing to answer, consider asking a more concrete question, instead of an open question.
+
+- Chameleon generates in chat format which means that the generated text will always be the "assistant's turn". You can enable a text completion generation by passing `return_for_text_completion=True` when calling the processor.
+
+> [!NOTE]
+> Chameleon implementation in Transformers uses a special image token to indicate where to merge image embeddings. For special image token we didn't add a new one but used one of the reserved tokens: `<reserved08707>`. You have to add `<image>` to your prompt in the place where the image should be embedded for correct generation.
+
+## Usage example
+
+### Single image inference
+
+Chameleon is a gated model so make sure to have access and login to Hugging Face Hub using a token. 
+Here's how to load the model and perform inference in half-precision (`torch.bfloat16`):
+
+```python
+from transformers import ChameleonProcessor, ChameleonForConditionalGeneration
+import torch
+from PIL import Image
+import requests
+
+processor = ChameleonProcessor.from_pretrained("facebook/chameleon-7b")
+model = ChameleonForConditionalGeneration.from_pretrained("facebook/chameleon-7b", torch_dtype=torch.bfloat16, device_map="cuda")
+
+# prepare image and text prompt
+url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
+image = Image.open(requests.get(url, stream=True).raw)
+prompt = "What do you see in this image?<image>"
+
+inputs = processor(prompt, image, return_tensors="pt").to(model.device)
+
+# autoregressively complete prompt
+output = model.generate(**inputs, max_new_tokens=50)
+print(processor.decode(output[0], skip_special_tokens=True))
+```
+
+### Multi image inference
+
+Chameleon can perform inference with multiple images as input, where images either belong to the same prompt or different prompts (in batched inference). Here is how you can do it:
+
+```python
+from transformers import ChameleonProcessor, ChameleonForConditionalGeneration
+import torch
+from PIL import Image
+import requests
+
+processor = ChameleonProcessor.from_pretrained("facebook/chameleon-7b")
+
+model = ChameleonForConditionalGeneration.from_pretrained("facebook/chameleon-7b", torch_dtype=torch.bfloat16, device_map="cuda")
+
+# Get three different images
+url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+image_stop = Image.open(requests.get(url, stream=True).raw)
+
+url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+image_cats = Image.open(requests.get(url, stream=True).raw)
+
+url = "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/snowman.jpg"
+image_snowman = Image.open(requests.get(url, stream=True).raw)
+
+# Prepare a batched prompt, where the first one is a multi-image prompt and the second is not
+prompts = [
+    "What do these images have in common?<image><image>",
+    "<image>What is shown in this image?"
+]
+
+# We can simply feed images in the order they have to be used in the text prompt
+# Each "<image>" token uses one image leaving the next for the subsequent "<image>" tokens
+inputs = processor(text=prompts, images=[image_stop, image_cats, image_snowman], padding=True, return_tensors="pt").to(device="cuda", dtype=torch.bfloat16)
+
+# Generate
+generate_ids = model.generate(**inputs, max_new_tokens=50)
+processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+```
+
+## Model optimization
+
+### Quantization using Bitsandbytes
+
+The model can be loaded in 8 or 4 bits, greatly reducing the memory requirements while maintaining the performance of the original model. First make sure to install bitsandbytes, `pip install bitsandbytes` and make sure to have access to a CUDA compatible GPU device. Simply change the snippet above with:
+
+```python
+from transformers import ChameleonForConditionalGeneration, BitsAndBytesConfig
+
+# specify how to quantize the model
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype=torch.bfloat16,
+)
+
+model = ChameleonForConditionalGeneration.from_pretrained("facebook/chameleon-7b", quantization_config=quantization_config, device_map="cuda")
+```
+
+### Use Flash-Attention 2 and SDPA to further speed-up generation
+
+The models supports both, Flash-Attention 2 and PyTorch's [`torch.nn.functional.scaled_dot_product_attention`](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention.html) which can be enables for optimization. SDPA is the default options when you load the model, If you want to switch for Flash Attention 2, first make sure to install flash-attn. Refer to the [original repository](https://github.com/Dao-AILab/flash-attention) regarding that package installation. Simply change the snippet above with:
+
+```python
+from transformers import ChameleonForConditionalGeneration
+
+model_id = "facebook/chameleon-7b"
+model = ChameleonForConditionalGeneration.from_pretrained(
+    model_id, 
+    torch_dtype=torch.bfloat16, 
+    low_cpu_mem_usage=True,
+    attn_implementation="flash_attention_2"
+).to(0)
+```
+
+## ChameleonConfig
+
+[[autodoc]] ChameleonConfig
+
+## ChameleonVQVAEConfig
+
+[[autodoc]] ChameleonVQVAEConfig
+
+## ChameleonProcessor
+
+[[autodoc]] ChameleonProcessor
+
+## ChameleonImageProcessor
+
+[[autodoc]] ChameleonImageProcessor
+    - preprocess
+
+## ChameleonVQVAE
+
+[[autodoc]] ChameleonVQVAE
+    - forward
+
+## ChameleonModel
+
+[[autodoc]] ChameleonModel
+    - forward
+
+## ChameleonForConditionalGeneration
+
+[[autodoc]] ChameleonForConditionalGeneration
+    - forward

docs/source/en/model_doc/clip.md

@@ -79,6 +79,123 @@ encode the text and prepare the images. The following example shows how to get t
 >>> probs = logits_per_image.softmax(dim=1)  # we can take the softmax to get the label probabilities
 ```
 
+
+### Combining CLIP and Flash Attention 2
+
+First, make sure to install the latest version of Flash Attention 2.
+
+```bash
+pip install -U flash-attn --no-build-isolation
+```
+
+Make also sure that you have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of flash-attn repository. Make also sure to load your model in half-precision (e.g. `torch.float16`)
+
+<Tip warning={true}>
+
+For small batch sizes, you might notice a slowdown in your model when using flash attention. Refer to the section [Expected speedups with Flash Attention and SDPA](#Expected-speedups-with-Flash-Attention-and-SDPA) below and select an appropriate attention implementation.
+
+</Tip>
+
+To load and run a model using Flash Attention 2, refer to the snippet below:
+
+```python
+>>> import torch
+>>> import requests
+>>> from PIL import Image
+
+>>> from transformers import CLIPProcessor, CLIPModel
+
+>>> device = "cuda"
+>>> torch_dtype = torch.float16
+
+>>> model = CLIPModel.from_pretrained(
+...     "openai/clip-vit-base-patch32",
+...     attn_implementation="flash_attention_2",
+...     device_map=device,
+...     torch_dtype=torch_dtype,
+... )
+>>> processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+>>> image = Image.open(requests.get(url, stream=True).raw)
+
+>>> inputs = processor(text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True)
+>>> inputs.to(device)
+
+>>> with torch.no_grad():
+...     with torch.autocast(device):
+...         outputs = model(**inputs)
+
+>>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
+>>> probs = logits_per_image.softmax(dim=1)  # we can take the softmax to get the label probabilities
+>>> print(probs)
+tensor([[0.9946, 0.0052]], device='cuda:0', dtype=torch.float16)
+```
+
+
+### Using Scaled Dot Product Attention (SDPA)
+
+PyTorch includes a native scaled dot-product attention (SDPA) operator as part of `torch.nn.functional`. This function 
+encompasses several implementations that can be applied depending on the inputs and the hardware in use. See the 
+[official documentation](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html) 
+or the [GPU Inference](https://huggingface.co/docs/transformers/main/en/perf_infer_gpu_one#pytorch-scaled-dot-product-attention)
+page for more information.
+
+SDPA is used by default for `torch>=2.1.1` when an implementation is available, but you may also set 
+`attn_implementation="sdpa"` in `from_pretrained()` to explicitly request SDPA to be used.
+
+```python
+from transformers import CLIPModel
+
+model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32", torch_dtype=torch.float16, attn_implementation="sdpa")
+```
+
+For the best speedups, we recommend loading the model in half-precision (e.g. `torch.float16` or `torch.bfloat16`).
+
+### Expected speedups with Flash Attention and SDPA
+
+On a local benchmark (NVIDIA A10G, PyTorch 2.3.1+cu121) with `float16`, we saw the following speedups during inference for `"openai/clip-vit-large-patch14"` checkpoint ([code](https://gist.github.com/qubvel/ac691a54e54f9fae8144275f866a7ff8)):
+
+#### CLIPTextModel
+
+|   Num text labels |   Eager (s/iter) |   FA2 (s/iter) |   FA2 speedup |   SDPA (s/iter) |   SDPA speedup |
+|------------------:|-----------------:|---------------:|--------------:|----------------:|---------------:|
+|                 4 |            0.009 |          0.012 |         0.737 |           0.007 |          1.269 |
+|                16 |            0.009 |          0.014 |         0.659 |           0.008 |          1.187 |
+|                32 |            0.018 |          0.021 |         0.862 |           0.016 |          1.142 |
+|                64 |            0.034 |          0.034 |         1.001 |           0.03  |          1.163 |
+|               128 |            0.063 |          0.058 |         1.09  |           0.054 |          1.174 |
+
+![clip_text_model_viz_3](https://github.com/user-attachments/assets/e9826b43-4e66-4f4c-952b-af4d90bd38eb)
+
+#### CLIPVisionModel
+
+|   Image batch size |   Eager (s/iter) |   FA2 (s/iter) |   FA2 speedup |   SDPA (s/iter) |   SDPA speedup |
+|-------------------:|-----------------:|---------------:|--------------:|----------------:|---------------:|
+|                  1 |            0.016 |          0.013 |         1.247 |           0.012 |          1.318 |
+|                  4 |            0.025 |          0.021 |         1.198 |           0.021 |          1.202 |
+|                 16 |            0.093 |          0.075 |         1.234 |           0.075 |          1.24  |
+|                 32 |            0.181 |          0.147 |         1.237 |           0.146 |          1.241 |
+
+![clip_image_model_viz_3](https://github.com/user-attachments/assets/50a36206-e3b9-4adc-ac8e-926b8b071d63)
+
+#### CLIPModel
+
+|   Image batch size |   Num text labels |   Eager (s/iter) |   FA2 (s/iter) |   FA2 speedup |   SDPA (s/iter) |   SDPA speedup |
+|-------------------:|------------------:|-----------------:|---------------:|--------------:|----------------:|---------------:|
+|                  1 |                 4 |            0.025 |          0.026 |         0.954 |           0.02  |          1.217 |
+|                  1 |                16 |            0.026 |          0.028 |         0.918 |           0.02  |          1.287 |
+|                  1 |                64 |            0.042 |          0.046 |         0.906 |           0.036 |          1.167 |
+|                  4 |                 4 |            0.028 |          0.033 |         0.849 |           0.024 |          1.189 |
+|                  4 |                16 |            0.034 |          0.035 |         0.955 |           0.029 |          1.169 |
+|                  4 |                64 |            0.059 |          0.055 |         1.072 |           0.05  |          1.179 |
+|                 16 |                 4 |            0.096 |          0.088 |         1.091 |           0.078 |          1.234 |
+|                 16 |                16 |            0.102 |          0.09  |         1.129 |           0.083 |          1.224 |
+|                 16 |                64 |            0.127 |          0.11  |         1.157 |           0.105 |          1.218 |
+|                 32 |                 4 |            0.185 |          0.159 |         1.157 |           0.149 |          1.238 |
+|                 32 |                16 |            0.19  |          0.162 |         1.177 |           0.154 |          1.233 |
+|                 32 |                64 |            0.216 |          0.181 |         1.19  |           0.176 |          1.228 |
+
 ## Resources
 
 A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with CLIP.

docs/source/en/model_doc/dac.md

@@ -0,0 +1,80 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# DAC
+
+## Overview
+
+
+The DAC model was proposed in [Descript Audio Codec: High-Fidelity Audio Compression with Improved RVQGAN](https://arxiv.org/abs/2306.06546) by Rithesh Kumar, Prem Seetharaman, Alejandro Luebs, Ishaan Kumar, Kundan Kumar.
+
+The Descript Audio Codec (DAC) model is a powerful tool for compressing audio data, making it highly efficient for storage and transmission. By compressing 44.1 KHz audio into tokens at just 8kbps bandwidth, the DAC model enables high-quality audio processing while significantly reducing the data footprint. This is particularly useful in scenarios where bandwidth is limited or storage space is at a premium, such as in streaming applications, remote conferencing, and archiving large audio datasets.
+
+The abstract from the paper is the following:
+
+*Language models have been successfully used to model natural signals, such as images, speech, and music. A key component of these models is a high quality neural compression model that can compress high-dimensional natural signals into lower dimensional discrete tokens. To that end, we introduce a high-fidelity universal neural audio compression algorithm that achieves ~90x compression of 44.1 KHz audio into tokens at just 8kbps bandwidth. We achieve this by combining advances in high-fidelity audio generation with better vector quantization techniques from the image domain, along with improved adversarial and reconstruction losses. We compress all domains (speech, environment, music, etc.) with a single universal model, making it widely applicable to generative modeling of all audio. We compare with competing audio compression algorithms, and find our method outperforms them significantly. We provide thorough ablations for every design choice, as well as open-source code and trained model weights. We hope our work can lay the foundation for the next generation of high-fidelity audio modeling.*
+
+This model was contributed by [Kamil Akesbi](https://huggingface.co/kamilakesbi).
+The original code can be found [here](https://github.com/descriptinc/descript-audio-codec/tree/main?tab=readme-ov-file).
+
+
+## Model structure
+
+The Descript Audio Codec (DAC) model is structured into three distinct stages:
+
+1. Encoder Model: This stage compresses the input audio, reducing its size while retaining essential information.
+2. Residual Vector Quantizer (RVQ) Model: Working in tandem with the encoder, this model quantizes the latent codes of the audio, refining the compression and ensuring high-quality reconstruction.
+3. Decoder Model: This final stage reconstructs the audio from its compressed form, restoring it to a state that closely resembles the original input.
+
+## Usage example 
+
+Here is a quick example of how to encode and decode an audio using this model: 
+
+```python 
+>>> from datasets import load_dataset, Audio
+>>> from transformers import DacModel, AutoProcessor
+>>> librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+>>> model = DacModel.from_pretrained("descript/dac_16khz")
+>>> processor = AutoProcessor.from_pretrained("descript/dac_16khz")
+>>> librispeech_dummy = librispeech_dummy.cast_column("audio", Audio(sampling_rate=processor.sampling_rate))
+>>> audio_sample = librispeech_dummy[-1]["audio"]["array"]
+>>> inputs = processor(raw_audio=audio_sample, sampling_rate=processor.sampling_rate, return_tensors="pt")
+
+>>> encoder_outputs = model.encode(inputs["input_values"])
+>>> # Get the intermediate audio codes
+>>> audio_codes = encoder_outputs.audio_codes
+>>> # Reconstruct the audio from its quantized representation
+>>> audio_values = model.decode(encoder_outputs.quantized_representation)
+>>> # or the equivalent with a forward pass
+>>> audio_values = model(inputs["input_values"]).audio_values
+```
+
+## DacConfig
+
+[[autodoc]] DacConfig
+
+## DacFeatureExtractor
+
+[[autodoc]] DacFeatureExtractor
+    - __call__
+
+## DacModel
+
+[[autodoc]] DacModel
+    - decode
+    - encode
+    - forward

docs/source/en/model_doc/depth_anything.md

@@ -20,6 +20,12 @@ rendered properly in your Markdown viewer.
 
 The Depth Anything model was proposed in [Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data](https://arxiv.org/abs/2401.10891) by Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, Hengshuang Zhao. Depth Anything is based on the [DPT](dpt) architecture, trained on ~62 million images, obtaining state-of-the-art results for both relative and absolute depth estimation.
 
+<Tip>
+
+[Depth Anything V2](depth_anything_v2) was released in June 2024. It uses the same architecture as Depth Anything and therefore it is compatible with all code examples and existing workflows. However, it leverages synthetic data and a larger capacity teacher model to achieve much finer and robust depth predictions.
+
+</Tip>
+
 The abstract from the paper is the following:
 
 *This work presents Depth Anything, a highly practical solution for robust monocular depth estimation. Without pursuing novel technical modules, we aim to build a simple yet powerful foundation model dealing with any images under any circumstances. To this end, we scale up the dataset by designing a data engine to collect and automatically annotate large-scale unlabeled data (~62M), which significantly enlarges the data coverage and thus is able to reduce the generalization error. We investigate two simple yet effective strategies that make data scaling-up promising. First, a more challenging optimization target is created by leveraging data augmentation tools. It compels the model to actively seek extra visual knowledge and acquire robust representations. Second, an auxiliary supervision is developed to enforce the model to inherit rich semantic priors from pre-trained encoders. We evaluate its zero-shot capabilities extensively, including six public datasets and randomly captured photos. It demonstrates impressive generalization ability. Further, through fine-tuning it with metric depth information from NYUv2 and KITTI, new SOTAs are set. Our better depth model also results in a better depth-conditioned ControlNet.*

docs/source/en/model_doc/depth_anything_v2.md

@@ -0,0 +1,115 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Depth Anything V2
+
+## Overview
+
+Depth Anything V2 was introduced in [the paper of the same name](https://arxiv.org/abs/2406.09414) by Lihe Yang et al. It uses the same architecture as the original [Depth Anything model](depth_anything), but uses synthetic data and a larger capacity teacher model to achieve much finer and robust depth predictions.
+
+The abstract from the paper is the following:
+
+*This work presents Depth Anything V2. Without pursuing fancy techniques, we aim to reveal crucial findings to pave the way towards building a powerful monocular depth estimation model. Notably, compared with V1, this version produces much finer and more robust depth predictions through three key practices: 1) replacing all labeled real images with synthetic images, 2) scaling up the capacity of our teacher model, and 3) teaching student models via the bridge of large-scale pseudo-labeled real images. Compared with the latest models built on Stable Diffusion, our models are significantly more efficient (more than 10x faster) and more accurate. We offer models of different scales (ranging from 25M to 1.3B params) to support extensive scenarios. Benefiting from their strong generalization capability, we fine-tune them with metric depth labels to obtain our metric depth models. In addition to our models, considering the limited diversity and frequent noise in current test sets, we construct a versatile evaluation benchmark with precise annotations and diverse scenes to facilitate future research.*
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/depth_anything_overview.jpg"
+alt="drawing" width="600"/>
+
+<small> Depth Anything overview. Taken from the <a href="https://arxiv.org/abs/2401.10891">original paper</a>.</small>
+
+The Depth Anything models were contributed by [nielsr](https://huggingface.co/nielsr).
+The original code can be found [here](https://github.com/DepthAnything/Depth-Anything-V2).
+
+## Usage example
+
+There are 2 main ways to use Depth Anything V2: either using the pipeline API, which abstracts away all the complexity for you, or by using the `DepthAnythingForDepthEstimation` class yourself.
+
+### Pipeline API
+
+The pipeline allows to use the model in a few lines of code:
+
+```python
+>>> from transformers import pipeline
+>>> from PIL import Image
+>>> import requests
+
+>>> # load pipe
+>>> pipe = pipeline(task="depth-estimation", model="depth-anything/Depth-Anything-V2-Small-hf")
+
+>>> # load image
+>>> url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
+>>> image = Image.open(requests.get(url, stream=True).raw)
+
+>>> # inference
+>>> depth = pipe(image)["depth"]
+```
+
+### Using the model yourself
+
+If you want to do the pre- and post-processing yourself, here's how to do that:
+
+```python
+>>> from transformers import AutoImageProcessor, AutoModelForDepthEstimation
+>>> import torch
+>>> import numpy as np
+>>> from PIL import Image
+>>> import requests
+
+>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+>>> image = Image.open(requests.get(url, stream=True).raw)
+
+>>> image_processor = AutoImageProcessor.from_pretrained("depth-anything/Depth-Anything-V2-Small-hf")
+>>> model = AutoModelForDepthEstimation.from_pretrained("depth-anything/Depth-Anything-V2-Small-hf")
+
+>>> # prepare image for the model
+>>> inputs = image_processor(images=image, return_tensors="pt")
+
+>>> with torch.no_grad():
+...     outputs = model(**inputs)
+...     predicted_depth = outputs.predicted_depth
+
+>>> # interpolate to original size
+>>> prediction = torch.nn.functional.interpolate(
+...     predicted_depth.unsqueeze(1),
+...     size=image.size[::-1],
+...     mode="bicubic",
+...     align_corners=False,
+... )
+
+>>> # visualize the prediction
+>>> output = prediction.squeeze().cpu().numpy()
+>>> formatted = (output * 255 / np.max(output)).astype("uint8")
+>>> depth = Image.fromarray(formatted)
+```
+
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with Depth Anything.
+
+- [Monocular depth estimation task guide](../tasks/depth_estimation)
+- [Depth Anything V2 demo](https://huggingface.co/spaces/depth-anything/Depth-Anything-V2).
+- A notebook showcasing inference with [`DepthAnythingForDepthEstimation`] can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/Depth%20Anything/Predicting_depth_in_an_image_with_Depth_Anything.ipynb). 🌎
+- [Core ML conversion of the `small` variant for use on Apple Silicon](https://huggingface.co/apple/coreml-depth-anything-v2-small).
+
+If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
+
+## DepthAnythingConfig
+
+[[autodoc]] DepthAnythingConfig
+
+## DepthAnythingForDepthEstimation
+
+[[autodoc]] DepthAnythingForDepthEstimation
+    - forward

docs/source/en/model_doc/dinov2.md

@@ -57,7 +57,7 @@ print((last_hidden_states - traced_outputs[0]).abs().max())
 
 ## Resources
 
-A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with DPT.
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with DINOv2.
 
 - Demo notebooks for DINOv2 can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/DINOv2). 🌎
 
@@ -72,6 +72,9 @@ If you're interested in submitting a resource to be included here, please feel f
 
 [[autodoc]] Dinov2Config
 
+<frameworkcontent>
+<pt>
+
 ## Dinov2Model
 
 [[autodoc]] Dinov2Model
@@ -81,3 +84,20 @@ If you're interested in submitting a resource to be included here, please feel f
 
 [[autodoc]] Dinov2ForImageClassification
     - forward
+
+</pt>
+<jax>
+
+## FlaxDinov2Model
+
+[[autodoc]] FlaxDinov2Model
+    - __call__
+
+
+## FlaxDinov2ForImageClassification
+
+[[autodoc]] FlaxDinov2ForImageClassification
+    - __call__
+
+</jax>
+</frameworkcontent>

docs/source/en/model_doc/falcon_mamba.md

@@ -0,0 +1,116 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# FalconMamba
+
+## Overview
+
+The FalconMamba model was proposed by TII UAE (Technology Innovation Institute) in their release.
+
+The abstract from the paper is the following:
+
+*We present FalconMamba, a new base large language model based on the novel Mamba architecture. FalconMamba is trained on 5.8 trillion tokens with carefully selected data mixtures. As a pure Mamba-based model, FalconMamba surpasses leading open-weight models based on Transformers, such as Mistral 7B, Llama3 8B, and Falcon2 11B. It is on par with Gemma 7B and outperforms models with different architecture designs, such as RecurrentGemma 9B. Currently, FalconMamba is the best-performing Mamba model in the literature at this scale, surpassing both existing Mamba and hybrid Mamba-Transformer models.
+Due to its architecture, FalconMamba is significantly faster at inference and requires substantially less memory for long sequence generation. Despite recent studies suggesting that hybrid Mamba-Transformer models outperform pure architecture designs, we argue and demonstrate that the pure Mamba design can achieve similar, even superior results compared to the hybrid design. We make the weights of our implementation of FalconMamba publicly available under a permissive license.*
+
+Tips:
+
+- FalconMamba is mostly based on Mamba architecutre, the same [tips and best practices](./mamba) would be relevant here.
+
+The model has been trained on approximtely 6T tokens consisting a mixture of many data sources such as RefineWeb, Cosmopedia and Math data.
+
+For more details about the training procedure and the architecture, have a look at [the technical paper of FalconMamba]() (coming soon).
+
+# Usage
+
+Below we demonstrate how to use the model:
+
+```python 
+from transformers import FalconMambaForCausalLM, AutoTokenizer
+import torch
+
+tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b")
+model = FalconMambaForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b")
+
+input_ids = tokenizer("Hey how are you doing?", return_tensors= "pt")["input_ids"]
+
+out = model.generate(input_ids, max_new_tokens=10)
+print(tokenizer.batch_decode(out))
+```
+
+The architecture is also compatible with `torch.compile` for faster generation:
+
+```python 
+from transformers import FalconMambaForCausalLM, AutoTokenizer
+import torch
+
+tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b")
+model = FalconMambaForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b", torch_dtype=torch.bfloat16).to(0)
+model = torch.compile(model)
+
+input_ids = tokenizer("Hey how are you doing?", return_tensors= "pt")["input_ids"]
+
+out = model.generate(input_ids, max_new_tokens=10)
+print(tokenizer.batch_decode(out))
+```
+
+If you have access to a GPU that is compatible with `bitsandbytes`, you can also quantize the model in 4-bit precision:
+
+```python 
+from transformers import FalconMambaForCausalLM, AutoTokenizer, BitsAndBytesConfig
+import torch
+
+tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b")
+quantization_config = BitsAndBytesConfig(load_in_4bit=True)
+model = FalconMambaForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b", quantization_config=quantization_config)
+
+input_ids = tokenizer("Hey how are you doing?", return_tensors= "pt")["input_ids"]
+
+out = model.generate(input_ids, max_new_tokens=10)
+print(tokenizer.batch_decode(out))
+```
+
+You can also play with the instruction fine-tuned model:
+
+```python 
+from transformers import FalconMambaForCausalLM, AutoTokenizer
+import torch
+
+tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-instruct")
+model = FalconMambaForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-instruct")
+
+# We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating
+messages = [
+    {"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
+]
+input_ids = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True).input_ids
+
+outputs = model.generate(input_ids)
+print(tokenizer.decode(outputs[0]))
+```
+
+## FalconMambaConfig
+
+[[autodoc]] FalconMambaConfig
+
+## FalconMambaModel
+
+[[autodoc]] FalconMambaModel
+    - forward
+
+## FalconMambaLMHeadModel
+
+[[autodoc]] FalconMambaForCausalLM
+    - forward

docs/source/en/model_doc/gemma2.md

@@ -0,0 +1,64 @@
+
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Gemma2
+
+## Overview
+
+The Gemma2 model was proposed in [Gemma2: Open Models Based on Gemini Technology and Research](https://blog.google/technology/developers/google-gemma-2/) by Gemma2 Team, Google.
+Two Gemma2 models are released, with parameters sizes of 9 billion (9B) and 27 billion (27B).
+
+The abstract from the blog post is the following:
+
+*Now we’re officially releasing Gemma 2 to researchers and developers globally. Available in both 9 billion (9B) and 27 billion (27B) parameter sizes, Gemma 2 is higher-performing and more efficient at inference than the first generation, with significant safety advancements built in. In fact, at 27B, it offers competitive alternatives to models more than twice its size, delivering the kind of performance that was only possible with proprietary models as recently as December.*
+
+Tips:
+
+- The original checkpoints can be converted using the conversion script `src/transformers/models/Gemma2/convert_Gemma2_weights_to_hf.py` 
+
+<Tip warning={true}>
+
+- Gemma2 uses sliding window attention every second layer, which makes it unsuitable for typical kv caching with [`~DynamicCache`] or tuples of tensors. To enable caching in Gemma2 forward call, you must initialize a [`~HybridCache`] instance and pass it as `past_key_values` to the forward call. Note, that you also have to prepare `cache_position` if the `past_key_values` already contains previous keys and values.
+
+</Tip>
+
+This model was contributed by [Arthur Zucker](https://huggingface.co/ArthurZ), [Pedro Cuenca](https://huggingface.co/pcuenq) and [Tom Arsen]().
+
+
+## Gemma2Config
+
+[[autodoc]] Gemma2Config
+
+## Gemma2Model
+
+[[autodoc]] Gemma2Model
+    - forward
+
+## Gemma2ForCausalLM
+
+[[autodoc]] Gemma2ForCausalLM
+    - forward
+
+## Gemma2ForSequenceClassification
+
+[[autodoc]] Gemma2ForSequenceClassification
+    - forward
+
+## Gemma2ForTokenClassification
+
+[[autodoc]] Gemma2ForTokenClassification
+    - forward

docs/source/en/model_doc/granite.md

@@ -0,0 +1,74 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ 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.
+
+-->
+
+# Granite
+
+## Overview
+
+The Granite model was proposed in [Power Scheduler: A Batch Size and Token Number Agnostic Learning Rate Scheduler](https://arxiv.org/abs/2408.13359) by Yikang Shen, Matthew Stallone, Mayank Mishra, Gaoyuan Zhang, Shawn Tan, Aditya Prasad, Adriana Meza Soria, David D. Cox and Rameswar Panda.
+
+PowerLM-3B is a 3B state-of-the-art small language model trained with the Power learning rate scheduler. It is trained on a wide range of open-source and synthetic datasets with permissive licenses. PowerLM-3B has shown promising results compared to other models in the size categories across various benchmarks, including natural language multi-choices, code generation, and math reasoning.
+
+The abstract from the paper is the following:
+
+*Finding the optimal learning rate for language model pretraining is a challenging task.
+This is not only because there is a complicated correlation between learning rate, batch size, number of training tokens, model size, and other hyperparameters but also because it is prohibitively expensive to perform a hyperparameter search for large language models with Billions or Trillions of parameters. Recent studies propose using small proxy models and small corpus to perform hyperparameter searches and transposing the optimal parameters to large models and large corpus. While the zero-shot transferability is theoretically and empirically proven for model size related hyperparameters, like depth and width, the zero-shot transfer from small corpus to large corpus is underexplored.
+In this paper, we study the correlation between optimal learning rate, batch size, and number of training tokens for the recently proposed WSD scheduler. After thousands of small experiments, we found a power-law relationship between variables and demonstrated its transferability across model sizes. Based on the observation, we propose a new learning rate scheduler, Power scheduler, that is agnostic about the number of training tokens and batch size. The experiment shows that combining the Power scheduler with Maximum Update Parameterization (\mup) can consistently achieve impressive performance with one set of hyperparameters regardless of the number of training tokens, batch size, model size, and even model architecture. Our 3B dense and MoE models trained with the Power scheduler achieve comparable performance as state-of-the-art small language models.
+We [open source](https://huggingface.co/collections/ibm/power-lm-66be64ae647ddf11b9808000) these pretrained models.*
+
+Tips:
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_path = "ibm/PowerLM-3b"
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+# drop device_map if running on CPU
+model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto")
+model.eval()
+
+# change input text as desired
+prompt = "Write a code to find the maximum value in a list of numbers."
+
+# tokenize the text
+input_tokens = tokenizer(prompt, return_tensors="pt")
+# generate output tokens
+output = model.generate(**input_tokens, max_new_tokens=100)
+# decode output tokens into text
+output = tokenizer.batch_decode(output)
+# loop over the batch to print, in this example the batch size is 1
+for i in output:
+    print(i)
+```
+
+This model was contributed by [mayank-mishra](https://huggingface.co/mayank-mishra).
+
+
+## GraniteConfig
+
+[[autodoc]] GraniteConfig
+
+## GraniteModel
+
+[[autodoc]] GraniteModel
+    - forward
+
+## GraniteForCausalLM
+
+[[autodoc]] GraniteForCausalLM
+    - forward

docs/source/en/model_doc/grounding-dino.md

@@ -41,33 +41,40 @@ The original code can be found [here](https://github.com/IDEA-Research/Grounding
 Here's how to use the model for zero-shot object detection:
 
 ```python
-import requests
+>>> import requests
 
-import torch
-from PIL import Image
-from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection, 
+>>> import torch
+>>> from PIL import Image
+>>> from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
 
-model_id = "IDEA-Research/grounding-dino-tiny"
+>>> model_id = "IDEA-Research/grounding-dino-tiny"
+>>> device = "cuda"
 
-processor = AutoProcessor.from_pretrained(model_id)
-model = AutoModelForZeroShotObjectDetection.from_pretrained(model_id).to(device)
+>>> processor = AutoProcessor.from_pretrained(model_id)
+>>> model = AutoModelForZeroShotObjectDetection.from_pretrained(model_id).to(device)
 
-image_url = "http://images.cocodataset.org/val2017/000000039769.jpg"
-image = Image.open(requests.get(image_url, stream=True).raw)
-# Check for cats and remote controls
-text = "a cat. a remote control."
+>>> image_url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+>>> image = Image.open(requests.get(image_url, stream=True).raw)
+>>> # Check for cats and remote controls
+>>> text = "a cat. a remote control."
 
-inputs = processor(images=image, text=text, return_tensors="pt").to(device)
-with torch.no_grad():
-    outputs = model(**inputs)
+>>> inputs = processor(images=image, text=text, return_tensors="pt").to(device)
+>>> with torch.no_grad():
+...     outputs = model(**inputs)
 
-results = processor.post_process_grounded_object_detection(
-    outputs,
-    inputs.input_ids,
-    box_threshold=0.4,
-    text_threshold=0.3,
-    target_sizes=[image.size[::-1]]
-)
+>>> results = processor.post_process_grounded_object_detection(
+...     outputs,
+...     inputs.input_ids,
+...     box_threshold=0.4,
+...     text_threshold=0.3,
+...     target_sizes=[image.size[::-1]]
+... )
+>>> print(results)
+[{'boxes': tensor([[344.6959,  23.1090, 637.1833, 374.2751],
+        [ 12.2666,  51.9145, 316.8582, 472.4392],
+        [ 38.5742,  70.0015, 176.7838, 118.1806]], device='cuda:0'),
+  'labels': ['a cat', 'a cat', 'a remote control'],
+  'scores': tensor([0.4785, 0.4381, 0.4776], device='cuda:0')}]
 ```
 
 ## Grounded SAM

docs/source/en/model_doc/hiera.md

@@ -0,0 +1,62 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Hiera
+
+## Overview
+
+Hiera was proposed in [Hiera: A Hierarchical Vision Transformer without the Bells-and-Whistles](https://arxiv.org/abs/2306.00989) by Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, Jitendra Malik, Yanghao Li, Christoph Feichtenhofer
+
+The paper introduces "Hiera," a hierarchical Vision Transformer that simplifies the architecture of modern hierarchical vision transformers by removing unnecessary components without compromising on accuracy or efficiency. Unlike traditional transformers that add complex vision-specific components to improve supervised classification performance, Hiera demonstrates that such additions, often termed "bells-and-whistles," are not essential for high accuracy. By leveraging a strong visual pretext task (MAE) for pretraining, Hiera retains simplicity and achieves superior accuracy and speed both in inference and training across various image and video recognition tasks. The approach suggests that spatial biases required for vision tasks can be effectively learned through proper pretraining, eliminating the need for added architectural complexity. 
+
+The abstract from the paper is the following:
+
+*Modern hierarchical vision transformers have added several vision-specific components in the pursuit of supervised classification performance. While these components lead to effective accuracies and attractive FLOP counts, the added complexity actually makes these transformers slower than their vanilla ViT counterparts. In this paper, we argue that this additional bulk is unnecessary. By pretraining with a strong visual pretext task (MAE), we can strip out all the bells-and-whistles from a state-of-the-art multi-stage vision transformer without losing accuracy. In the process, we create Hiera, an extremely simple hierarchical vision transformer that is more accurate than previous models while being significantly faster both at inference and during training. We evaluate Hiera on a variety of tasks for image and video recognition. Our code and models are available at https://github.com/facebookresearch/hiera.*
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/hiera_overview.png"
+alt="drawing" width="600"/>
+
+<small> Hiera architecture. Taken from the <a href="https://arxiv.org/abs/2306.00989">original paper.</a> </small>
+
+This model was a joint contibution by [EduardoPacheco](https://huggingface.co/EduardoPacheco) and [namangarg110](https://huggingface.co/namangarg110). The original code can be found [here] (https://github.com/facebookresearch/hiera).
+
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with Hiera. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
+
+<PipelineTag pipeline="image-classification"/>
+
+- [`HieraForImageClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification.ipynb).
+- See also: [Image classification task guide](../tasks/image_classification)
+
+## HieraConfig
+
+[[autodoc]] HieraConfig
+
+## HieraModel
+
+[[autodoc]] HieraModel
+    - forward
+
+## HieraForPreTraining
+
+[[autodoc]] HieraForPreTraining
+    - forward
+  
+## HieraForImageClassification
+
+[[autodoc]] HieraForImageClassification
+    - forward

docs/source/en/model_doc/llama3.md

@@ -16,6 +16,15 @@ rendered properly in your Markdown viewer.
 
 # Llama3
 
+```py3
+import transformers
+import torch
+
+model_id = "meta-llama/Meta-Llama-3-8B"
+
+pipeline = transformers.pipeline("text-generation", model=model_id, model_kwargs={"torch_dtype": torch.bfloat16}, device_map="auto")
+pipeline("Hey how are you doing today?")
+```
 
 ## Overview
 
@@ -66,20 +75,7 @@ model = AutoModelForCausalLM.from_pretrained("/output/path")
 Note that executing the script requires enough CPU RAM to host the whole model in float16 precision (even if the biggest versions
 come in several checkpoints they each contain a part of each weight of the model, so we need to load them all in RAM). For the 75B model, it's thus 145GB of RAM needed.
 
-
 - When using Flash Attention 2 via `attn_implementation="flash_attention_2"`, don't pass `torch_dtype` to the `from_pretrained` class method and use Automatic Mixed-Precision training. When using `Trainer`, it is simply specifying either `fp16` or `bf16` to `True`. Otherwise, make sure you are using `torch.autocast`. This is required because the Flash Attention only support `fp16` and `bf16` data type.
 
-## Quick usage
-
-```py3
-import transformers
-import torch
-
-model_id = "meta-llama/Meta-Llama-3-8B"
-
-pipeline = transformers.pipeline("text-generation", model=model_id, model_kwargs={"torch_dtype": torch.bfloat16}, device_map="auto")
-pipeline("Hey how are you doing today?")
-```
-
 ## Resources
-A ton of cool resources are already available on the documentation page of [~llama2], inviting contributors to add new resources curated for Llama3 here! 🤗
+A ton of cool resources are already available on the documentation page of [Llama2](./llama2), inviting contributors to add new resources curated for Llama3 here! 🤗

docs/source/en/model_doc/llava.md

@@ -40,8 +40,55 @@ The original code can be found [here](https://github.com/haotian-liu/LLaVA/tree/
 
 - Note the model has not been explicitly trained to process multiple images in the same prompt, although this is technically possible, you may experience inaccurate results.
 
-- For better results, we recommend users to prompt the model with the correct prompt format: 
+- For better results, we recommend users to use the processor's `apply_chat_template()` method to format your prompt correctly. For that you need to construct a conversation history, passing in a plain string will not format your prompt. Each message in the conversation history for chat templates is a dictionary with keys "role" and "content". The "content" should be a list of dictionaries, for "text" and "image" modalities, as follows:
 
+```python
+from transformers import AutoProcessor
+
+processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")
+
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": "What’s shown in this image?"},
+            ],
+    },
+    {
+        "role": "assistant",
+        "content": [{"type": "text", "text": "This image shows a red stop sign."},]
+    },
+    {
+
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "Describe the image in more details."},
+        ],
+    },
+]
+
+text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
+
+# Note that the template simply formats your prompt, you still have to tokenize it and obtain pixel values for your images
+print(text_prompt)
+>>> "USER: <image>\n<What’s shown in this image? ASSISTANT: This image shows a red stop sign.</s>USER: Describe the image in more details. ASSISTANT:"
+```
+
+- If you want to construct a chat prompt yourself, below is a list of prompt formats accepted by each llava checkpoint:
+
+[llava-interleave models](https://huggingface.co/collections/llava-hf/llava-interleave-668e19a97da0036aad4a2f19) requires the following format:
+```bash
+"<|im_start|>user <image>\nWhat is shown in this image?<|im_end|><|im_start|>assistant"
+```
+
+For multiple turns conversation:
+
+```bash
+"<|im_start|>user <image>\n<prompt1><|im_end|><|im_start|>assistant <answer1><|im_end|><|im_start|>user <image>\n<prompt1><|im_end|><|im_start|>assistant "
+```
+
+[llava-1.5 models](https://huggingface.co/collections/llava-hf/llava-15-65f762d5b6941db5c2ba07e0) requires the following format:
 ```bash
 "USER: <image>\n<prompt> ASSISTANT:"
 ```
@@ -52,6 +99,7 @@ For multiple turns conversation:
 "USER: <image>\n<prompt1> ASSISTANT: <answer1></s>USER: <prompt2> ASSISTANT: <answer2></s>USER: <prompt3> ASSISTANT:"
 ```
 
+
 ### Using Flash Attention 2
 
 Flash Attention 2 is an even faster, optimized version of the previous optimization, please refer to the [Flash Attention 2 section of performance docs](https://huggingface.co/docs/transformers/perf_infer_gpu_one).

docs/source/en/model_doc/llava_next.md

@@ -46,26 +46,79 @@ The original code can be found [here](https://github.com/haotian-liu/LLaVA/tree/
 
 - We advise users to use `padding_side="left"` when computing batched generation as it leads to more accurate results. Simply make sure to call `processor.tokenizer.padding_side = "left"` before generating.
 
-- Note that each checkpoint has been trained with a specific prompt format, depending on which large language model (LLM) was used. Below, we list the correct prompt formats to use for the text prompt "What is shown in this image?":
+<Tip warning={true}>
 
+- Llava-Next uses different number of patches for images and thus has to pad the inputs inside modeling code, aside from the padding done when processing the inputs. The default setting is "left-padding" if model is in `eval()` mode, otherwise "right-padding".
+
+</Tip>
+
+
+- Note that each checkpoint has been trained with a specific prompt format, depending on which large language model (LLM) was used. You can use the processor's `apply_chat_template` to format your prompts correctly. For that you have to construct a conversation history, passing a plain string will not format your prompt. Each message in the conversation history for chat templates is a dictionary with keys "role" and "content". The "content" should be a list of dictionaries, for "text" and "image" modalities. Below is an example of how to do that and the list of formats accepted by each checkpoint.
+
+We will use [llava-v1.6-mistral-7b-hf](https://huggingface.co/llava-hf/llava-v1.6-mistral-7b-hf) and a conversation history of text and image. Each content field has to be a list of dicts, as follows:
+
+```python
+from transformers import LlavaNextProcessor
+
+processor = LlavaNextProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")
+
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": "What’s shown in this image?"},
+        ],
+    },
+    {
+        "role": "assistant",
+        "content": [{"type": "text", "text": "This image shows a red stop sign."},]
+    },
+    {
+
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "Describe the image in more details."},
+        ],
+    },
+]
+
+text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
+
+# Note that the template simply formats your prompt, you still have to tokenize it and obtain pixel values for your images
+print(text_prompt)
+>>> "[INST] <image>\nWhat's shown in this image? [/INST] This image shows a red stop sign. [INST] Describe the image in more details. [/INST]"
+```
+
+- If you want to construct a chat prompt yourself, below is a list of possible formats
+.
 [llava-v1.6-mistral-7b-hf](https://huggingface.co/llava-hf/llava-v1.6-mistral-7b-hf) requires the following format:
-
 ```bash
 "[INST] <image>\nWhat is shown in this image? [/INST]"
 ```
 
 [llava-v1.6-vicuna-7b-hf](https://huggingface.co/llava-hf/llava-v1.6-vicuna-7b-hf) and [llava-v1.6-vicuna-13b-hf](https://huggingface.co/llava-hf/llava-v1.6-vicuna-13b-hf) require the following format:
-
 ```bash
 "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. USER: <image>\nWhat is shown in this image? ASSISTANT:"
 ```
 
 [llava-v1.6-34b-hf](https://huggingface.co/llava-hf/llava-v1.6-34b-hf) requires the following format:
-
 ```bash
 "<|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nWhat is shown in this image?<|im_end|><|im_start|>assistant\n"
 ```
 
+[llama3-llava-next-8b-hf](https://huggingface.co/llava-hf/llava-next-8b-hf) requires the following format:
+
+```bash
+"<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful language and vision assistant. You are able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language.<|eot_id|><|start_header_id|><|start_header_id|>user<|end_header_id|>\n\n<image>\nWhat is shown in this image?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+```
+
+[llava-next-72b-hf](https://huggingface.co/llava-hf/llava-next-72b-hf) and [llava-next-110b-hf](https://huggingface.co/llava-hf/llava-next-110b-hf) require the following format:
+
+```bash
+"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<image>\nWhat is shown in this image?<|im_end|>\n<|im_start|>assistant\n"
+```
+
 ## Usage example
 
 ### Single image inference
@@ -86,8 +139,17 @@ model.to("cuda:0")
 # prepare image and text prompt, using the appropriate prompt template
 url = "https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true"
 image = Image.open(requests.get(url, stream=True).raw)
-prompt = "[INST] <image>\nWhat is shown in this image? [/INST]"
 
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": "What is shown in this image?"},
+        ],
+    },
+]
+prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
 inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")
 
 # autoregressively complete prompt
@@ -120,15 +182,47 @@ image_cats = Image.open(requests.get(url, stream=True).raw)
 url = "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/snowman.jpg"
 image_snowman = Image.open(requests.get(url, stream=True).raw)
 
-# Prepare a batched prompt, where the first one is a multi-turn conversation and the second is not
-prompt = [
-    "[INST] <image>\nWhat is shown in this image? [/INST] There is a red stop sign in the image. [INST] <image>\nWhat about this image? How many cats do you see [/INST]",
-    "[INST] <image>\nWhat is shown in this image? [/INST]"
+# Prepare a batch of two prompts, where the first one is a multi-turn conversation and the second is not
+conversation_1 = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": "What is shown in this image?"},
+            ],
+    },
+    {
+        "role": "assistant",
+        "content": [
+            {"type": "text", "text": "There is a red stop sign in the image."},
+            ],
+    },
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": "What about this image? How many cats do you see?"},
+            ],
+    },
 ]
 
+conversation_2 = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": "What is shown in this image?"},
+            ],
+    },
+]
+
+prompt_1 = processor.apply_chat_template(conversation_1, add_generation_prompt=True)
+prompt_2 = processor.apply_chat_template(conversation_2, add_generation_prompt=True)
+prompts = [prompt_1, prompt_2]
+
 # We can simply feed images in the order they have to be used in the text prompt
 # Each "<image>" token uses one image leaving the next for the subsequent "<image>" tokens
-inputs = processor(text=prompt, images=[image_stop, image_cats, image_snowman], padding=True, return_tensors="pt").to(model.device)
+inputs = processor(text=prompts, images=[image_stop, image_cats, image_snowman], padding=True, return_tensors="pt").to(model.device)
 
 # Generate
 generate_ids = model.generate(**inputs, max_new_tokens=30)

docs/source/en/model_doc/llava_next_video.md

@@ -43,6 +43,13 @@ The original code can be found [here](https://github.com/LLaVA-VL/LLaVA-NeXT/tre
 
 - We advise users to use `padding_side="left"` when computing batched generation as it leads to more accurate results. Simply make sure to call `processor.tokenizer.padding_side = "left"` before generating.
 
+<Tip warning={true}>
+
+- Llava-Next uses different number of patches for images and thus has to pad the inputs inside modeling code, aside from the padding done when processing the inputs. The default setting is "left-padding" if model is in `eval()` mode, otherwise "right-padding".
+
+</Tip>
+
+
 - Note that each checkpoint has been trained with a specific prompt format, depending on which large language model (LLM) was used. You can use tokenizer's `apply_chat_template` to format your prompts correctly. Below is an example of how to do that.
 
 We will use [LLaVA-NeXT-Video-7B-hf](https://huggingface.co/llava-hf/LLaVA-NeXT-Video-7B-hf) and a conversation history of videos and images. Each content field has to be a list of dicts, as follows:

docs/source/en/model_doc/mamba2.md

@@ -0,0 +1,106 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Mamba 2
+
+## Overview
+
+The Mamba2 model was proposed in [Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality](https://arxiv.org/abs/2405.21060) by Tri Dao and Albert Gu. It is a State Space Model similar to Mamba 1, with better performances in a simplified architecture. 
+
+
+The abstract from the paper is the following:
+
+*While Transformers have been the main architecture behind deep learning's success in language modeling, state-space models (SSMs) such as Mamba have recently been shown to match or outperform Transformers at small to medium scale. We show that these families of models are actually quite closely related, and develop a rich framework of theoretical connections between SSMs and variants of attention, connected through various decompositions of a well-studied class of structured semiseparable matrices. Our state space duality (SSD) framework allows us to design a new architecture (Mamba-2) whose core layer is an a refinement of Mamba's selective SSM that is 2-8X faster, while continuing to be competitive with Transformers on language modeling.*
+
+Tips:
+
+This version should support all implementations of Mamba 2, and in particular [Mamba-2 codestral](https://huggingface.co/mistralai/Mamba-Codestral-7B-v0.1) from Mistral AI. In particular, mamba 2 codestral was released with a number of `groups` equal to 8, which can be thought intuitively as similar to the number of kv heads in an attention-based model. 
+This model has two different forward passes, `torch_forward` or `cuda_kernels_forward`. The latter uses the original cuda kernels if they are found in your environment, and is slower on the prefill i.e. requires a "warmup run" due to high cpu overhead, see [here](https://github.com/state-spaces/mamba/issues/389#issuecomment-2171755306) and [also here](https://github.com/state-spaces/mamba/issues/355#issuecomment-2147597457). Without compilation, the `torch_forward` implementation is faster by a factor 3 to 4. Further, there are no positional embeddings in this model, but there is an `attention_mask` and a specific logic to mask out hidden states in two places in the case of batched generation, see [here](https://github.com/state-spaces/mamba/issues/66#issuecomment-1863563829) as well. Due to this, in addition to the reimplementation of mamba2 kernels, batched generation and cached generation are expected to have slight discrepancies. Further, the results given by the cuda kernels or the torch forward are expected to be slightly different. The SSM algorithm heavily relies on tensor contractions, which have matmul equivalents but the order of operations is slightly different, making the difference greater at smaller precisions. 
+Another note, shutdown of hidden states corresponding to padding tokens is done in 2 places and mostly has been tested with left-padding. Right-padding will propagate noise down the line and is not guaranteed to yield satisfactory results. `tokenizer.padding_side = "left"` ensures you are using the correct padding side.
+
+This model was contributed by [Molbap](https://huggingface.co/Molbap), with tremendous help from [Anton Vlasjuk](https://github.com/vasqu).
+The original code can be found [here](https://github.com/state-spaces/mamba).
+
+
+# Usage
+
+### A simple generation example: 
+```python 
+from transformers import Mamba2Config, Mamba2ForCausalLM, AutoTokenizer
+import torch
+model_id = 'mistralai/Mamba-Codestral-7B-v0.1'
+tokenizer = AutoTokenizer.from_pretrained(model_id, revision='refs/pr/9', from_slow=True, legacy=False)
+model = Mamba2ForCausalLM.from_pretrained(model_id, revision='refs/pr/9')
+input_ids = tokenizer("Hey how are you doing?", return_tensors= "pt")["input_ids"]
+
+out = model.generate(input_ids, max_new_tokens=10)
+print(tokenizer.batch_decode(out))
+```
+
+Here's a draft script for finetuning: 
+```python 
+from trl import SFTTrainer
+from peft import LoraConfig
+from transformers import AutoTokenizer, Mamba2ForCausalLM, TrainingArguments
+model_id = 'mistralai/Mamba-Codestral-7B-v0.1'
+tokenizer = AutoTokenizer.from_pretrained(model_id, revision='refs/pr/9', from_slow=True, legacy=False)
+tokenizer.pad_token = tokenizer.eos_token
+tokenizer.padding_side = "left" #enforce padding side left
+
+model = Mamba2ForCausalLM.from_pretrained(model_id, revision='refs/pr/9')
+dataset = load_dataset("Abirate/english_quotes", split="train")
+# Without CUDA kernels, batch size of 2 occupies one 80GB device
+# but precision can be reduced.
+# Experiments and trials welcome!
+training_args = TrainingArguments(
+    output_dir="./results",
+    num_train_epochs=3,
+    per_device_train_batch_size=2,
+    logging_dir='./logs',
+    logging_steps=10,
+    learning_rate=2e-3
+)
+lora_config =  LoraConfig(
+        r=8,
+        target_modules=["embeddings", "in_proj", "out_proj"],
+        task_type="CAUSAL_LM",
+        bias="none"
+)
+trainer = SFTTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    args=training_args,
+    peft_config=lora_config,
+    train_dataset=dataset,
+    dataset_text_field="quote",
+)
+trainer.train()
+```
+
+
+## Mamba2Config
+
+[[autodoc]] Mamba2Config
+
+## Mamba2Model
+
+[[autodoc]] Mamba2Model
+    - forward
+
+## Mamba2LMHeadModel
+
+[[autodoc]] Mamba2ForCausalLM
+    - forward

docs/source/en/model_doc/marian.md

@@ -105,7 +105,7 @@ from huggingface_hub import list_models
 
 model_list = list_models()
 org = "Helsinki-NLP"
-model_ids = [x.modelId for x in model_list if x.modelId.startswith(org)]
+model_ids = [x.id for x in model_list if x.id.startswith(org)]
 suffix = [x.split("/")[1] for x in model_ids]
 old_style_multi_models = [f"{org}/{s}" for s in suffix if s != s.lower()]
 ```

docs/source/en/model_doc/nemotron.md

@@ -0,0 +1,148 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+-->
+
+# Nemotron
+
+## Nemotron
+
+### License
+
+The use of this model is governed by the [NVIDIA AI Foundation Models Community License Agreement](https://developer.nvidia.com/downloads/nv-ai-foundation-models-license).
+
+### Description
+
+Nemotron-4 is a family of enterprise ready generative text models compatible with [NVIDIA NeMo Framework](https://www.nvidia.com/en-us/ai-data-science/generative-ai/nemo-framework/).
+
+NVIDIA NeMo is an end-to-end, cloud-native platform to build, customize, and deploy generative AI models anywhere. It includes training and inferencing frameworks, guardrailing toolkits, data curation tools, and pretrained models, offering enterprises an easy, cost-effective, and fast way to adopt generative AI. To get access to NeMo Framework, please sign up at [this link](https://developer.nvidia.com/nemo-framework/join).
+
+### References
+
+[Announcement Blog](https://developer.nvidia.com/blog/nvidia-ai-foundation-models-build-custom-enterprise-chatbots-and-co-pilots-with-production-ready-llms/)
+
+### Model Architecture
+
+**Architecture Type:** Transformer
+
+**Network Architecture:** Transformer Decoder (auto-regressive language model).
+
+## Minitron
+
+### Minitron 4B Base
+
+Minitron is a family of small language models (SLMs) obtained by pruning NVIDIA's [Nemotron-4 15B](https://arxiv.org/abs/2402.16819) model. We prune model embedding size, attention heads, and MLP intermediate dimension, following which, we perform continued training with distillation to arrive at the final models.
+
+Deriving the Minitron 8B and 4B models from the base 15B model using our approach requires up to **40x fewer training tokens** per model compared to training from scratch; this results in **compute cost savings of 1.8x** for training the full model family (15B, 8B, and 4B). Minitron models exhibit up to a 16% improvement in MMLU scores compared to training from scratch, perform comparably to other community models such as Mistral 7B, Gemma 7B and Llama-3 8B, and outperform state-of-the-art compression techniques from the literature. Please refer to our [arXiv paper](https://arxiv.org/abs/2407.14679) for more details.
+
+Minitron models are for research and development only.
+
+### HuggingFace Quickstart
+
+The following code provides an example of how to load the Minitron-4B model and use it to perform text generation.
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+# Load the tokenizer and model
+model_path = 'nvidia/Minitron-4B-Base'
+tokenizer  = AutoTokenizer.from_pretrained(model_path)
+
+device = 'cuda'
+dtype  = torch.bfloat16
+model  = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=dtype, device_map=device)
+
+# Prepare the input text
+prompt = 'Complete the paragraph: our solar system is'
+inputs = tokenizer.encode(prompt, return_tensors='pt').to(model.device)
+
+# Generate the output
+outputs = model.generate(inputs, max_length=20)
+
+# Decode and print the output
+output_text = tokenizer.decode(outputs[0])
+print(output_text)
+```
+
+### License
+
+Minitron is released under the [NVIDIA Open Model License Agreement](https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf).
+
+### Evaluation Results
+
+*5-shot performance.* Language Understanding evaluated using [Massive Multitask Language Understanding](https://arxiv.org/abs/2009.03300):
+
+| Average |
+| :---- |
+| 58.6 |
+
+*Zero-shot performance.* Evaluated using select datasets from the [LM Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) with additions:
+
+| HellaSwag | Winogrande | GSM8K| ARC-C | XLSum |
+| :------------- | :------------- | :------------- | :------------- | :------------- |
+| 75.0 | 74.0 | 24.1  | 50.9 | 29.5
+
+
+*Code generation performance*. Evaluated using [HumanEval](https://github.com/openai/human-eval):
+
+| p@1, 0-Shot |
+| :------------- |
+| 23.3 |
+
+Please refer to our [paper](https://arxiv.org/abs/2407.14679) for the full set of results.
+
+### Citation
+
+If you find our work helpful, please consider citing our paper:
+```
+@article{minitron2024,
+      title={Compact Language Models via Pruning and Knowledge Distillation},
+      author={Saurav Muralidharan and Sharath Turuvekere Sreenivas and Raviraj Joshi and Marcin Chochowski and Mostofa Patwary and Mohammad Shoeybi and Bryan Catanzaro and Jan Kautz and Pavlo Molchanov},
+      journal={arXiv preprint arXiv:2407.14679},
+      year={2024},
+      url={https://arxiv.org/abs/2407.14679},
+}
+```
+
+## NemotronConfig
+
+[[autodoc]] NemotronConfig
+
+
+## NemotronModel
+
+[[autodoc]] NemotronModel
+    - forward
+
+
+## NemotronForCausalLM
+
+[[autodoc]] NemotronForCausalLM
+    - forward
+
+## NemotronForSequenceClassification
+
+[[autodoc]] NemotronForSequenceClassification
+    - forward
+
+
+## NemotronForQuestionAnswering
+
+[[autodoc]] NemotronForQuestionAnswering
+    - forward
+
+
+## NemotronForTokenClassification
+
+[[autodoc]] NemotronForTokenClassification
+    - forward

docs/source/en/model_doc/nllb.md

@@ -101,7 +101,7 @@ for the list of all BCP-47 in the Flores 200 dataset.
 >>> inputs = tokenizer(article, return_tensors="pt")
 
 >>> translated_tokens = model.generate(
-...     **inputs, forced_bos_token_id=tokenizer.lang_code_to_id["fra_Latn"], max_length=30
+...     **inputs, forced_bos_token_id=tokenizer.convert_tokens_to_ids("fra_Latn"), max_length=30
 ... )
 >>> tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
 Le chef de l'ONU dit qu'il n'y a pas de solution militaire en Syrie
@@ -126,7 +126,7 @@ See example below for a translation from romanian to german:
 >>> inputs = tokenizer(article, return_tensors="pt")
 
 >>> translated_tokens = model.generate(
-...     **inputs, forced_bos_token_id=tokenizer.lang_code_to_id["deu_Latn"], max_length=30
+...     **inputs, forced_bos_token_id=tokenizer.convert_tokens_to_ids("deu_Latn"), max_length=30
 ... )
 >>> tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
 UN-Chef sagt, es gibt keine militärische Lösung in Syrien
@@ -175,7 +175,7 @@ To load a model using Flash Attention 2, we can pass the argument `attn_implemen
 >>> inputs = tokenizer(article, return_tensors="pt").to("cuda")
 
 >>> translated_tokens = model.generate(
-...     **inputs, forced_bos_token_id=tokenizer.lang_code_to_id["deu_Latn"], max_length=30
+...     **inputs, forced_bos_token_id=tokenizer.convert_tokens_to_ids("deu_Latn"), max_length=30
 ... )
 >>> tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
 "UN-Chef sagt, es gibt keine militärische Lösung in Syrien"
@@ -187,4 +187,4 @@ Below is an expected speedup diagram that compares pure inference time between t
 
 <div style="text-align: center">
 <img src="https://huggingface.co/datasets/visheratin/documentation-images/resolve/main/nllb-speedup.webp">
-</div>
+</div>

docs/source/en/model_doc/prophetnet.md

@@ -51,7 +51,7 @@ The Authors' code can be found [here](https://github.com/microsoft/ProphetNet).
 
 - ProphetNet is a model with absolute position embeddings so it's usually advised to pad the inputs on the right rather than
   the left.
-- The model architecture is based on the original Transformer, but replaces the “standard” self-attention mechanism in the decoder by a a main self-attention mechanism and a self and n-stream (predict) self-attention mechanism.
+- The model architecture is based on the original Transformer, but replaces the “standard” self-attention mechanism in the decoder by a main self-attention mechanism and a self and n-stream (predict) self-attention mechanism.
 
 ## Resources
 

docs/source/en/model_doc/qwen2.md

@@ -18,7 +18,7 @@ rendered properly in your Markdown viewer.
 
 ## Overview
 
-Qwen2 is the new model series of large language models from the Qwen team. Previously, we released the Qwen series, including Qwen-72B, Qwen-1.8B, Qwen-VL, Qwen-Audio, etc.
+Qwen2 is the new model series of large language models from the Qwen team. Previously, we released the Qwen series, including Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B, Qwen2-57B-A14B, Qwen2-72B, Qwen2-Audio, etc.
 
 ### Model Details
 
@@ -27,16 +27,16 @@ Qwen2 is a language model series including decoder language models of different
 
 ## Usage tips
 
-`Qwen2-7B-beta` and `Qwen2-7B-Chat-beta` can be found on the [Huggingface Hub](https://huggingface.co/Qwen)
+`Qwen2-7B` and `Qwen2-7B-Instruct` can be found on the [Huggingface Hub](https://huggingface.co/Qwen)
 
-In the following, we demonstrate how to use `Qwen2-7B-Chat-beta` for the inference. Note that we have used the ChatML format for dialog, in this demo we show how to leverage `apply_chat_template` for this purpose.
+In the following, we demonstrate how to use `Qwen2-7B-Instruct` for the inference. Note that we have used the ChatML format for dialog, in this demo we show how to leverage `apply_chat_template` for this purpose.
 
 ```python
 >>> from transformers import AutoModelForCausalLM, AutoTokenizer
 >>> device = "cuda" # the device to load the model onto
 
->>> model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-7B-Chat", device_map="auto")
->>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen1.5-7B-Chat")
+>>> model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2-7B-Instruct", device_map="auto")
+>>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-7B-Instruct")
 
 >>> prompt = "Give me a short introduction to large language model."
 

docs/source/en/model_doc/qwen2_audio.md

@@ -0,0 +1,198 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Qwen2Audio
+
+## Overview
+
+The Qwen2-Audio is the new model series of large audio-language models from the Qwen team. Qwen2-Audio is capable of accepting various audio signal inputs and performing audio analysis or direct textual responses with regard to speech instructions. We introduce two distinct audio interaction modes:
+
+* voice chat: users can freely engage in voice interactions with Qwen2-Audio without text input
+* audio analysis: users could provide audio and text instructions for analysis during the interaction
+
+It was proposed in [Qwen2-Audio Technical Report](https://arxiv.org/abs/2407.10759) by Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, Chang Zhou, Jingren Zhou. 
+
+The abstract from the paper is the following:
+
+*We introduce the latest progress of Qwen-Audio, a large-scale audio-language model called Qwen2-Audio, which is capable of accepting various audio signal inputs and performing audio analysis or direct textual responses with regard to speech instructions. In contrast to complex hierarchical tags, we have simplified the pre-training process by utilizing natural language prompts for different data and tasks, and have further expanded the data volume. We have boosted the instruction-following capability of Qwen2-Audio and implemented two distinct audio interaction modes for voice chat and audio analysis. In the voice chat mode, users can freely engage in voice interactions with Qwen2-Audio without text input. In the audio analysis mode, users could provide audio and text instructions for analysis during the interaction. Note that we do not use any system prompts to switch between voice chat and audio analysis modes. Qwen2-Audio is capable of intelligently comprehending the content within audio and following voice commands to respond appropriately. For instance, in an audio segment that simultaneously contains sounds, multi-speaker conversations, and a voice command, Qwen2-Audio can directly understand the command and provide an interpretation and response to the audio. Additionally, DPO has optimized the model's performance in terms of factuality and adherence to desired behavior. According to the evaluation results from AIR-Bench, Qwen2-Audio outperformed previous SOTAs, such as Gemini-1.5-pro, in tests focused on audio-centric instruction-following capabilities. Qwen2-Audio is open-sourced with the aim of fostering the advancement of the multi-modal language community. *
+
+
+## Usage tips
+
+`Qwen2-Audio-7B` and `Qwen2-Audio-7B-Instruct` can be found on the [Huggingface Hub](https://huggingface.co/Qwen)
+
+In the following, we demonstrate how to use `Qwen2-Audio-7B-Instruct` for the inference, supporting both voice chat and audio analysis modes. Note that we have used the ChatML format for dialog, in this demo we show how to leverage `apply_chat_template` for this purpose.
+
+### Voice Chat Inference
+In the voice chat mode, users can freely engage in voice interactions with Qwen2-Audio without text input:
+```python
+from io import BytesIO
+from urllib.request import urlopen
+import librosa
+from transformers import Qwen2AudioForConditionalGeneration, AutoProcessor
+
+processor = AutoProcessor.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct")
+model = Qwen2AudioForConditionalGeneration.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct", device_map="auto")
+
+conversation = [
+    {"role": "user", "content": [
+        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/guess_age_gender.wav"},
+    ]},
+    {"role": "assistant", "content": "Yes, the speaker is female and in her twenties."},
+    {"role": "user", "content": [
+        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/translate_to_chinese.wav"},
+    ]},
+]
+text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
+audios = []
+for message in conversation:
+    if isinstance(message["content"], list):
+        for ele in message["content"]:
+            if ele["type"] == "audio":
+                audios.append(librosa.load(
+                    BytesIO(urlopen(ele['audio_url']).read()), 
+                    sr=processor.feature_extractor.sampling_rate)[0]
+                )
+
+inputs = processor(text=text, audios=audios, return_tensors="pt", padding=True)
+inputs.input_ids = inputs.input_ids.to("cuda")
+
+generate_ids = model.generate(**inputs, max_length=256)
+generate_ids = generate_ids[:, inputs.input_ids.size(1):]
+
+response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+```
+
+### Audio Analysis Inference
+In the audio analysis, users could provide both audio and text instructions for analysis:
+```python
+from io import BytesIO
+from urllib.request import urlopen
+import librosa
+from transformers import Qwen2AudioForConditionalGeneration, AutoProcessor
+
+processor = AutoProcessor.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct")
+model = Qwen2AudioForConditionalGeneration.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct", device_map="auto")
+
+conversation = [
+    {'role': 'system', 'content': 'You are a helpful assistant.'}, 
+    {"role": "user", "content": [
+        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/glass-breaking-151256.mp3"},
+        {"type": "text", "text": "What's that sound?"},
+    ]},
+    {"role": "assistant", "content": "It is the sound of glass shattering."},
+    {"role": "user", "content": [
+        {"type": "text", "text": "What can you do when you hear that?"},
+    ]},
+    {"role": "assistant", "content": "Stay alert and cautious, and check if anyone is hurt or if there is any damage to property."},
+    {"role": "user", "content": [
+        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/1272-128104-0000.flac"},
+        {"type": "text", "text": "What does the person say?"},
+    ]},
+]
+text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
+audios = []
+for message in conversation:
+    if isinstance(message["content"], list):
+        for ele in message["content"]:
+            if ele["type"] == "audio":
+                audios.append(
+                    librosa.load(
+                        BytesIO(urlopen(ele['audio_url']).read()), 
+                        sr=processor.feature_extractor.sampling_rate)[0]
+                )
+
+inputs = processor(text=text, audios=audios, return_tensors="pt", padding=True)
+inputs.input_ids = inputs.input_ids.to("cuda")
+
+generate_ids = model.generate(**inputs, max_length=256)
+generate_ids = generate_ids[:, inputs.input_ids.size(1):]
+
+response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+```
+
+### Batch Inference
+We also support batch inference:
+```python
+from io import BytesIO
+from urllib.request import urlopen
+import librosa
+from transformers import Qwen2AudioForConditionalGeneration, AutoProcessor
+
+processor = AutoProcessor.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct")
+model = Qwen2AudioForConditionalGeneration.from_pretrained("Qwen/Qwen2-Audio-7B-Instruct", device_map="auto")
+
+conversation1 = [
+    {"role": "user", "content": [
+        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/glass-breaking-151256.mp3"},
+        {"type": "text", "text": "What's that sound?"},
+    ]},
+    {"role": "assistant", "content": "It is the sound of glass shattering."},
+    {"role": "user", "content": [
+        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/f2641_0_throatclearing.wav"},
+        {"type": "text", "text": "What can you hear?"},
+    ]}
+]
+
+conversation2 = [
+    {"role": "user", "content": [
+        {"type": "audio", "audio_url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/1272-128104-0000.flac"},
+        {"type": "text", "text": "What does the person say?"},
+    ]},
+]
+
+conversations = [conversation1, conversation2]
+
+text = [processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False) for conversation in conversations]
+
+audios = []
+for conversation in conversations:
+    for message in conversation:
+        if isinstance(message["content"], list):
+            for ele in message["content"]:
+                if ele["type"] == "audio":
+                    audios.append(
+                        librosa.load(
+                            BytesIO(urlopen(ele['audio_url']).read()), 
+                            sr=processor.feature_extractor.sampling_rate)[0]
+                    )
+
+inputs = processor(text=text, audios=audios, return_tensors="pt", padding=True)
+inputs['input_ids'] = inputs['input_ids'].to("cuda")
+inputs.input_ids = inputs.input_ids.to("cuda")
+
+generate_ids = model.generate(**inputs, max_length=256)
+generate_ids = generate_ids[:, inputs.input_ids.size(1):]
+
+response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+```
+
+## Qwen2AudioConfig
+
+[[autodoc]] Qwen2AudioConfig
+
+## Qwen2AudioConfig
+
+[[autodoc]] Qwen2AudioEncoderConfig
+
+## Qwen2AudioProcessor
+
+[[autodoc]] Qwen2AudioProcessor
+
+## Qwen2AudioForConditionalGeneration
+
+[[autodoc]] Qwen2AudioForConditionalGeneration
+    - forward

docs/source/en/model_doc/qwen2_vl.md

@@ -0,0 +1,329 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Qwen2_VL
+
+
+## Overview
+
+The [Qwen2_VL](https://qwenlm.github.io/blog/qwen2-vl/) is a major update to our [Qwen-VL](https://arxiv.org/pdf/2308.12966) model from the Qwen team. 
+
+The abstract from the blog is the following:
+
+*This blog introduces Qwen2-VL, an advanced version of the Qwen-VL model that has undergone significant enhancements over the past year. Key improvements include enhanced image comprehension, advanced video understanding, integrated visual agent functionality, and expanded multilingual support. The model architecture has been optimized for handling arbitrary image resolutions through Naive Dynamic Resolution support and utilizes Multimodal Rotary Position Embedding (M-ROPE) to effectively process both 1D textual and multi-dimensional visual data. This updated model demonstrates competitive performance against leading AI systems like GPT-4o and Claude 3.5 Sonnet in vision-related tasks and ranks highly among open-source models in text capabilities. These advancements make Qwen2-VL a versatile tool for various applications requiring robust multimodal processing and reasoning abilities.*
+
+
+## Usage example
+
+### Single Media inference
+
+The model can accept both images and videos as input. Here's an example code for inference.
+
+```python
+
+from PIL import Image
+import requests
+import torch
+from torchvision import io
+from typing import Dict
+from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
+
+# Load the model in half-precision on the available device(s)
+model = Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2-VL-7B-Instruct", device_map="auto")
+processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
+
+# Image
+url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
+image = Image.open(requests.get(url, stream=True).raw)
+
+conversation = [
+    {
+        "role":"user",
+        "content":[
+            {
+                "type":"image",
+            },
+            {
+                "type":"text",
+                "text":"Describe this image."
+            }
+        ]
+    }
+]
+
+
+# Preprocess the inputs
+text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
+# Excepted output: '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Describe this image.<|im_end|>\n<|im_start|>assistant\n'
+
+inputs = processor(text=[text_prompt], images=[image], padding=True, return_tensors="pt")
+inputs = inputs.to('cuda')
+
+# Inference: Generation of the output
+output_ids = model.generate(**inputs, max_new_tokens=128)
+generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, output_ids)]
+output_text = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
+print(output_text)
+
+
+
+# Video
+def fetch_video(ele: Dict, nframe_factor=2):
+    if isinstance(ele['video'], str):
+        def round_by_factor(number: int, factor: int) -> int:
+            return round(number / factor) * factor
+
+        video = ele["video"]
+        if video.startswith("file://"):
+            video = video[7:]
+
+        video, _, info = io.read_video(
+            video,
+            start_pts=ele.get("video_start", 0.0),
+            end_pts=ele.get("video_end", None),
+            pts_unit="sec",
+            output_format="TCHW",
+        )
+        assert not ("fps" in ele and "nframes" in ele), "Only accept either `fps` or `nframes`"
+        if "nframes" in ele:
+            nframes = round_by_factor(ele["nframes"], nframe_factor)
+        else:
+            fps = ele.get("fps", 1.0)
+            nframes = round_by_factor(video.size(0) / info["video_fps"] * fps, nframe_factor)
+        idx = torch.linspace(0, video.size(0) - 1, nframes, dtype=torch.int64)
+        return video[idx]
+
+video_info = {"type": "video", "video": "/path/to/video.mp4", "fps": 1.0}
+video = fetch_video(video_info)
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "video"},
+            {"type": "text", "text": "What happened in the video?"},
+        ],
+    }
+]
+
+# Preprocess the inputs
+text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
+# Excepted output: '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|><|video_pad|><|vision_end|>What happened in the video?<|im_end|>\n<|im_start|>assistant\n'
+
+inputs = processor(text=[text_prompt], videos=[video], padding=True, return_tensors="pt")
+inputs = inputs.to('cuda')
+
+# Inference: Generation of the output
+output_ids = model.generate(**inputs, max_new_tokens=128)
+generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, output_ids)]
+output_text = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
+print(output_text)
+
+```
+
+
+### Batch Mixed Media Inference
+
+The model can batch inputs composed of mixed samples of various types such as images, videos, and text. Here is an example.
+
+```python
+
+image1 = Image.open("/path/to/image1.jpg")
+image2 = Image.open("/path/to/image2.jpg")
+image3 = Image.open("/path/to/image3.jpg")
+image4 = Image.open("/path/to/image4.jpg")
+image5 = Image.open("/path/to/image5.jpg")
+video = fetch_video({
+    "type": "video",
+    "video": "/path/to/video.mp4",
+    "fps": 1.0
+})
+
+# Conversation for the first image
+conversation1 = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": "Describe this image."}
+        ]
+    }
+]
+
+# Conversation with two images
+conversation2 = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "image"},
+            {"type": "text", "text": "What is written in the pictures?"}
+        ]
+    }
+]
+
+# Conversation with pure text
+conversation3 = [
+    {
+        "role": "user",
+        "content": "who are you?"
+    }
+]
+
+
+# Conversation with mixed midia
+conversation4 = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "image"},
+            {"type": "video"},
+            {"type": "text", "text": "What are the common elements in these medias?"},
+        ],
+    }
+]
+
+conversations = [conversation1, conversation2, conversation3, conversation4]
+# Preparation for batch inference
+texts = [processor.apply_chat_template(msg, add_generation_prompt=True) for msg in conversations]
+inputs = processor(
+    text=texts,
+    images=[image1, image2, image3, image4, image5],
+    videos=[video],
+    padding=True,
+    return_tensors="pt",
+)
+inputs = inputs.to('cuda')
+
+# Batch Inference
+output_ids = model.generate(**inputs, max_new_tokens=128)
+generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, output_ids)]
+output_text = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
+print(output_text)
+```
+
+### Usage Tips
+
+#### Image Resolution for performance boost
+
+The model supports a wide range of resolution inputs. By default, it uses the native resolution for input, but higher resolutions can enhance performance at the cost of more computation. Users can set the minimum and maximum number of pixels to achieve an optimal configuration for their needs.
+
+```python
+
+min_pixels = 224*224
+max_pixels = 2048*2048
+processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels)
+
+```
+
+
+
+#### Multiple Image Inputs
+
+By default, images and video content are directly included in the conversation. When handling multiple images, it's helpful to add labels to the images and videos for better reference. Users can control this behavior with the following settings:
+
+
+
+```python
+
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"}, 
+            {"type": "text", "text": "Hello, how are you?"}
+        ]
+    },
+    {
+        "role": "assistant",
+        "content": "I'm doing well, thank you for asking. How can I assist you today?"
+    },
+    {
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "Can you describe these images and video?"}, 
+            {"type": "image"}, 
+            {"type": "image"}, 
+            {"type": "video"}, 
+            {"type": "text", "text": "These are from my vacation."}
+        ]
+    },
+    {
+        "role": "assistant",
+        "content": "I'd be happy to describe the images and video for you. Could you please provide more context about your vacation?"
+    },
+    {
+        "role": "user",
+        "content": "It was a trip to the mountains. Can you see the details in the images and video?"
+    }
+]
+
+# default:
+prompt_without_id = processor.apply_chat_template(conversation, add_generation_prompt=True)
+# Excepted output: '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Hello, how are you?<|im_end|>\n<|im_start|>assistant\nI'm doing well, thank you for asking. How can I assist you today?<|im_end|>\n<|im_start|>user\nCan you describe these images and video?<|vision_start|><|image_pad|><|vision_end|><|vision_start|><|image_pad|><|vision_end|><|vision_start|><|video_pad|><|vision_end|>These are from my vacation.<|im_end|>\n<|im_start|>assistant\nI'd be happy to describe the images and video for you. Could you please provide more context about your vacation?<|im_end|>\n<|im_start|>user\nIt was a trip to the mountains. Can you see the details in the images and video?<|im_end|>\n<|im_start|>assistant\n'
+
+
+# add ids
+prompt_with_id = processor.apply_chat_template(conversation, add_generation_prompt=True, add_vision_id=True)
+# Excepted output: '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nPicture 1: <|vision_start|><|image_pad|><|vision_end|>Hello, how are you?<|im_end|>\n<|im_start|>assistant\nI'm doing well, thank you for asking. How can I assist you today?<|im_end|>\n<|im_start|>user\nCan you describe these images and video?Picture 2: <|vision_start|><|image_pad|><|vision_end|>Picture 3: <|vision_start|><|image_pad|><|vision_end|>Video 1: <|vision_start|><|video_pad|><|vision_end|>These are from my vacation.<|im_end|>\n<|im_start|>assistant\nI'd be happy to describe the images and video for you. Could you please provide more context about your vacation?<|im_end|>\n<|im_start|>user\nIt was a trip to the mountains. Can you see the details in the images and video?<|im_end|>\n<|im_start|>assistant\n'
+
+```
+
+#### Flash-Attention 2 to speed up generation
+
+First, make sure to install the latest version of Flash Attention 2:
+
+```bash
+pip install -U flash-attn --no-build-isolation
+```
+
+Also, you should have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of the [flash attention repository](https://github.com/Dao-AILab/flash-attention). FlashAttention-2 can only be used when a model is loaded in `torch.float16` or `torch.bfloat16`.
+
+To load and run a model using Flash Attention-2, simply add `attn_implementation="flash_attention_2"` when loading the model as follows:
+
+```python
+from transformers import Qwen2VLForConditionalGeneration
+
+model = Qwen2VLForConditionalGeneration.from_pretrained(
+    "Qwen/Qwen2-VL-7B-Instruct", 
+    torch_dtype=torch.bfloat16, 
+    attn_implementation="flash_attention_2",
+)
+```
+
+
+## Qwen2VLConfig
+
+[[autodoc]] Qwen2VLConfig
+
+## Qwen2VLImageProcessor
+
+[[autodoc]] Qwen2VLImageProcessor
+    - preprocess
+
+## Qwen2VLProcessor
+
+[[autodoc]] Qwen2VLProcessor
+
+## Qwen2VLModel
+
+[[autodoc]] Qwen2VLModel
+    - forward
+
+## Qwen2VLForConditionalGeneration
+
+[[autodoc]] Qwen2VLForConditionalGeneration
+    - forward

docs/source/en/model_doc/roberta.md

@@ -51,19 +51,19 @@ This model was contributed by [julien-c](https://huggingface.co/julien-c). The o
 
 ## Usage tips
 
-- This implementation is the same as [`BertModel`] with a tiny embeddings tweak as well as a setup
-  for Roberta pretrained models.
-- RoBERTa has the same architecture as BERT, but uses a byte-level BPE as a tokenizer (same as GPT-2) and uses a
+- This implementation is the same as [`BertModel`] with a minor tweak to the embeddings, as well as a setup
+  for RoBERTa pretrained models.
+- RoBERTa has the same architecture as BERT but uses a byte-level BPE as a tokenizer (same as GPT-2) and uses a
   different pretraining scheme.
-- RoBERTa doesn't have `token_type_ids`, you don't need to indicate which token belongs to which segment. Just
-  separate your segments with the separation token `tokenizer.sep_token` (or `</s>`)
-- Same as BERT with better pretraining tricks:
+- RoBERTa doesn't have `token_type_ids`, so you don't need to indicate which token belongs to which segment. Just
+  separate your segments with the separation token `tokenizer.sep_token` (or `</s>`).
+- RoBERTa is similar to BERT but with better pretraining techniques:
 
-    * dynamic masking: tokens are masked differently at each epoch, whereas BERT does it once and for all
-    * together to reach 512 tokens (so the sentences are in an order than may span several documents)
-    * train with larger batches
-    * use BPE with bytes as a subunit and not characters (because of unicode characters)
-- [CamemBERT](camembert) is a wrapper around RoBERTa. Refer to this page for usage examples.
+    * Dynamic masking: tokens are masked differently at each epoch, whereas BERT does it once and for all.
+    * Sentence packing: Sentences are packed together to reach 512 tokens (so the sentences are in an order that may span several documents).
+    * Larger batches: Training uses larger batches.
+    * Byte-level BPE vocabulary: Uses BPE with bytes as a subunit instead of characters, accommodating Unicode characters.
+- [CamemBERT](camembert) is a wrapper around RoBERTa. Refer to its model page for usage examples.
 
 ## Resources
 

docs/source/en/model_doc/rt_detr.md

@@ -27,6 +27,11 @@ The abstract from the paper is the following:
 
 *Recently, end-to-end transformer-based detectors (DETRs) have achieved remarkable performance. However, the issue of the high computational cost of DETRs has not been effectively addressed, limiting their practical application and preventing them from fully exploiting the benefits of no post-processing, such as non-maximum suppression (NMS). In this paper, we first analyze the influence of NMS in modern real-time object detectors on inference speed, and establish an end-to-end speed benchmark. To avoid the inference delay caused by NMS, we propose a Real-Time DEtection TRansformer (RT-DETR), the first real-time end-to-end object detector to our best knowledge. Specifically, we design an efficient hybrid encoder to efficiently process multi-scale features by decoupling the intra-scale interaction and cross-scale fusion, and propose IoU-aware query selection to improve the initialization of object queries. In addition, our proposed detector supports flexibly adjustment of the inference speed by using different decoder layers without the need for retraining, which facilitates the practical application of real-time object detectors. Our RT-DETR-L achieves 53.0% AP on COCO val2017 and 114 FPS on T4 GPU, while RT-DETR-X achieves 54.8% AP and 74 FPS, outperforming all YOLO detectors of the same scale in both speed and accuracy. Furthermore, our RT-DETR-R50 achieves 53.1% AP and 108 FPS, outperforming DINO-Deformable-DETR-R50 by 2.2% AP in accuracy and by about 21 times in FPS.*
 
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/rt_detr_overview.png"
+alt="drawing" width="600"/>
+
+<small> RT-DETR performance relative to YOLO models. Taken from the <a href="https://arxiv.org/abs/2304.08069">original paper.</a> </small>
+
 The model version was contributed by [rafaelpadilla](https://huggingface.co/rafaelpadilla) and [sangbumchoi](https://github.com/SangbumChoi). The original code can be found [here](https://github.com/lyuwenyu/RT-DETR/).
 
 
@@ -66,6 +71,16 @@ remote: 0.95 [40.11, 73.44, 175.96, 118.48]
 remote: 0.92 [333.73, 76.58, 369.97, 186.99]
 ```
 
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with RT-DETR.
+
+<PipelineTag pipeline="object-detection"/>
+
+- Scripts for finetuning [`RTDetrForObjectDetection`] with [`Trainer`] or [Accelerate](https://huggingface.co/docs/accelerate/index) can be found [here](https://github.com/huggingface/transformers/tree/main/examples/pytorch/object-detection).
+- See also: [Object detection task guide](../tasks/object_detection).
+- Notebooks regarding inference and fine-tuning RT-DETR on a custom dataset can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/RT-DETR). 🌎
+
 ## RTDetrConfig
 
 [[autodoc]] RTDetrConfig

docs/source/en/model_doc/sam.md

@@ -34,7 +34,7 @@ Tips:
 - The model predicts much better results if input 2D points and/or input bounding boxes are provided
 - You can prompt multiple points for the same image, and predict a single mask. 
 - Fine-tuning the model is not supported yet
-- According to the paper, textual input should be also supported. However, at this time of writing this seems to be not supported according to [the official repository](https://github.com/facebookresearch/segment-anything/issues/4#issuecomment-1497626844). 
+- According to the paper, textual input should be also supported. However, at this time of writing this seems not to be supported according to [the official repository](https://github.com/facebookresearch/segment-anything/issues/4#issuecomment-1497626844). 
 
 
 This model was contributed by [ybelkada](https://huggingface.co/ybelkada) and [ArthurZ](https://huggingface.co/ArthurZ).

docs/source/en/model_doc/siglip.md

@@ -27,7 +27,7 @@ The abstract from the paper is the following:
 ## Usage tips
 
 - Usage of SigLIP is similar to [CLIP](clip). The main difference is the training loss, which does not require a global view of all the pairwise similarities of images and texts within a batch. One needs to apply the sigmoid activation function to the logits, rather than the softmax.
-- Training is not yet supported. If you want to fine-tune SigLIP or train from scratch, refer to the loss function from [OpenCLIP](https://github.com/mlfoundations/open_clip/blob/73ad04ae7fb93ede1c02dc9040a828634cb1edf1/src/open_clip/loss.py#L307), which leverages various `torch.distributed` utilities.
+- Training is supported but does not use `torch.distributed` utilities which may limit the scalability of batch size. However, DDP and FDSP works on single-node multi-gpu setup.
 - When using the standalone [`SiglipTokenizer`] or [`SiglipProcessor`], make sure to pass `padding="max_length"` as that's how the model was trained.
 - To get the same results as the pipeline, a prompt template of "This is a photo of {label}." should be used.
 
@@ -107,6 +107,88 @@ A list of official Hugging Face and community (indicated by 🌎) resources to h
 
 If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
 
+
+## Combining SigLIP and Flash Attention 2
+
+First, make sure to install the latest version of Flash Attention 2.
+
+```bash
+pip install -U flash-attn --no-build-isolation
+```
+
+Make also sure that you have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of flash-attn repository. Make also sure to load your model in half-precision (e.g. `torch.float16``)
+
+To load and run a model using Flash Attention 2, refer to the snippet below:
+
+```python
+>>> import torch
+>>> import requests
+>>> from PIL import Image
+>>> from transformers import SiglipProcessor, SiglipModel
+>>> device = "cuda" # the device to load the model onto
+
+>>> model = SiglipModel.from_pretrained(
+...     "google/siglip-so400m-patch14-384",
+...     attn_implementation="flash_attention_2",
+...     torch_dtype=torch.float16,
+...     device_map=device,
+... )
+>>> processor = SiglipProcessor.from_pretrained("google/siglip-so400m-patch14-384")
+
+>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+>>> image = Image.open(requests.get(url, stream=True).raw)
+
+>>> candidate_labels = ["2 cats", "2 dogs"]
+# follows the pipeline prompt template to get same results
+>>> candidate_labels = [f'This is a photo of {label}.' for label in candidate_labels]
+# important: we pass `padding=max_length` since the model was trained with this
+>>> inputs = processor(text=candidate_labels, images=image, padding="max_length", return_tensors="pt")
+>>> inputs.to(device)
+
+>>> with torch.no_grad():
+...     with torch.autocast(device):
+...         outputs = model(**inputs)
+
+>>> logits_per_image = outputs.logits_per_image
+>>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
+>>> print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
+51.3% that image 0 is 'This is a photo of 2 cats.'
+```
+
+
+## Using Scaled Dot Product Attention (SDPA)
+
+PyTorch includes a native scaled dot-product attention (SDPA) operator as part of `torch.nn.functional`. This function 
+encompasses several implementations that can be applied depending on the inputs and the hardware in use. See the 
+[official documentation](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html) 
+or the [GPU Inference](https://huggingface.co/docs/transformers/main/en/perf_infer_gpu_one#pytorch-scaled-dot-product-attention)
+page for more information.
+
+You may set `attn_implementation="sdpa"` in `from_pretrained()` to explicitly request SDPA to be used. Make sure you have `torch>=2.1.1`.
+
+```python
+>>> from transformers import SiglipModel
+
+>>> model = SiglipModel.from_pretrained(
+...     "google/siglip-so400m-patch14-384",
+...     attn_implementation="sdpa",
+...     torch_dtype=torch.float16,
+...     device_map=device,
+... )
+```
+
+For the best speedups, we recommend loading the model in half-precision (e.g. `torch.float16` or `torch.bfloat16`).
+
+
+## Expected speedups
+
+Below is an expected speedup diagram that compares inference time between the native implementation in transformers using `google/siglip-so400m-patch14-384` checkpoint in `float16` precision and the Flash Attention 2 / SDPA version of the model using different batch sizes.
+
+<div style="text-align: center">
+<img src="https://i.imgur.com/cWm4rsn.png">
+</div>
+
+
 ## SiglipConfig
 
 [[autodoc]] SiglipConfig

docs/source/en/model_doc/video_llava.md

@@ -98,7 +98,7 @@ indices = np.arange(0, total_frames, total_frames / 8).astype(int)
 video = read_video_pyav(container, indices)
 
 # For better results, we recommend to prompt the model in the following format
-prompt = "USER: <video>Why is this funny? ASSISTANT:"
+prompt = "USER: <video>\nWhy is this funny? ASSISTANT:"
 inputs = processor(text=prompt, videos=video, return_tensors="pt")
 
 out = model.generate(**inputs, max_new_tokens=60)
@@ -108,7 +108,7 @@ processor.batch_decode(out, skip_special_tokens=True, clean_up_tokenization_spac
 For multiple turns conversation change the prompt format to:
 
 ```bash
-"USER: <video>What do you see in this video? ASSISTANT: A baby reading a book. USER: Why is the it funny? ASSISTANT:"
+"USER: <video>\nWhat do you see in this video? ASSISTANT: A baby reading a book. USER: Why is the it funny? ASSISTANT:"
 ```
 
 ### Mixed Media Mode
@@ -123,7 +123,7 @@ import requests
 # Load and image and write a new prompt
 url = "http://images.cocodataset.org/val2017/000000039769.jpg"
 image = Image.open(requests.get(url, stream=True).raw)
-prompt = "USER: <image> How many cats are there in the image? ASSISTANT: There are two cats. USER: <video>Why is this video funny? ASSISTANT:"
+prompt = "USER: <image>\nHow many cats are there in the image? ASSISTANT: There are two cats. USER: <video>\nWhy is this video funny? ASSISTANT:"
 
 inputs = processor(text=prompt, images=image, videos=clip, padding=True, return_tensors="pt")
 

docs/source/en/model_doc/vipllava.md

@@ -26,7 +26,12 @@ The abstract from the paper is the following:
 
 *While existing large vision-language multimodal models focus on whole image understanding, there is a prominent gap in achieving region-specific comprehension. Current approaches that use textual coordinates or spatial encodings often fail to provide a user-friendly interface for visual prompting. To address this challenge, we introduce a novel multimodal model capable of decoding arbitrary visual prompts. This allows users to intuitively mark images and interact with the model using natural cues like a "red bounding box" or "pointed arrow". Our simple design directly overlays visual markers onto the RGB image, eliminating the need for complex region encodings, yet achieves state-of-the-art performance on region-understanding tasks like Visual7W, PointQA, and Visual Commonsense Reasoning benchmark. Furthermore, we present ViP-Bench, a comprehensive benchmark to assess the capability of models in understanding visual prompts across multiple dimensions, enabling future research in this domain. Code, data, and model are publicly available.*
 
-Tips:
+The original code can be found [here](https://github.com/mu-cai/ViP-LLaVA).
+
+This model was contributed by [Younes Belkada](https://huggingface.co/ybelkada)
+
+
+## Usage tips:
 
 - The architecture is similar than llava architecture except that the multi-modal projector takes a set of concatenated vision hidden states and has an additional layernorm layer on that module.
 
@@ -34,22 +39,51 @@ Tips:
 
 - Note the model has not been explicitly trained to process multiple images in the same prompt, although this is technically possible, you may experience inaccurate results.
 
-- For better results, we recommend users to prompt the model with the correct prompt format: 
+- For better results, we recommend users to use the processor's `apply_chat_template()` method to format your prompt correctly. For that you need to construct a conversation history, passing in a plain string will not format your prompt. Each message in the conversation history for chat templates is a dictionary with keys "role" and "content". The "content" should be a list of dictionaries, for "text" and "image" modalities, as follows:
 
+```python
+from transformers import AutoProcessor
+
+processor = AutoProcessor.from_pretrained("llava-hf/vip-llava-7b-hf")
+
+conversation = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": "What’s shown in this image?"},
+         ,
+    },
+    {
+        "role": "assistant",
+        "content": [{"type": "text", "text": "This image shows a red stop sign."},]
+    },
+    {
+
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "Describe the image in more details."},
+        ],
+    },
+]
+
+text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
+
+# Note that the template simply formats your prompt, you still have to tokenize it and obtain pixel values for your images
+print(text_prompt)
+>>> "###Human: <image>\nWhat’s shown in this image?###Assistant: This image shows a red stop sign.###Human: Describe the image in more details.###Assistant:"
+```
+
+- If you want to construct a chat prompt yourself, below is a list of prompt formats accepted by VipLLaVa checkpoints:
 ```bash
 A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.###Human: <image>\n<prompt>###Assistant:
 ```
 
 For multiple turns conversation:
-
 ```bash
 A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.###Human: <image>\n<prompt1>###Assistant: <answer1>###Human: <prompt2>###Assistant:
 ```
 
-The original code can be found [here](https://github.com/mu-cai/ViP-LLaVA).
-
-This model was contributed by [Younes Belkada](https://huggingface.co/ybelkada)
-
 
 ## VipLlavaConfig
 

docs/source/en/model_doc/vits.md

@@ -93,12 +93,33 @@ from transformers import VitsTokenizer
 tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")
 print(tokenizer.is_uroman)
 ```
+If the is_uroman attribute is `True`, the tokenizer will automatically apply the `uroman` package to your text inputs, but you need to install uroman if not already installed using:  
+```
+pip install --upgrade uroman
+```
+Note: Python version required to use `uroman` as python package should be >= `3.10`. 
+You can use the tokenizer as usual without any additional preprocessing steps:
+```python
+import torch
+from transformers import VitsTokenizer, VitsModel, set_seed
+import os
+import subprocess
 
-If required, you should apply the uroman package to your text inputs **prior** to passing them to the `VitsTokenizer`, 
-since currently the tokenizer does not support performing the pre-processing itself.  
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-kor")
+model = VitsModel.from_pretrained("facebook/mms-tts-kor")
+text = "이봐 무슨 일이야"
+inputs = tokenizer(text=text, return_tensors="pt")
 
+set_seed(555)  # make deterministic
+with torch.no_grad():
+   outputs = model(inputs["input_ids"])
+
+waveform = outputs.waveform[0]
+```
+If you don't want to upgrade to python >= `3.10`, then you can use the `uroman` perl package to pre-process the text inputs to the Roman alphabet.
 To do this, first clone the uroman repository to your local machine and set the bash variable `UROMAN` to the local path:
 
+
 ```bash
 git clone https://github.com/isi-nlp/uroman.git
 cd uroman

docs/source/en/model_doc/whisper.md

@@ -27,6 +27,27 @@ The abstract from the paper is the following:
 This model was contributed by [Arthur Zucker](https://huggingface.co/ArthurZ). The Tensorflow version of this model was contributed by [amyeroberts](https://huggingface.co/amyeroberts).
 The original code can be found [here](https://github.com/openai/whisper).
 
+## Quick usage
+
+You can run Whisper in less than 4 lines of code and transcribe in less than a minute!
+
+```python
+# pip install transformers torch
+
+import torch
+from transformers import pipeline
+
+whisper = pipeline("automatic-speech-recognition", "openai/whisper-large-v3", torch_dtype=torch.float16, device="cuda:0")
+
+transcription = whisper("<audio_file.mp3>")
+
+print(transcription["text"])
+```
+
+Voila! You can swap the model with any [Whisper checkpoints](https://huggingface.co/models?other=whisper&sort=downloads) on the Hugging Face Hub with the same pipeline based on your needs.
+
+Bonus: You can replace `"cuda"` with `"mps"` to make it seamlessly work on Macs.
+
 ## Usage tips
 
 - The model usually performs well without requiring any finetuning.
@@ -52,8 +73,6 @@ Here is a step-by-step guide to transcribing an audio sample using a pre-trained
 >>> # Select an audio file and read it:
 >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
 >>> audio_sample = ds[0]["audio"]
->>> waveform = audio_sample["array"]
->>> sampling_rate = audio_sample["sampling_rate"]
 
 >>> # Load the Whisper model in Hugging Face format:
 >>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
@@ -61,7 +80,7 @@ Here is a step-by-step guide to transcribing an audio sample using a pre-trained
 
 >>> # Use the model and processor to transcribe the audio:
 >>> input_features = processor(
-...     waveform, sampling_rate=sampling_rate, return_tensors="pt"
+...     audio_sample["array"], sampling_rate=audio_sample["sampling_rate"], return_tensors="pt"
 ... ).input_features
 
 >>> # Generate token ids
@@ -74,6 +93,50 @@ Here is a step-by-step guide to transcribing an audio sample using a pre-trained
 ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
 ```
 
+Whisper is compatible with the following optimisations for both short and long-form generation:
+- [PyTorch Scaled Dot Product Attention (SDPA)](../perf_infer_gpu_one#pytorch-scaled-dot-product-attention): flash attention and memory-efficient attention kernels. Enabled by default for `torch>=2.1.1`.
+- [Flash Attention 2](../perf_infer_gpu_one#flashattention-2): improved implementation of flash attention through better parallelism and work partitioning. 
+- [torch.compile](../llm_optims#static-kv-cache-and-torchcompile): JIT-compile the forward pass to dispatch to efficient fused kernels.
+
+As an example, the following codesnippet enables SDPA and `torch.compile` for up to 5x faster inference:
+
+```python
+>>> from datasets import load_dataset
+>>> from transformers import WhisperProcessor, WhisperForConditionalGeneration
+
+>>> # Select an audio file and read it:
+>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+>>> audio_sample = ds[0]["audio"]
+
+>>> # Load the Whisper model with SDPA attention
+>>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+>>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en", attn_implementation="sdpa")
+
+>>> # Enable static cache and compile the forward pass
+>>> model.generation_config.cache_implementation = "static"
+>>> model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
+
+>>> # Use the model and processor to transcribe the audio:
+>>> input_features = processor(
+...     audio_sample["array"], sampling_rate=audio_sample["sampling_rate"], return_tensors="pt"
+... ).input_features
+
+>>> # Compile the forward pass
+>>> for _ in range(2):
+>>>     model.generate(input_features)
+
+>>> # Generate token ids using compiled graph (fast!)
+>>> predicted_ids = model.generate(input_features)
+
+>>> # Decode token ids to text
+>>> transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
+
+>>> transcription[0]
+' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
+```
+
+For more details on each optimisation, refer to the documentation linked above.
+
 ## Resources
 
 A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with Whisper. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.

docs/source/en/model_doc/zoedepth.md

@@ -0,0 +1,108 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# ZoeDepth
+
+## Overview
+
+The ZoeDepth model was proposed in [ZoeDepth: Zero-shot Transfer by Combining Relative and Metric Depth](https://arxiv.org/abs/2302.12288) by Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, Matthias Müller. ZoeDepth extends the [DPT](dpt) framework for metric (also called absolute) depth estimation. ZoeDepth is pre-trained on 12 datasets using relative depth and fine-tuned on two domains (NYU and KITTI) using metric depth. A lightweight head is used with a novel bin adjustment design called metric bins module for each domain. During inference, each input image is automatically routed to the appropriate head using a latent classifier.
+
+The abstract from the paper is the following:
+
+*This paper tackles the problem of depth estimation from a single image. Existing work either focuses on generalization performance disregarding metric scale, i.e. relative depth estimation, or state-of-the-art results on specific datasets, i.e. metric depth estimation. We propose the first approach that combines both worlds, leading to a model with excellent generalization performance while maintaining metric scale. Our flagship model, ZoeD-M12-NK, is pre-trained on 12 datasets using relative depth and fine-tuned on two datasets using metric depth. We use a lightweight head with a novel bin adjustment design called metric bins module for each domain. During inference, each input image is automatically routed to the appropriate head using a latent classifier. Our framework admits multiple configurations depending on the datasets used for relative depth pre-training and metric fine-tuning. Without pre-training, we can already significantly improve the state of the art (SOTA) on the NYU Depth v2 indoor dataset. Pre-training on twelve datasets and fine-tuning on the NYU Depth v2 indoor dataset, we can further improve SOTA for a total of 21% in terms of relative absolute error (REL). Finally, ZoeD-M12-NK is the first model that can jointly train on multiple datasets (NYU Depth v2 and KITTI) without a significant drop in performance and achieve unprecedented zero-shot generalization performance to eight unseen datasets from both indoor and outdoor domains.*
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/zoedepth_architecture_bis.png"
+alt="drawing" width="600"/>
+
+<small> ZoeDepth architecture. Taken from the <a href="https://arxiv.org/abs/2302.12288">original paper.</a> </small>
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr).
+The original code can be found [here](https://github.com/isl-org/ZoeDepth).
+
+## Usage tips
+
+- ZoeDepth is an absolute (also called metric) depth estimation model, unlike DPT which is a relative depth estimation model. This means that ZoeDepth is able to estimate depth in metric units like meters.
+
+The easiest to perform inference with ZoeDepth is by leveraging the [pipeline API](../main_classes/pipelines.md):
+
+```python
+from transformers import pipeline
+from PIL import Image
+import requests
+
+url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+image = Image.open(requests.get(url, stream=True).raw)
+
+pipe = pipeline(task="depth-estimation", model="Intel/zoedepth-nyu-kitti")
+result = pipe(image)
+depth = result["depth"]
+```
+
+Alternatively, one can also perform inference using the classes:
+
+```python
+from transformers import AutoImageProcessor, ZoeDepthForDepthEstimation
+import torch
+import numpy as np
+from PIL import Image
+import requests
+
+url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+image = Image.open(requests.get(url, stream=True).raw)
+
+image_processor = AutoImageProcessor.from_pretrained("Intel/zoedepth-nyu-kitti")
+model = ZoeDepthForDepthEstimation.from_pretrained("Intel/zoedepth-nyu-kitti")
+
+# prepare image for the model
+inputs = image_processor(images=image, return_tensors="pt")
+
+with torch.no_grad():
+    outputs = model(**inputs)
+    predicted_depth = outputs.predicted_depth
+
+# interpolate to original size
+prediction = torch.nn.functional.interpolate(
+    predicted_depth.unsqueeze(1),
+    size=image.size[::-1],
+    mode="bicubic",
+    align_corners=False,
+)
+
+# visualize the prediction
+output = prediction.squeeze().cpu().numpy()
+formatted = (output * 255 / np.max(output)).astype("uint8")
+depth = Image.fromarray(formatted)
+```
+
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with ZoeDepth.
+
+- A demo notebook regarding inference with ZoeDepth models can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/ZoeDepth). 🌎
+
+## ZoeDepthConfig
+
+[[autodoc]] ZoeDepthConfig
+
+## ZoeDepthImageProcessor
+
+[[autodoc]] ZoeDepthImageProcessor
+    - preprocess
+
+## ZoeDepthForDepthEstimation
+
+[[autodoc]] ZoeDepthForDepthEstimation
+    - forward

docs/source/en/model_memory_anatomy.md

@@ -42,7 +42,7 @@ In total, we get 512 sequences each with length 512 and store them in a [`~datas
 >>> seq_len, dataset_size = 512, 512
 >>> dummy_data = {
 ...     "input_ids": np.random.randint(100, 30000, (dataset_size, seq_len)),
-...     "labels": np.random.randint(0, 1, (dataset_size)),
+...     "labels": np.random.randint(0, 2, (dataset_size)),
 ... }
 >>> ds = Dataset.from_dict(dummy_data)
 >>> ds.set_format("pt")

docs/source/en/model_sharing.md

@@ -77,7 +77,7 @@ Then use `notebook_login` to sign-in to the Hub, and follow the link [here](http
 
 To ensure your model can be used by someone working with a different framework, we recommend you convert and upload your model with both PyTorch and TensorFlow checkpoints. While users are still able to load your model from a different framework if you skip this step, it will be slower because 🤗 Transformers will need to convert the checkpoint on-the-fly.
 
-Converting a checkpoint for another framework is easy. Make sure you have PyTorch and TensorFlow installed (see [here](installation) for installation instructions), and then find the specific model for your task in the other framework. 
+Converting a checkpoint for another framework is easy. Make sure you have PyTorch and TensorFlow installed (see [here](installation) for installation instructions), and then find the specific model for your task in the other framework.
 
 <frameworkcontent>
 <pt>

docs/source/en/perf_hardware.md

@@ -116,7 +116,7 @@ Each new generation provides a faster bandwidth, e.g. here is a quote from [Nvid
 
 So the higher `X` you get in the report of `NVX` in the output of `nvidia-smi topo -m` the better. The generation will depend on your GPU architecture.
 
-Let's compare the execution of a openai-community/gpt2 language model training over a small sample of wikitext.
+Let's compare the execution of an openai-community/gpt2 language model training over a small sample of wikitext.
 
 The results are:
 

docs/source/en/perf_infer_gpu_one.md

@@ -39,15 +39,19 @@ FlashAttention-2 is experimental and may change considerably in future versions.
 FlashAttention-2 is currently supported for the following architectures:
 * [Bark](https://huggingface.co/docs/transformers/model_doc/bark#transformers.BarkModel)
 * [Bart](https://huggingface.co/docs/transformers/model_doc/bart#transformers.BartModel)
+* [Chameleon](https://huggingface.co/docs/transformers/model_doc/chameleon#transformers.Chameleon)
+* [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPModel)
 * [Cohere](https://huggingface.co/docs/transformers/model_doc/cohere#transformers.CohereModel)
 * [Dbrx](https://huggingface.co/docs/transformers/model_doc/dbrx#transformers.DbrxModel)
 * [DistilBert](https://huggingface.co/docs/transformers/model_doc/distilbert#transformers.DistilBertModel)
 * [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel)
+* [Gemma2](https://huggingface.co/docs/transformers/model_doc/gemma2#transformers.Gemma2Model)
 * [GPT2](https://huggingface.co/docs/transformers/model_doc/gpt2)
 * [GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode#transformers.GPTBigCodeModel)
 * [GPTNeo](https://huggingface.co/docs/transformers/model_doc/gpt_neo#transformers.GPTNeoModel)
 * [GPTNeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox#transformers.GPTNeoXModel)
 * [GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj#transformers.GPTJModel)
+* [Granite](https://huggingface.co/docs/transformers/model_doc/granite#transformers.GraniteModel)
 * [Idefics2](https://huggingface.co/docs/transformers/model_doc/idefics2#transformers.Idefics2Model)
 * [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
 * [JetMoe](https://huggingface.co/docs/transformers/model_doc/jetmoe#transformers.JetMoeModel)
@@ -64,6 +68,7 @@ FlashAttention-2 is currently supported for the following architectures:
 * [Mixtral](https://huggingface.co/docs/transformers/model_doc/mixtral#transformers.MixtralModel)
 * [Musicgen](https://huggingface.co/docs/transformers/model_doc/musicgen#transformers.MusicgenModel)
 * [MusicGen Melody](https://huggingface.co/docs/transformers/model_doc/musicgen_melody#transformers.MusicgenMelodyModel)
+* [Nemotron](https://huggingface.co/docs/transformers/model_doc/nemotron)
 * [NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)
 * [OLMo](https://huggingface.co/docs/transformers/model_doc/olmo#transformers.OlmoModel)
 * [OPT](https://huggingface.co/docs/transformers/model_doc/opt#transformers.OPTModel)
@@ -72,12 +77,15 @@ FlashAttention-2 is currently supported for the following architectures:
 * [StableLm](https://huggingface.co/docs/transformers/model_doc/stablelm#transformers.StableLmModel)
 * [Starcoder2](https://huggingface.co/docs/transformers/model_doc/starcoder2#transformers.Starcoder2Model)
 * [Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2#transformers.Qwen2Model)
+* [Qwen2Audio](https://huggingface.co/docs/transformers/model_doc/qwen2_audio#transformers.Qwen2AudioEncoder)
 * [Qwen2MoE](https://huggingface.co/docs/transformers/model_doc/qwen2_moe#transformers.Qwen2MoeModel)
+* [Qwen2VL](https://huggingface.co/docs/transformers/model_doc/qwen2_vl#transformers.Qwen2VLModel)
 * [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel)
 * [Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2#transformers.Wav2Vec2Model)
 * [Hubert](https://huggingface.co/docs/transformers/model_doc/hubert#transformers.HubertModel)
 * [data2vec_audio](https://huggingface.co/docs/transformers/main/en/model_doc/data2vec#transformers.Data2VecAudioModel)
 * [Sew](https://huggingface.co/docs/transformers/main/en/model_doc/sew#transformers.SEWModel)
+* [SigLIP](https://huggingface.co/docs/transformers/model_doc/siglip)
 * [UniSpeech](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech#transformers.UniSpeechModel)
 * [unispeech_sat](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech-sat#transformers.UniSpeechSatModel)
 
@@ -196,21 +204,34 @@ For now, Transformers supports SDPA inference and training for the following arc
 * [Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer#transformers.ASTModel)
 * [Bart](https://huggingface.co/docs/transformers/model_doc/bart#transformers.BartModel)
 * [Bert](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertModel)
+* [CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert#transformers.CamembertModel)
+* [Chameleon](https://huggingface.co/docs/transformers/model_doc/chameleon#transformers.Chameleon)
+* [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPModel)
 * [Cohere](https://huggingface.co/docs/transformers/model_doc/cohere#transformers.CohereModel)
+* [data2vec_audio](https://huggingface.co/docs/transformers/main/en/model_doc/data2vec#transformers.Data2VecAudioModel)
 * [Dbrx](https://huggingface.co/docs/transformers/model_doc/dbrx#transformers.DbrxModel)
 * [DeiT](https://huggingface.co/docs/transformers/model_doc/deit#transformers.DeiTModel)
 * [Dpr](https://huggingface.co/docs/transformers/model_doc/dpr#transformers.DprReader)
 * [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
 * [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel)
+* [Gemma2](https://huggingface.co/docs/transformers/model_doc/gemma2#transformers.Gemma2Model)
 * [GPT2](https://huggingface.co/docs/transformers/model_doc/gpt2)
 * [GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode#transformers.GPTBigCodeModel)
 * [GPTNeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox#transformers.GPTNeoXModel)
+* [Hubert](https://huggingface.co/docs/transformers/model_doc/hubert#transformers.HubertModel)
+* [Idefics](https://huggingface.co/docs/transformers/model_doc/idefics#transformers.IdeficsModel)
+* [Granite](https://huggingface.co/docs/transformers/model_doc/granite#transformers.GraniteModel)
 * [JetMoe](https://huggingface.co/docs/transformers/model_doc/jetmoe#transformers.JetMoeModel)
 * [Jamba](https://huggingface.co/docs/transformers/model_doc/jamba#transformers.JambaModel)
 * [Llama](https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel)
+* [Mistral](https://huggingface.co/docs/transformers/model_doc/mistral#transformers.MistralModel)
+* [Mixtral](https://huggingface.co/docs/transformers/model_doc/mixtral#transformers.MixtralModel)
+* [Musicgen](https://huggingface.co/docs/transformers/model_doc/musicgen#transformers.MusicgenModel)
+* [MusicGen Melody](https://huggingface.co/docs/transformers/model_doc/musicgen_melody#transformers.MusicgenMelodyModel)
 * [OLMo](https://huggingface.co/docs/transformers/model_doc/olmo#transformers.OlmoModel)
 * [PaliGemma](https://huggingface.co/docs/transformers/model_doc/paligemma#transformers.PaliGemmaForConditionalGeneration)
 * [Phi](https://huggingface.co/docs/transformers/model_doc/phi#transformers.PhiModel)
+* [Phi3](https://huggingface.co/docs/transformers/model_doc/phi3#transformers.Phi3Model)
 * [Idefics](https://huggingface.co/docs/transformers/model_doc/idefics#transformers.IdeficsModel)
 * [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel)
 * [Mistral](https://huggingface.co/docs/transformers/model_doc/mistral#transformers.MistralModel)
@@ -218,20 +239,29 @@ For now, Transformers supports SDPA inference and training for the following arc
 * [StableLm](https://huggingface.co/docs/transformers/model_doc/stablelm#transformers.StableLmModel)
 * [Starcoder2](https://huggingface.co/docs/transformers/model_doc/starcoder2#transformers.Starcoder2Model)
 * [Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2#transformers.Qwen2Model)
+* [Qwen2Audio](https://huggingface.co/docs/transformers/model_doc/qwen2_audio#transformers.Qwen2AudioEncoder)
 * [Qwen2MoE](https://huggingface.co/docs/transformers/model_doc/qwen2_moe#transformers.Qwen2MoeModel)
+* [RoBERTa](https://huggingface.co/docs/transformers/model_doc/roberta#transformers.RobertaModel)
+* [Sew](https://huggingface.co/docs/transformers/main/en/model_doc/sew#transformers.SEWModel)
+* [SigLIP](https://huggingface.co/docs/transformers/model_doc/siglip)
+* [StableLm](https://huggingface.co/docs/transformers/model_doc/stablelm#transformers.StableLmModel)
+* [Starcoder2](https://huggingface.co/docs/transformers/model_doc/starcoder2#transformers.Starcoder2Model)
+* [UniSpeech](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech#transformers.UniSpeechModel)
+* [unispeech_sat](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech-sat#transformers.UniSpeechSatModel)
+* [RoBERTa](https://huggingface.co/docs/transformers/model_doc/roberta#transformers.RobertaModel)
+* [Qwen2VL](https://huggingface.co/docs/transformers/model_doc/qwen2_vl#transformers.Qwen2VLModel)
 * [Musicgen](https://huggingface.co/docs/transformers/model_doc/musicgen#transformers.MusicgenModel)
 * [MusicGen Melody](https://huggingface.co/docs/transformers/model_doc/musicgen_melody#transformers.MusicgenMelodyModel)
+* [Nemotron](https://huggingface.co/docs/transformers/model_doc/nemotron)
 * [ViT](https://huggingface.co/docs/transformers/model_doc/vit#transformers.ViTModel)
 * [ViTHybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid#transformers.ViTHybridModel)
 * [ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae#transformers.ViTMAEModel)
 * [ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn#transformers.ViTMSNModel)
 * [VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae#transformers.VideoMAEModell)
 * [wav2vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2#transformers.Wav2Vec2Model)
-* [Hubert](https://huggingface.co/docs/transformers/model_doc/hubert#transformers.HubertModel)
-* [data2vec_audio](https://huggingface.co/docs/transformers/main/en/model_doc/data2vec#transformers.Data2VecAudioModel)
-* [Sew](https://huggingface.co/docs/transformers/main/en/model_doc/sew#transformers.SEWModel)
-* [UniSpeech](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech#transformers.UniSpeechModel)
-* [unispeech_sat](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech-sat#transformers.UniSpeechSatModel)
+* [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel)
+* [XLM-RoBERTa](https://huggingface.co/docs/transformers/model_doc/xlm-roberta#transformers.XLMRobertaModel)
+* [XLM-RoBERTa-XL](https://huggingface.co/docs/transformers/model_doc/xlm-roberta-xl#transformers.XLMRobertaXLModel)
 * [YOLOS](https://huggingface.co/docs/transformers/model_doc/yolos#transformers.YolosModel)
 
 

docs/source/en/perf_torch_compile.md

@@ -98,7 +98,7 @@ Below you can find the list of the models we benchmarked.
 - [google/vit-base-patch16-224](https://huggingface.co/google/vit-base-patch16-224)
 - [microsoft/beit-base-patch16-224-pt22k-ft22k](https://huggingface.co/microsoft/beit-base-patch16-224-pt22k-ft22k)
 - [facebook/convnext-large-224](https://huggingface.co/facebook/convnext-large-224)
-- [microsoft/resnet-50](https://huggingface.co/)
+- [microsoft/resnet-50](https://huggingface.co/microsoft/resnet-50)
 
 **Image Segmentation** 
 - [nvidia/segformer-b0-finetuned-ade-512-512](https://huggingface.co/nvidia/segformer-b0-finetuned-ade-512-512)

docs/source/en/perf_train_cpu_many.md

@@ -155,13 +155,20 @@ This example assumes that you have:
 The snippet below is an example of a Dockerfile that uses a base image that supports distributed CPU training and then
 extracts a Transformers release to the `/workspace` directory, so that the example scripts are included in the image:
 ```dockerfile
-FROM intel/ai-workflows:torch-2.0.1-huggingface-multinode-py3.9
+FROM intel/intel-optimized-pytorch:2.3.0-pip-multinode
+
+RUN apt-get update -y && \
+    apt-get install -y --no-install-recommends --fix-missing \
+    google-perftools \
+    libomp-dev
 
 WORKDIR /workspace
 
 # Download and extract the transformers code
-ARG HF_TRANSFORMERS_VER="4.35.2"
-RUN mkdir transformers && \
+ARG HF_TRANSFORMERS_VER="4.44.0"
+RUN pip install --no-cache-dir \
+    transformers==${HF_TRANSFORMERS_VER} && \
+    mkdir transformers && \
     curl -sSL --retry 5 https://github.com/huggingface/transformers/archive/refs/tags/v${HF_TRANSFORMERS_VER}.tar.gz | tar -C transformers --strip-components=1 -xzf -
 ```
 The image needs to be built and copied to the cluster's nodes or pushed to a container registry prior to deploying the
@@ -189,7 +196,6 @@ apiVersion: "kubeflow.org/v1"
 kind: PyTorchJob
 metadata:
   name: transformers-pytorchjob
-  namespace: kubeflow
 spec:
   elasticPolicy:
     rdzvBackend: c10d
@@ -206,32 +212,27 @@ spec:
             - name: pytorch
               image: <image name>:<tag>  # Specify the docker image to use for the worker pods
               imagePullPolicy: IfNotPresent
-              command:
-                - torchrun
-                - /workspace/transformers/examples/pytorch/question-answering/run_qa.py
-                - --model_name_or_path
-                - "google-bert/bert-large-uncased"
-                - --dataset_name
-                - "squad"
-                - --do_train
-                - --do_eval
-                - --per_device_train_batch_size
-                - "12"
-                - --learning_rate
-                - "3e-5"
-                - --num_train_epochs
-                - "2"
-                - --max_seq_length
-                - "384"
-                - --doc_stride
-                - "128"
-                - --output_dir
-                - "/tmp/pvc-mount/output"
-                - --no_cuda
-                - --ddp_backend
-                - "ccl"
-                - --use_ipex
-                - --bf16  # Specify --bf16 if your hardware supports bfloat16
+              command: ["/bin/bash", "-c"]
+              args:
+                - >-
+                  cd /workspace/transformers;
+                  pip install -r /workspace/transformers/examples/pytorch/question-answering/requirements.txt;
+                  source /usr/local/lib/python3.10/dist-packages/oneccl_bindings_for_pytorch/env/setvars.sh;
+                  torchrun /workspace/transformers/examples/pytorch/question-answering/run_qa.py \
+                    --model_name_or_path distilbert/distilbert-base-uncased \
+                    --dataset_name squad \
+                    --do_train \
+                    --do_eval \
+                    --per_device_train_batch_size 12 \
+                    --learning_rate 3e-5 \
+                    --num_train_epochs 2 \
+                    --max_seq_length 384 \
+                    --doc_stride 128 \
+                    --output_dir /tmp/pvc-mount/output_$(date +%Y%m%d_%H%M%S) \
+                    --no_cuda \
+                    --ddp_backend ccl \
+                    --bf16 \
+                    --use_ipex;
               env:
               - name: LD_PRELOAD
                 value: "/usr/lib/x86_64-linux-gnu/libtcmalloc.so.4.5.9:/usr/local/lib/libiomp5.so"
@@ -244,13 +245,13 @@ spec:
               - name: CCL_WORKER_COUNT
                 value: "1"
               - name: OMP_NUM_THREADS  # Can be tuned for optimal performance
--                value: "56"
+                value: "240"
               resources:
                 limits:
-                  cpu: 200  # Update the CPU and memory limit values based on your nodes
+                  cpu: 240  # Update the CPU and memory limit values based on your nodes
                   memory: 128Gi
                 requests:
-                  cpu: 200  # Update the CPU and memory request values based on your nodes
+                  cpu: 240  # Update the CPU and memory request values based on your nodes
                   memory: 128Gi
               volumeMounts:
               - name: pvc-volume
@@ -258,8 +259,8 @@ spec:
               - mountPath: /dev/shm
                 name: dshm
           restartPolicy: Never
-          nodeSelector:  #  Optionally use the node selector to specify what types of nodes to use for the workers
-            node-type: spr
+          nodeSelector:  # Optionally use nodeSelector to match a certain node label for the worker pods
+            node-type: gnr
           volumes:
           - name: pvc-volume
             persistentVolumeClaim:
@@ -287,10 +288,12 @@ set the same CPU and memory amounts for both the resource limits and requests.
 After the PyTorchJob spec has been updated with values appropriate for your cluster and training job, it can be deployed
 to the cluster using:
 ```bash
-kubectl create -f pytorchjob.yaml
+export NAMESPACE=<specify your namespace>
+
+kubectl create -f pytorchjob.yaml -n ${NAMESPACE}
 ```
 
-The `kubectl get pods -n kubeflow` command can then be used to list the pods in the `kubeflow` namespace. You should see
+The `kubectl get pods -n ${NAMESPACE}` command can then be used to list the pods in your namespace. You should see
 the worker pods for the PyTorchJob that was just deployed. At first, they will probably have a status of "Pending" as
 the containers get pulled and created, then the status should change to "Running".
 ```
@@ -303,13 +306,13 @@ transformers-pytorchjob-worker-3                         1/1     Running
 ...
 ```
 
-The logs for worker can be viewed using `kubectl logs -n kubeflow <pod name>`. Add `-f` to stream the logs, for example:
+The logs for worker can be viewed using `kubectl logs <pod name> -n ${NAMESPACE}`. Add `-f` to stream the logs, for example:
 ```bash
-kubectl logs -n kubeflow transformers-pytorchjob-worker-0 -f
+kubectl logs transformers-pytorchjob-worker-0 -n ${NAMESPACE} -f
 ```
 
 After the training job completes, the trained model can be copied from the PVC or storage location. When you are done
-with the job, the PyTorchJob resource can be deleted from the cluster using `kubectl delete -f pytorchjob.yaml`.
+with the job, the PyTorchJob resource can be deleted from the cluster using `kubectl delete -f pytorchjob.yaml -n ${NAMESPACE}`.
 
 ## Summary
 

docs/source/en/perf_train_gpu_one.md

@@ -41,21 +41,22 @@ hyperparameter tuning, you should determine which batch size yields the best res
 
 The methods and tools covered in this guide can be classified based on the effect they have on the training process:
 
-| Method/tool                                                | Improves training speed | Optimizes memory utilization |
-|:-----------------------------------------------------------|:------------------------|:-----------------------------|
-| [Batch size choice](#batch-size-choice)                    | Yes                     | Yes                          |
-| [Gradient accumulation](#gradient-accumulation)            | No                      | Yes                          |
-| [Gradient checkpointing](#gradient-checkpointing)          | No                      | Yes                          |
-| [Mixed precision training](#mixed-precision-training)      | Yes                     | (No)                         |
-| [Optimizer choice](#optimizer-choice)                      | Yes                     | Yes                          |
-| [Data preloading](#data-preloading)                        | Yes                     | No                           |
-| [DeepSpeed Zero](#deepspeed-zero)                          | No                      | Yes                          |
-| [torch.compile](#using-torchcompile)                       | Yes                     | No                           |
-| [Parameter-Efficient Fine Tuning (PEFT)](#using--peft)            | No                      | Yes                          |
+| Method/tool                                                                                                                                             | Improves training speed | Optimizes memory utilization |
+|:--------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------|:-----------------------------|
+| [Batch size choice](#batch-size-choice)                                                                                                                 | Yes                     | Yes                          |
+| [Gradient accumulation](#gradient-accumulation)                                                                                                         | No                      | Yes                          |
+| [Gradient checkpointing](#gradient-checkpointing)                                                                                                       | No                      | Yes                          |
+| [Mixed precision training](#mixed-precision-training)                                                                                                   | Yes                     | Maybe*                       |
+| [torch_empty_cache_steps](https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.TrainingArguments.torch_empty_cache_steps) | No                      | Yes                          |
+| [Optimizer choice](#optimizer-choice)                                                                                                                   | Yes                     | Yes                          |
+| [Data preloading](#data-preloading)                                                                                                                     | Yes                     | No                           |
+| [DeepSpeed Zero](#deepspeed-zero)                                                                                                                       | No                      | Yes                          |
+| [torch.compile](#using-torchcompile)                                                                                                                    | Yes                     | No                           |
+| [Parameter-Efficient Fine Tuning (PEFT)](#using--peft)                                                                                                  | No                      | Yes                          |
  
 <Tip>
 
-Note: when using mixed precision with a small model and a large batch size, there will be some memory savings but with a 
+*Note: when using mixed precision with a small model and a large batch size, there will be some memory savings but with a 
 large model and a small batch size, the memory use will be larger.
 
 </Tip>

docs/source/en/pipeline_tutorial.md

@@ -54,7 +54,7 @@ speech-to-text.
 Not the result you had in mind? Check out some of the [most downloaded automatic speech recognition models](https://huggingface.co/models?pipeline_tag=automatic-speech-recognition&sort=trending) 
 on the Hub to see if you can get a better transcription.
 
-Let's try the [Whisper large-v2](https://huggingface.co/openai/whisper-large) model from OpenAI. Whisper was released 
+Let's try the [Whisper large-v2](https://huggingface.co/openai/whisper-large-v2) model from OpenAI. Whisper was released 
 2 years later than Wav2Vec2, and was trained on close to 10x more data. As such, it beats Wav2Vec2 on most downstream 
 benchmarks. It also has the added benefit of predicting punctuation and casing, neither of which are possible with  
 Wav2Vec2.
@@ -113,7 +113,9 @@ This will work regardless of whether you are using PyTorch or Tensorflow.
 transcriber = pipeline(model="openai/whisper-large-v2", device=0)
 ```
 
-If the model is too large for a single GPU and you are using PyTorch, you can set `device_map="auto"` to automatically 
+If the model is too large for a single GPU and you are using PyTorch, you can set `torch_dtype='float16'` to enable FP16 precision inference. Usually this would not cause significant performance drops but make sure you evaluate it on your models!
+
+Alternatively, you can set `device_map="auto"` to automatically 
 determine how to load and store the model weights. Using the `device_map` argument requires the 🤗 [Accelerate](https://huggingface.co/docs/accelerate)
 package:
 
@@ -342,4 +344,3 @@ gr.Interface.from_pipeline(pipe).launch()
 
 By default, the web demo runs on a local server. If you'd like to share it with others, you can generate a temporary public
 link by setting `share=True` in `launch()`. You can also host your demo on [Hugging Face Spaces](https://huggingface.co/spaces) for a permanent link. 
-

docs/source/en/preprocessing.md

@@ -471,7 +471,7 @@ from [`DetrImageProcessor`] and define a custom `collate_fn` to batch images tog
 
 ## Multimodal
 
-For tasks involving multimodal inputs, you'll need a [processor](main_classes/processors) to prepare your dataset for the model. A processor couples together two processing objects such as as tokenizer and feature extractor.
+For tasks involving multimodal inputs, you'll need a [processor](main_classes/processors) to prepare your dataset for the model. A processor couples together two processing objects such as tokenizer and feature extractor.
 
 Load the [LJ Speech](https://huggingface.co/datasets/lj_speech) dataset (see the 🤗 [Datasets tutorial](https://huggingface.co/docs/datasets/load_hub) for more details on how to load a dataset) to see how you can use a processor for automatic speech recognition (ASR):
 

docs/source/en/quantization/fbgemm_fp8.md

@@ -0,0 +1,58 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# FBGEMM FP8
+
+With FBGEMM FP8 quantization method, you can quantize your model in FP8 (W8A8):
+- the weights will be quantized in 8bit (FP8) per channel
+- the activation will be quantized in 8bit (FP8) per token
+
+It relies on the [FBGEMM](https://github.com/pytorch/FBGEMM) library which provides efficient low-precision general matrix multiplication for small batch sizes and support for accuracy-loss minimizing techniques such as row-wise quantization and outlier-aware quantization. 
+
+> [!TIP]
+> You need a GPU with compute capability>=9 (e.g. H100) 
+
+Before you begin, make sure the following libraries are installed with their latest version:
+
+```bash
+pip install --upgrade accelerate fbgemm-gpu torch
+```
+
+If you are having issues with fbgemm-gpu and torch library, you might need to install the nighlty release. You can follow the instruction [here](https://pytorch.org/FBGEMM/fbgemm_gpu-development/InstallationInstructions.html#fbgemm-gpu-install-libraries:~:text=found%20here.-,Install%20the%20FBGEMM_GPU%20Package,-Install%20through%20PyTorch)
+
+
+```py
+from transformers import FbgemmFp8Config, AutoModelForCausalLM, AutoTokenizer
+
+model_name = "meta-llama/Meta-Llama-3-8B"
+quantization_config = FbgemmFp8Config()
+quantized_model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", quantization_config=quantization_config)
+
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+output = quantized_model.generate(**input_ids, max_new_tokens=10)
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+
+A quantized model can be saved via "saved_pretrained" and be reused again via the "from_pretrained".
+
+```py
+quant_path = "/path/to/save/quantized/model"
+model.save_pretrained(quant_path)
+model = AutoModelForCausalLM.from_pretrained(quant_path, device_map="auto")
+```

docs/source/en/quantization/overview.md

@@ -55,4 +55,5 @@ Use the table below to help you decide which quantization method to use.
 | [GPTQ](./gptq)                                | 🔴                       | 🔴   | 🟢        | 🟢              | 🔴                     | 🔴                       | 2 - 3 - 4 - 8          | 🟢                                   | 🟢            | 🟢                      | https://github.com/AutoGPTQ/AutoGPTQ        |
 | [HQQ](./hqq)                                 | 🟢                       | 🟢    | 🟢        | 🔴              | 🔴                     | 🟢                       | 1 - 8          | 🟢                                   | 🔴            | 🟢                      | https://github.com/mobiusml/hqq/            |
 | [Quanto](./quanto)                              | 🟢                       | 🟢   | 🟢        | 🔴              | 🟢                     | 🟢                       | 2 / 4 / 8      | 🔴                                   | 🔴            | 🟢                      | https://github.com/huggingface/quanto       |
-
+| [FBGEMM_FP8](./fbgemm_fp8.md)                              | 🟢                       | 🔴    | 🟢        | 🔴              | 🔴                      | 🔴                        | 8      | 🔴                                   | 🟢            | 🟢                      | https://github.com/pytorch/FBGEMM       |
+| [torchao](./torchao.md)                              | 🟢                       |     | 🟢        | 🔴              | partial support (int4 weight only)       |                       | 4 / 8      |                                   | 🟢🔴           | 🟢                      | https://github.com/pytorch/ao       |

docs/source/en/quantization/torchao.md

@@ -0,0 +1,45 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# TorchAO
+
+[TorchAO](https://github.com/pytorch/ao) is an architecture optimization library for PyTorch, it provides high performance dtypes, optimization techniques and kernels for inference and training, featuring composability with native PyTorch features like `torch.compile`, FSDP etc.. Some benchmark numbers can be found [here](https://github.com/pytorch/ao/tree/main?tab=readme-ov-file#without-intrusive-code-changes)
+
+Before you begin, make sure the following libraries are installed with their latest version:
+
+```bash
+pip install --upgrade torch torchao
+```
+
+
+```py
+from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+
+model_name = "meta-llama/Meta-Llama-3-8B"
+# We support int4_weight_only, int8_weight_only and int8_dynamic_activation_int8_weight
+# More examples and documentations for arguments can be found in https://github.com/pytorch/ao/tree/main/torchao/quantization#other-available-quantization-techniques
+quantization_config = TorchAoConfig("int4_weight_only", group_size=128)
+quantized_model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", quantization_config=quantization_config)
+
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+# compile the quantizd model to get speedup
+import torchao
+torchao.quantization.utils.recommended_inductor_config_setter()
+quantized_model = torch.compile(quantized_model, mode="max-autotune")
+
+output = quantized_model.generate(**input_ids, max_new_tokens=10)
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+
+torchao quantization is implemented with tensor subclasses, currently it does not work with huggingface serialization, both the safetensor option and [non-safetensor option](https://github.com/huggingface/transformers/issues/32364), we'll update here with instructions when it's working.

docs/source/en/tasks/image_text_to_text.md

@@ -0,0 +1,232 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Image-text-to-text
+
+[[open-in-colab]]
+
+Image-text-to-text models, also known as vision language models (VLMs), are language models that take an image input. These models can tackle various tasks, from visual question answering to image segmentation. This task shares many similarities with image-to-text, but with some overlapping use cases like image captioning. Image-to-text models only take image inputs and often accomplish a specific task, whereas VLMs take open-ended text and image inputs and are more generalist models.
+
+In this guide, we provide a brief overview of VLMs and show how to use them with Transformers for inference.
+
+To begin with, there are multiple types of VLMs:
+- base models used for fine-tuning
+- chat fine-tuned models for conversation
+- instruction fine-tuned models
+
+This guide focuses on inference with an instruction-tuned model. 
+
+Let's begin installing the dependencies.
+
+```bash
+pip install -q transformers accelerate flash_attn 
+```
+
+Let's initialize the model and the processor. 
+
+```python
+from transformers import AutoProcessor, Idefics2ForConditionalGeneration
+import torch
+
+device = torch.device("cuda")
+model = Idefics2ForConditionalGeneration.from_pretrained(
+    "HuggingFaceM4/idefics2-8b",
+    torch_dtype=torch.bfloat16,
+    attn_implementation="flash_attention_2",
+).to(device)
+
+processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-8b")
+```
+
+This model has a [chat template](./chat_templating) that helps user parse chat outputs. Moreover, the model can also accept multiple images as input in a single conversation or message. We will now prepare the inputs. 
+
+The image inputs look like the following.
+
+<div class="flex justify-center">
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.png" alt="Two cats sitting on a net"/>
+</div>
+
+<div class="flex justify-center">
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg" alt="A bee on a pink flower"/>
+</div>
+
+
+```python
+from PIL import Image
+import requests
+
+img_urls =["https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.png",
+           "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"]
+images = [Image.open(requests.get(img_urls[0], stream=True).raw),
+          Image.open(requests.get(img_urls[1], stream=True).raw)]
+```
+
+Below is an example of the chat template. We can feed conversation turns and the last message as an input by appending it at the end of the template. 
+
+
+```python
+messages = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": "What do we see in this image?"},
+        ]
+    },
+    {
+        "role": "assistant",
+        "content": [
+            {"type": "text", "text": "In this image we can see two cats on the nets."},
+        ]
+    },
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": "And how about this image?"},
+        ]
+    },       
+]
+```
+
+We will now call the processors' [`~ProcessorMixin.apply_chat_template`] method to preprocess its output along with the image inputs.
+
+```python
+prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
+inputs = processor(text=prompt, images=[images[0], images[1]], return_tensors="pt").to(device)
+```
+
+We can now pass the preprocessed inputs to the model.
+
+```python
+with torch.no_grad():
+    generated_ids = model.generate(**inputs, max_new_tokens=500)
+generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+print(generated_texts)
+## ['User: What do we see in this image? \nAssistant: In this image we can see two cats on the nets. \nUser: And how about this image? \nAssistant: In this image we can see flowers, plants and insect.']
+```
+
+## Streaming
+
+We can use [text streaming](./generation_strategies#streaming) for a better generation experience. Transformers supports streaming with the [`TextStreamer`] or [`TextIteratorStreamer`] classes. We will use the [`TextIteratorStreamer`] with IDEFICS-8B.
+
+Assume we have an application that keeps chat history and takes in the new user input. We will preprocess the inputs as usual and initialize [`TextIteratorStreamer`] to handle the generation in a separate thread. This allows you to stream the generated text tokens in real-time. Any generation arguments can be passed to [`TextIteratorStreamer`].
+
+
+```python
+import time
+from transformers import TextIteratorStreamer
+from threading import Thread
+
+def model_inference(
+    user_prompt,
+    chat_history,
+    max_new_tokens,
+    images
+):
+    user_prompt = {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "text", "text": user_prompt},
+        ]
+    }
+    chat_history.append(user_prompt)
+    streamer = TextIteratorStreamer(
+        processor.tokenizer,
+        skip_prompt=True,
+        timeout=5.0,
+    )
+
+    generation_args = {
+        "max_new_tokens": max_new_tokens,
+        "streamer": streamer,
+        "do_sample": False
+    }
+
+    # add_generation_prompt=True makes model generate bot response
+    prompt = processor.apply_chat_template(chat_history, add_generation_prompt=True)
+    inputs = processor(
+        text=prompt,
+        images=images,
+        return_tensors="pt",
+    ).to(device)
+    generation_args.update(inputs)
+
+    thread = Thread(
+        target=model.generate,
+        kwargs=generation_args,
+    )
+    thread.start()
+
+    acc_text = ""
+    for text_token in streamer:
+        time.sleep(0.04)
+        acc_text += text_token
+        if acc_text.endswith("<end_of_utterance>"):
+            acc_text = acc_text[:-18]
+        yield acc_text
+    
+    thread.join()
+```
+
+Now let's call the `model_inference` function we created and stream the values. 
+
+```python
+generator = model_inference(
+    user_prompt="And what is in this image?",
+    chat_history=messages,
+    max_new_tokens=100,
+    images=images
+)
+
+for value in generator:
+  print(value)
+
+# In
+# In this
+# In this image ...
+```
+
+## Fit models in smaller hardware
+
+VLMs are often large and need to be optimized to fit in smaller hardware. Transformers supports many model quantization libraries, and here we will only show int8 quantization with [Quanto](./quantization/quanto#quanto). int8 quantization offers memory improvements up to 75 percent (if all weights are quantized). However it is no free lunch, since 8-bit is not a CUDA-native precision, the weights are quantized back and forth on the fly, which adds up to latency. 
+
+First, install dependencies.
+
+```bash
+pip install -U quanto bitsandbytes
+```
+
+To quantize a model during loading, we need to first create [`QuantoConfig`]. Then load the model as usual, but pass `quantization_config` during model initialization.
+
+```python
+from transformers import Idefics2ForConditionalGeneration, AutoTokenizer, QuantoConfig
+
+model_id = "HuggingFaceM4/idefics2-8b"
+quantization_config = QuantoConfig(weights="int8")
+quantized_model = Idefics2ForConditionalGeneration.from_pretrained(model_id, device_map="cuda", quantization_config=quantization_config)
+```
+
+And that's it, we can use the model the same way with no changes. 
+
+## Further Reading
+
+Here are some more resources for the image-text-to-text task.
+
+- [Image-text-to-text task page](https://huggingface.co/tasks/image-text-to-text) covers model types, use cases, datasets, and more. 
+- [Vision Language Models Explained](https://huggingface.co/blog/vlms) is a blog post that covers everything about vision language models and supervised fine-tuning using [TRL](https://huggingface.co/docs/trl/en/index).

docs/source/en/tasks/monocular_depth_estimation.md

@@ -23,23 +23,26 @@ a single camera viewpoint.
 Monocular depth estimation has various applications, including 3D reconstruction, augmented reality, autonomous driving,
 and robotics. It is a challenging task as it requires the model to understand the complex relationships between objects
 in the scene and the corresponding depth information, which can be affected by factors such as lighting conditions,
-occlusion, and texture.
+occlusion, and texture. 
+
+There are two main depth estimation categories:
+
+- **Absolute depth estimation**: This task variant aims to provide exact depth measurements from the camera. The term is used interchangeably with metric depth estimation, where depth is provided in precise measurements in meters or feet. Absolute depth estimation models output depth maps with numerical values that represent real-world distances.
+
+- **Relative depth estimation**: Relative depth estimation aims to predict the depth order of objects or points in a scene without providing the precise measurements. These models output a depth map that indicates which parts of the scene are closer or farther relative to each other without the actual distances to A and B.
+
+In this guide, we will see how to infer with [Depth Anything V2](https://huggingface.co/depth-anything/Depth-Anything-V2-Large), a state-of-the-art zero-shot relative depth estimation model, and [ZoeDepth](https://huggingface.co/docs/transformers/main/en/model_doc/zoedepth), an absolute depth estimation model.
 
 <Tip>
 
-To see all architectures and checkpoints compatible with this task, we recommend checking the [task-page](https://huggingface.co/tasks/depth-anything)
+Check the [Depth Estimation](https://huggingface.co/tasks/depth-estimation) task page to view all compatible architectures and checkpoints.
 
 </Tip>
 
-In this guide you'll learn how to:
-
-* create a depth estimation pipeline
-* run depth estimation inference by hand
-
-Before you begin, make sure you have all the necessary libraries installed:
+Before we begin, we need to install the latest version of Transformers:
 
 ```bash
-pip install -q transformers
+pip install -q -U transformers
 ```
 
 ## Depth estimation pipeline
@@ -49,9 +52,11 @@ Instantiate a pipeline from a [checkpoint on the Hugging Face Hub](https://huggi
 
 ```py
 >>> from transformers import pipeline
+>>> import torch
 
->>> checkpoint = "vinvino02/glpn-nyu"
->>> depth_estimator = pipeline("depth-estimation", model=checkpoint)
+>>> device = "cuda" if torch.cuda.is_available() else "cpu"
+>>> checkpoint = "depth-anything/Depth-Anything-V2-base-hf"
+>>> pipe = pipeline("depth-estimation", model=checkpoint, device=device)
 ```
 
 Next, choose an image to analyze:
@@ -60,19 +65,19 @@ Next, choose an image to analyze:
 >>> from PIL import Image
 >>> import requests
 
->>> url = "https://unsplash.com/photos/HwBAsSbPBDU/download?ixid=MnwxMjA3fDB8MXxzZWFyY2h8MzR8fGNhciUyMGluJTIwdGhlJTIwc3RyZWV0fGVufDB8MHx8fDE2Nzg5MDEwODg&force=true&w=640"
+>>> url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"
 >>> image = Image.open(requests.get(url, stream=True).raw)
 >>> image
 ```
 
 <div class="flex justify-center">
-     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/depth-estimation-example.jpg" alt="Photo of a busy street"/>
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg" alt="Photo of a bee"/>
 </div>
 
 Pass the image to the pipeline.
 
 ```py
->>> predictions = depth_estimator(image)
+>>> predictions = pipe(image)
 ```
 
 The pipeline returns a dictionary with two entries. The first one, called `predicted_depth`, is a tensor with the values
@@ -99,17 +104,17 @@ Here we'll use the same checkpoint as before:
 ```py
 >>> from transformers import AutoImageProcessor, AutoModelForDepthEstimation
 
->>> checkpoint = "vinvino02/glpn-nyu"
+>>> checkpoint = "Intel/zoedepth-nyu-kitti"
 
 >>> image_processor = AutoImageProcessor.from_pretrained(checkpoint)
->>> model = AutoModelForDepthEstimation.from_pretrained(checkpoint)
+>>> model = AutoModelForDepthEstimation.from_pretrained(checkpoint).to(device)
 ```
 
 Prepare the image input for the model using the `image_processor` that will take care of the necessary image transformations
 such as resizing and normalization:
 
 ```py
->>> pixel_values = image_processor(image, return_tensors="pt").pixel_values
+>>> pixel_values = image_processor(image, return_tensors="pt").pixel_values.to(device)
 ```
 
 Pass the prepared inputs through the model:
@@ -119,28 +124,100 @@ Pass the prepared inputs through the model:
 
 >>> with torch.no_grad():
 ...     outputs = model(pixel_values)
-...     predicted_depth = outputs.predicted_depth
 ```
 
-Visualize the results:
+Let's post-process and visualize the results. 
+
+We need to pad and then resize the outputs so that predicted depth map has the same dimension as the original image. After resizing we will remove the padded regions from the depth. 
 
 ```py
 >>> import numpy as np
+>>> import torch.nn.functional as F
 
->>> # interpolate to original size
->>> prediction = torch.nn.functional.interpolate(
-...     predicted_depth.unsqueeze(1),
-...     size=image.size[::-1],
-...     mode="bicubic",
-...     align_corners=False,
-... ).squeeze()
->>> output = prediction.numpy()
+>>> predicted_depth = outputs.predicted_depth.unsqueeze(dim=1)
+>>> height, width = pixel_values.shape[2:]
 
->>> formatted = (output * 255 / np.max(output)).astype("uint8")
->>> depth = Image.fromarray(formatted)
->>> depth
+>>> height_padding_factor = width_padding_factor = 3
+>>> pad_h = int(np.sqrt(height/2) * height_padding_factor)
+>>> pad_w = int(np.sqrt(width/2) * width_padding_factor)
+
+>>> if predicted_depth.shape[-2:] != pixel_values.shape[-2:]:
+>>>    predicted_depth = F.interpolate(predicted_depth, size= (height, width), mode='bicubic', align_corners=False)
+
+>>> if pad_h > 0:
+     predicted_depth = predicted_depth[:, :, pad_h:-pad_h,:]
+>>> if pad_w > 0:
+     predicted_depth = predicted_depth[:, :, :, pad_w:-pad_w]
 ```
 
+We can now visualize the results (the function below is taken from the [GaussianObject](https://github.com/GaussianObject/GaussianObject/blob/ad6629efadb57902d5f8bc0fa562258029a4bdf1/pred_monodepth.py#L11) framework).
+
+```py
+import matplotlib
+
+def colorize(value, vmin=None, vmax=None, cmap='gray_r', invalid_val=-99, invalid_mask=None, background_color=(128, 128, 128, 255), gamma_corrected=False, value_transform=None):
+    """Converts a depth map to a color image.
+
+    Args:
+        value (torch.Tensor, numpy.ndarry): Input depth map. Shape: (H, W) or (1, H, W) or (1, 1, H, W). All singular dimensions are squeezed
+        vmin (float, optional): vmin-valued entries are mapped to start color of cmap. If None, value.min() is used. Defaults to None.
+        vmax (float, optional):  vmax-valued entries are mapped to end color of cmap. If None, value.max() is used. Defaults to None.
+        cmap (str, optional): matplotlib colormap to use. Defaults to 'magma_r'.
+        invalid_val (int, optional): Specifies value of invalid pixels that should be colored as 'background_color'. Defaults to -99.
+        invalid_mask (numpy.ndarray, optional): Boolean mask for invalid regions. Defaults to None.
+        background_color (tuple[int], optional): 4-tuple RGB color to give to invalid pixels. Defaults to (128, 128, 128, 255).
+        gamma_corrected (bool, optional): Apply gamma correction to colored image. Defaults to False.
+        value_transform (Callable, optional): Apply transform function to valid pixels before coloring. Defaults to None.
+
+    Returns:
+        numpy.ndarray, dtype - uint8: Colored depth map. Shape: (H, W, 4)
+    """
+    if isinstance(value, torch.Tensor):
+        value = value.detach().cpu().numpy()
+
+    value = value.squeeze()
+    if invalid_mask is None:
+        invalid_mask = value == invalid_val
+    mask = np.logical_not(invalid_mask)
+
+    # normalize
+    vmin = np.percentile(value[mask],2) if vmin is None else vmin
+    vmax = np.percentile(value[mask],85) if vmax is None else vmax
+    if vmin != vmax:
+        value = (value - vmin) / (vmax - vmin)  # vmin..vmax
+    else:
+        # Avoid 0-division
+        value = value * 0.
+
+    # squeeze last dim if it exists
+    # grey out the invalid values
+
+    value[invalid_mask] = np.nan
+    cmapper = matplotlib.colormaps.get_cmap(cmap)
+    if value_transform:
+        value = value_transform(value)
+        # value = value / value.max()
+    value = cmapper(value, bytes=True)  # (nxmx4)
+
+    # img = value[:, :, :]
+    img = value[...]
+    img[invalid_mask] = background_color
+
+    #     return img.transpose((2, 0, 1))
+    if gamma_corrected:
+        # gamma correction
+        img = img / 255
+        img = np.power(img, 2.2)
+        img = img * 255
+        img = img.astype(np.uint8)
+    return img
+
+>>> result = colorize(predicted_depth.cpu().squeeze().numpy())
+>>> Image.fromarray(result)
+```
+
+
+
 <div class="flex justify-center">
-     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/depth-visualization.png" alt="Depth estimation visualization"/>
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/depth-visualization-zoe.png" alt="Depth estimation visualization"/>
 </div>

docs/source/en/tasks/translation.md

@@ -90,7 +90,7 @@ The next step is to load a T5 tokenizer to process the English-French language p
 The preprocessing function you want to create needs to:
 
 1. Prefix the input with a prompt so T5 knows this is a translation task. Some models capable of multiple NLP tasks require prompting for specific tasks.
-2. Tokenize the input (English) and target (French) separately because you can't tokenize French text with a tokenizer pretrained on an English vocabulary.
+2. Set the target language (French) in the `text_target` parameter to ensure the tokenizer processes the target text correctly. If you don't set `text_target`, the tokenizer processes the target text as English.
 3. Truncate sequences to be no longer than the maximum length set by the `max_length` parameter.
 
 ```py

docs/source/en/testing.md

@@ -191,7 +191,7 @@ RUN_SLOW=1 pytest -m accelerate_tests tests/models/opt/test_modeling_opt.py
 ### Run documentation tests
 
 In order to test whether the documentation examples are correct, you should check that the `doctests` are passing.
-As an example, let's use [`WhisperModel.forward`'s docstring](https://github.com/huggingface/transformers/blob/main/src/transformers/models/whisper/modeling_whisper.py#L1017-L1035):
+As an example, let's use [`WhisperModel.forward`'s docstring](https://github.com/huggingface/transformers/blob/1124d95dbb1a3512d3e80791d73d0f541d1d7e9f/src/transformers/models/whisper/modeling_whisper.py#L1591-L1609)
 
 ```python
 r"""
@@ -1011,7 +1011,7 @@ slow models to do qualitative testing. To see the use of these simply look for *
 grep tiny tests examples
 ```
 
-Here is a an example of a [script](https://github.com/huggingface/transformers/tree/main/scripts/fsmt/fsmt-make-tiny-model.py) that created the tiny model
+Here is an example of a [script](https://github.com/huggingface/transformers/tree/main/scripts/fsmt/fsmt-make-tiny-model.py) that created the tiny model
 [stas/tiny-wmt19-en-de](https://huggingface.co/stas/tiny-wmt19-en-de). You can easily adjust it to your specific
 model's architecture.
 

docs/source/en/tf_xla.md

@@ -157,7 +157,7 @@ Execution time -- 79.0 ms
 
 Execution time -- 78.9 ms
 ```
-The first call to `xla_generate()` is time-consuming because of tracing, but the successive calls are orders of magnitude faster. Keep in mind that any change in the generation options at any point with trigger re-tracing and thus leading to slow-downs in the generation time. 
+The first call to `xla_generate()` is time-consuming because of tracing, but the successive calls are orders of magnitude faster. Keep in mind that any change in the generation options at any point will trigger re-tracing and thus leading to slow-downs in the generation time. 
 
 We didn’t cover all the text generation options 🤗 Transformers provides in this document. We encourage you to read the documentation for advanced use cases.
 
@@ -171,4 +171,4 @@ Here, we leave you with some additional resources if you want to delve deeper in
 * Recommended posts for learning more about XLA and TensorFlow graphs in general:
     * [XLA: Optimizing Compiler for Machine Learning](https://www.tensorflow.org/xla)
     * [Introduction to graphs and tf.function](https://www.tensorflow.org/guide/intro_to_graphs)
-    * [Better performance with tf.function](https://www.tensorflow.org/guide/function) 
+    * [Better performance with tf.function](https://www.tensorflow.org/guide/function)