Release: v4.52.0

* fix attn mask

* attn viz doesn't show yello cubes between images

* bucketize made it hard with different number of crops

* fixup

.circleci/config.yml

@@ -7,6 +7,18 @@ parameters:
     nightly:
         type: boolean
         default: false
+    GHA_Actor:
+        type: string
+        default: ""
+    GHA_Action:
+        type: string
+        default: ""
+    GHA_Event:
+        type: string
+        default: ""
+    GHA_Meta:
+        type: string
+        default: ""
 
 jobs:
     # Ensure running with CircleCI/huggingface
@@ -31,8 +43,10 @@ jobs:
         parallelism: 1
         steps:
             - checkout
-            - run: if [[ "$CIRCLE_PULL_REQUEST" == "" && "$CIRCLE_BRANCH" != "main" && "$CIRCLE_BRANCH" != *-release ]]; then echo "Not a PR, not the main branch and not a release branch, skip test!"; circleci-agent step halt; fi
-            - run: 'curl -L -H "Accept: application/vnd.github+json" -H "X-GitHub-Api-Version: 2022-11-28" https://api.github.com/repos/$CIRCLE_PROJECT_USERNAME/$CIRCLE_PROJECT_REPONAME/pulls/${CIRCLE_PULL_REQUEST##*/} >> github.txt'
+            - run: python3 utils/extract_pr_number_from_circleci.py > pr_number.txt
+            - run: echo $(cat pr_number.txt)
+            - run: if [[ "$(cat pr_number.txt)" == "" && "$CIRCLE_BRANCH" != "main" && "$CIRCLE_BRANCH" != *-release ]]; then echo "Not a PR, not the main branch and not a release branch, skip test!"; circleci-agent step halt; fi
+            - run: 'curl -L -H "Accept: application/vnd.github+json" -H "X-GitHub-Api-Version: 2022-11-28" https://api.github.com/repos/$CIRCLE_PROJECT_USERNAME/$CIRCLE_PROJECT_REPONAME/pulls/$(cat pr_number.txt) >> github.txt'
             - run: cat github.txt
             - run: (python3 -c 'import json; from datetime import datetime; fp = open("github.txt"); data = json.load(fp); fp.close(); f = "%Y-%m-%dT%H:%M:%SZ"; created = datetime.strptime(data["created_at"], f); updated = datetime.strptime(data["updated_at"], f); s = (updated - created).total_seconds(); print(int(s))' || true) > elapsed.txt
             - run: if [ "$(cat elapsed.txt)" == "" ]; then echo 60 > elapsed.txt; fi

.circleci/create_circleci_config.py

@@ -28,6 +28,8 @@ COMMON_ENV_VARIABLES = {
     "TRANSFORMERS_IS_CI": True,
     "PYTEST_TIMEOUT": 120,
     "RUN_PIPELINE_TESTS": False,
+    # will be adjust in `CircleCIJob.to_dict`.
+    "RUN_FLAKY": True,
 }
 # Disable the use of {"s": None} as the output is way too long, causing the navigation on CircleCI impractical
 COMMON_PYTEST_OPTIONS = {"max-worker-restart": 0, "vvv": None, "rsfE":None}
@@ -108,6 +110,7 @@ class CircleCIJob:
             print(f"Using {self.docker_image} docker image")
         if self.install_steps is None:
             self.install_steps = ["uv venv && uv pip install ."]
+        self.install_steps.append("uv venv && uv pip install git+https://github.com/ydshieh/pytest.git@8.3.5-ydshieh git+https://github.com/ydshieh/pluggy.git@1.5.0-ydshieh")
         if self.pytest_options is None:
             self.pytest_options = {}
         if isinstance(self.tests_to_run, str):
@@ -126,6 +129,8 @@ class CircleCIJob:
 
     def to_dict(self):
         env = COMMON_ENV_VARIABLES.copy()
+        # Do not run tests decorated by @is_flaky on pull requests
+        env['RUN_FLAKY'] = os.environ.get("CIRCLE_PULL_REQUEST", "") == ""
         env.update(self.additional_env)
 
         job = {
@@ -393,7 +398,12 @@ def create_circleci_config(folder=None):
         "parameters": {
             # Only used to accept the parameters from the trigger
             "nightly": {"type": "boolean", "default": False},
-            "tests_to_run": {"type": "string", "default": ''},
+            # Only used to accept the parameters from GitHub Actions trigger
+            "GHA_Actor": {"type": "string", "default": ""},
+            "GHA_Action": {"type": "string", "default": ""},
+            "GHA_Event": {"type": "string", "default": ""},
+            "GHA_Meta": {"type": "string", "default": ""},
+            "tests_to_run": {"type": "string", "default": ""},
             **{j.job_name + "_test_list":{"type":"string", "default":''} for j in jobs},
             **{j.job_name + "_parallelism":{"type":"integer", "default":1} for j in jobs},
         },

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -16,7 +16,7 @@ body:
     id: system-info
     attributes:
       label: System Info
-      description: Please share your system info with us. You can run the command `transformers-cli env` and copy-paste its output below.
+      description: Please share your system info with us. You can run the command `transformers env` and copy-paste its output below.
       placeholder: transformers version, platform, python version, ...
     validations:
       required: true

.github/ISSUE_TEMPLATE/migration.yml

@@ -6,7 +6,7 @@ body:
     id: system-info
     attributes:
       label: System Info
-      description: Please share your system info with us. You can run the command `transformers-cli env` and copy-paste its output below.
+      description: Please share your system info with us. You can run the command `transformers env` and copy-paste its output below.
       render: shell
       placeholder: transformers version, platform, python version, ...
     validations:

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

@@ -54,7 +54,7 @@ jobs:
       - name: Create model files
         run: |
           . ~/venv/bin/activate
-          transformers-cli add-new-model-like --config_file tests/fixtures/add_distilbert_like_config.json --path_to_repo .
+          transformers add-new-model-like --config_file tests/fixtures/add_distilbert_like_config.json --path_to_repo .
           make style
           make fix-copies
 

.github/workflows/change_pr_to_draft.yml

@@ -1,25 +0,0 @@
-name: Change PR to draft
-
-on:
-  pull_request_target:
-    types: [opened, reopened]
-
-jobs:
-  convert_pr_to_draft:
-    runs-on: ubuntu-22.04
-    name: Convert PR to draft
-    permissions:
-      pull-requests: write
-      contents: write
-    if: github.event.pull_request.draft == false
-    steps:
-      - name: Convert PR to draft
-        shell: bash
-        env:
-          PR_NUMBER: ${{ github.event.number }}
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          REPO: ${{ github.repository }}
-        run: |
-          echo $PR_NUMBER
-          gh pr ready $PR_NUMBER --repo $REPO --undo
-          gh pr comment $PR_NUMBER --repo $REPO --body "Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. The CI will be paused while the PR is in draft mode. When it is ready for review, please click the \`Ready for review\` button (at the bottom of the PR page). This will assign reviewers and trigger CI."

.github/workflows/check_failed_model_tests.yml

@@ -29,7 +29,7 @@ jobs:
   run_models_gpu:
     name: " "
     runs-on:
-      group: aws-g4dn-2xlarge-cache
+      group: aws-g4dn-4xlarge-cache
     container:
       image: ${{ inputs.docker }}
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/

.github/workflows/doctest_job.yml

@@ -28,7 +28,7 @@ jobs:
       matrix:
         split_keys: ${{ fromJson(inputs.split_keys) }}
     runs-on: 
-      group: aws-g4dn-2xlarge-cache
+      group: aws-g4dn-4xlarge-cache
     container:
       image: huggingface/transformers-all-latest-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/

.github/workflows/doctests.yml

@@ -15,7 +15,7 @@ jobs:
   setup:
     name: Setup
     runs-on: 
-      group: aws-g4dn-2xlarge-cache
+      group: aws-g4dn-4xlarge-cache
     container:
       image: huggingface/transformers-all-latest-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/

.github/workflows/model_jobs.yml

@@ -107,7 +107,7 @@ jobs:
         run: |
           echo "${{ inputs.machine_type }}"
 
-          if [ "${{ inputs.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
+          if [ "${{ inputs.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
             machine_type=single-gpu
           elif [ "${{ inputs.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
             machine_type=multi-gpu

.github/workflows/new_model_pr_merged_notification.yml

@@ -59,7 +59,7 @@ jobs:
                   "type": "section",
                   "text": {
                     "type": "mrkdwn",
-                    "text": "<https://github.com/huggingface/transformers/commit/${{ env.COMMIT_SHA }}|New model: ${{ env.NEW_MODEL }}> GH_ArthurZucker, GH_lysandrejik, GH_ydshieh"
+                    "text": "<https://github.com/huggingface/transformers/commit/${{ env.COMMIT_SHA }}|New model: ${{ env.NEW_MODEL }}> GH_ArthurZucker, GH_lysandrejik, GH_ydshieh\ncommit SHA: ${{ env.COMMIT_SHA }}"
                   }
                 }
               ]

.github/workflows/pr-style-bot.yml

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

.github/workflows/self-comment-ci.yml

@@ -29,7 +29,7 @@ jobs:
     runs-on: ubuntu-22.04
     name: Get PR number
     # For security: only allow team members to run
-    if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "qubvel", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "muellerzr", "eustlb", "MekkCyber"]'), github.actor) && (startsWith(github.event.comment.body, 'run-slow') || startsWith(github.event.comment.body, 'run slow') || startsWith(github.event.comment.body, 'run_slow')) }}
+    if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "qubvel", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "muellerzr", "eustlb", "MekkCyber", "manueldeprada"]'), github.actor) && (startsWith(github.event.comment.body, 'run-slow') || startsWith(github.event.comment.body, 'run slow') || startsWith(github.event.comment.body, 'run_slow')) }}
     outputs:
       PR_NUMBER: ${{ steps.set_pr_number.outputs.PR_NUMBER }}
     steps:
@@ -145,7 +145,7 @@ jobs:
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           MODELS: ${{ needs.get-tests.outputs.models }}
-          BODY: "This comment contains run-slow, running the specified jobs:\n\nmodels: ${{ needs.get-tests.outputs.models }}\nquantizations: ${{ needs.get-tests.outputs.quantizations }}"
+          BODY: "\n\nmodels: ${{ needs.get-tests.outputs.models }}\nquantizations: ${{ needs.get-tests.outputs.quantizations }}"
         run: |
           gh api \
             --method POST \
@@ -185,7 +185,7 @@ jobs:
       fail-fast: false
       matrix:
         folders: ${{ fromJson(needs.get-tests.outputs.models) }}
-        machine_type: [aws-g4dn-2xlarge-cache, aws-g4dn-12xlarge-cache]
+        machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
     runs-on:
        group: '${{ matrix.machine_type }}'
     container:
@@ -239,7 +239,7 @@ jobs:
         shell: bash
         run: |
           echo "${{ matrix.machine_type }}"
-          if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
+          if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
             machine_type=single-gpu
           elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
             machine_type=multi-gpu
@@ -292,7 +292,7 @@ jobs:
       fail-fast: false
       matrix:
         folders: ${{ fromJson(needs.get-tests.outputs.quantizations) }}
-        machine_type: [aws-g4dn-2xlarge-cache, aws-g4dn-12xlarge-cache]
+        machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
     runs-on:
       group: '${{ matrix.machine_type }}'
     container:
@@ -338,7 +338,7 @@ jobs:
         shell: bash
         run: |
           echo "${{ matrix.machine_type }}"
-          if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
+          if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
             machine_type=single-gpu
           elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
             machine_type=multi-gpu

.github/workflows/self-scheduled.yml

@@ -49,7 +49,7 @@ jobs:
     name: Setup
     strategy:
       matrix:
-        machine_type: [aws-g4dn-2xlarge-cache, aws-g4dn-12xlarge-cache]
+        machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
     runs-on:
       group: '${{ matrix.machine_type }}'
     container:
@@ -107,7 +107,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        machine_type: [aws-g4dn-2xlarge-cache, aws-g4dn-12xlarge-cache]
+        machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
         slice_id: ${{ fromJSON(needs.setup.outputs.slice_ids) }}
     uses: ./.github/workflows/model_jobs.yml
     with:
@@ -125,7 +125,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        machine_type: [aws-g4dn-2xlarge-cache, aws-g4dn-12xlarge-cache]
+        machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
         slice_id: [0, 1]
     uses: ./.github/workflows/model_jobs.yml
     with:
@@ -143,7 +143,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        machine_type: [aws-g4dn-2xlarge-cache, aws-g4dn-12xlarge-cache]
+        machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
     runs-on:
       group: '${{ matrix.machine_type }}'
     container:
@@ -177,7 +177,7 @@ jobs:
         run: |
           echo "${{ matrix.machine_type }}"
 
-          if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
+          if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
             machine_type=single-gpu
           elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
             machine_type=multi-gpu
@@ -211,7 +211,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        machine_type: [aws-g4dn-2xlarge-cache, aws-g4dn-12xlarge-cache]
+        machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
     runs-on:
       group: '${{ matrix.machine_type }}'
     container:
@@ -246,7 +246,7 @@ jobs:
         run: |
           echo "${{ matrix.machine_type }}"
 
-          if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
+          if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
             machine_type=single-gpu
           elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
             machine_type=multi-gpu
@@ -280,7 +280,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        machine_type: [aws-g4dn-2xlarge-cache]
+        machine_type: [aws-g4dn-4xlarge-cache]
     runs-on:
       group: '${{ matrix.machine_type }}'
     container:
@@ -314,7 +314,7 @@ jobs:
         run: |
           echo "${{ matrix.machine_type }}"
 
-          if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
+          if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
             machine_type=single-gpu
           elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
             machine_type=multi-gpu
@@ -349,7 +349,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        machine_type: [aws-g4dn-2xlarge-cache, aws-g4dn-12xlarge-cache]
+        machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
     runs-on:
       group: '${{ matrix.machine_type }}'
     container:
@@ -411,7 +411,7 @@ jobs:
         run: |
           echo "${{ matrix.machine_type }}"
 
-          if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
+          if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
             machine_type=single-gpu
           elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
             machine_type=multi-gpu
@@ -448,7 +448,7 @@ jobs:
       fail-fast: false
       matrix:
         folders: ${{ fromJson(needs.setup.outputs.quantization_matrix) }}
-        machine_type: [aws-g4dn-2xlarge-cache, aws-g4dn-12xlarge-cache]
+        machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
     runs-on:
       group: '${{ matrix.machine_type }}'
     container:
@@ -491,7 +491,7 @@ jobs:
         run: |
           echo "${{ matrix.machine_type }}"
 
-          if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
+          if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
             machine_type=single-gpu
           elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
             machine_type=multi-gpu

.github/workflows/ssh-runner.yml

@@ -35,7 +35,7 @@ jobs:
         shell: bash
         run: |
           if [[ "${{ github.event.inputs.num_gpus }}" == "single" && "${{ github.event.inputs.runner_type }}" == "t4" ]]; then
-            echo "RUNNER=aws-g4dn-2xlarge-cache" >> $GITHUB_ENV
+            echo "RUNNER=aws-g4dn-4xlarge-cache" >> $GITHUB_ENV
           elif [[ "${{ github.event.inputs.num_gpus }}" == "multi" && "${{ github.event.inputs.runner_type }}" == "t4" ]]; then
             echo "RUNNER=aws-g4dn-12xlarge-cache" >> $GITHUB_ENV
           elif [[ "${{ github.event.inputs.num_gpus }}" == "single" && "${{ github.event.inputs.runner_type }}" == "a10" ]]; then

CONTRIBUTING.md

@@ -78,7 +78,7 @@ Once you've confirmed the bug hasn't already been reported, please include the f
 To get the OS and software versions automatically, run the following command:
 
 ```bash
-transformers-cli env
+transformers env
 ```
 
 You can also run the same command from the root of the repository:

Makefile

@@ -79,7 +79,7 @@ fixup: modified_only_fixup extra_style_checks autogenerate_code repo-consistency
 
 fix-copies:
 	python utils/check_copies.py --fix_and_overwrite
-	python utils/check_modular_conversion.py  --fix_and_overwrite
+	python utils/check_modular_conversion.py --fix_and_overwrite
 	python utils/check_dummies.py --fix_and_overwrite
 	python utils/check_doctest_list.py --fix_and_overwrite
 	python utils/check_docstrings.py --fix_and_overwrite

README.md

@@ -78,7 +78,6 @@ Create and activate a virtual environment with [venv](https://docs.python.org/3/
 # venv
 python -m venv .my-env
 source .my-env/bin/activate
-
 # uv
 uv venv .my-env
 source .my-env/bin/activate
@@ -88,10 +87,10 @@ Install Transformers in your virtual environment.
 
 ```py
 # pip
-pip install transformers
+pip install "transformers[torch]"
 
 # uv
-uv pip install transformers
+uv pip install "transformers[torch]"
 ```
 
 Install Transformers from source if you want the latest changes in the library or are interested in contributing. However, the *latest* version may not be stable. Feel free to open an [issue](https://github.com/huggingface/transformers/issues) if you encounter an error.
@@ -99,7 +98,12 @@ Install Transformers from source if you want the latest changes in the library o
 ```shell
 git clone https://github.com/huggingface/transformers.git
 cd transformers
-pip install .
+
+# pip
+pip install .[torch]
+
+# uv
+uv pip install .[torch]
 ```
 
 ## Quickstart
@@ -121,7 +125,7 @@ To chat with a model, the usage pattern is the same. The only difference is you
 > [!TIP]
 > You can also chat with a model directly from the command line.
 > ```shell
-> transformers-cli chat --model_name_or_path Qwen/Qwen2.5-0.5B-Instruct
+> transformers chat Qwen/Qwen2.5-0.5B-Instruct
 > ```
 
 ```py

docker/transformers-all-latest-gpu/Dockerfile

@@ -71,6 +71,9 @@ RUN python3 -m pip install --no-cache-dir g2p-en
 # For Some bitsandbytes tests
 RUN python3 -m pip install --no-cache-dir einops
 
+# `kernels` may give different outputs (within 1e-5 range) even with the same model (weights) and the same inputs
+RUN python3 -m pip uninstall -y kernels
+
 # 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

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

@@ -45,6 +45,9 @@ RUN python3 -m pip uninstall -y deepspeed
 # TODO: Find out why test fail.
 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
 
+# `kernels` may give different outputs (within 1e-5 range) even with the same model (weights) and the same inputs
+RUN python3 -m pip uninstall -y kernels
+
 # 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

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

@@ -57,6 +57,9 @@ RUN python3 -m pip uninstall -y deepspeed
 #RUN git clone https://github.com/pytorch/TensorRT.git
 #RUN cd TensorRT/py && python3 setup.py install --fx-only
 
+# `kernels` may give different outputs (within 1e-5 range) even with the same model (weights) and the same inputs
+RUN python3 -m pip uninstall -y kernels
+
 # 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

docker/transformers-pytorch-gpu/Dockerfile

@@ -28,6 +28,9 @@ RUN python3 -m pip uninstall -y tensorflow flax
 RUN python3 -m pip install --no-cache-dir git+https://github.com/facebookresearch/detectron2.git pytesseract
 RUN python3 -m pip install -U "itsdangerous<2.1.0"
 
+# `kernels` may give different outputs (within 1e-5 range) even with the same model (weights) and the same inputs
+RUN python3 -m pip uninstall -y kernels
+
 # 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

docker/transformers-quantization-latest-gpu/Dockerfile

@@ -90,6 +90,9 @@ RUN python3 -m pip install --no-cache-dir "auto-round>=0.5.0"
 # Add transformers in editable mode
 RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-torch]
 
+# `kernels` may give different outputs (within 1e-5 range) even with the same model (weights) and the same inputs
+RUN python3 -m pip uninstall -y kernels
+
 # 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

docs/source/de/add_new_model.md

@@ -95,7 +95,7 @@ wie der Code geschrieben werden sollte :-)
 1. Der Vorwärtsdurchlauf Ihres Modells sollte vollständig in die Modellierungsdatei geschrieben werden und dabei völlig unabhängig von anderen
    Modellen in der Bibliothek. Wenn Sie einen Block aus einem anderen Modell wiederverwenden möchten, kopieren Sie den Code und fügen ihn mit einem
    `# Kopiert von` ein (siehe [hier](https://github.com/huggingface/transformers/blob/v4.17.0/src/transformers/models/roberta/modeling_roberta.py#L160)
-   für ein gutes Beispiel und [hier](pr_checks#check-copies) für weitere Dokumentation zu Copied from). 
+   für ein gutes Beispiel und [hier](pr_checks#check-copies) für weitere Dokumentation zu Copied from).
 2. Der Code sollte vollständig verständlich sein, auch für einen Nicht-Muttersprachler. Das heißt, Sie sollten
    beschreibende Variablennamen wählen und Abkürzungen vermeiden. Ein Beispiel: `activation` ist `act` vorzuziehen.
    Von Variablennamen mit nur einem Buchstaben wird dringend abgeraten, es sei denn, es handelt sich um einen Index in einer for-Schleife.
@@ -402,7 +402,7 @@ Andernfalls beginnen wir mit der Erstellung eines neuen Modells. Wir empfehlen d
 ein bestehendes Modell:
 
 ```bash
-transformers-cli add-new-model-like
+transformers add-new-model-like
 ```
 
 Sie werden mit einem Fragebogen aufgefordert, die grundlegenden Informationen Ihres Modells einzugeben.

docs/source/de/contributing.md

@@ -63,7 +63,7 @@ Wenn Sie sich vergewissert haben, dass der Fehler noch nicht gemeldet wurde, geb
 Um das Betriebssystem und die Softwareversionen automatisch auszugeben, führen Sie den folgenden Befehl aus:
 
 ```bash
-transformers-cli env
+transformers env
 ```
 
 Sie können denselben Befehl auch im Hauptverzeichnis des Repositorys ausführen:

docs/source/en/_toctree.yml

@@ -21,6 +21,8 @@
       title: Adding a new model to Transformers
     - local: modular_transformers
       title: Modular Transformers
+    - local: auto_docstring
+      title: Document your models
     - local: task_summary
       title: What 🤗 Transformers can do
     - local: tasks_explained
@@ -37,6 +39,8 @@
       title: Tokenizers
     - local: image_processors
       title: Image processors
+    - local: video_processors
+      title: Video processors
     - local: backbones
       title: Backbones
     - local: feature_extractors
@@ -149,6 +153,8 @@
       title: TPU
     - local: perf_train_special
       title: Apple Silicon
+    - local: perf_train_gaudi
+      title: Intel Gaudi
     - local: perf_hardware
       title: Build your own machine
     title: Hardware
@@ -358,7 +364,9 @@
       title: Feature Extractor
     - local: main_classes/image_processor
       title: Image Processor
-    title: Main classes
+    - local: main_classes/video_processor
+      title: Video Processor
+    title: Main Classes
   - sections:
     - sections:
       - local: model_doc/albert
@@ -493,12 +501,16 @@
         title: Granite
       - local: model_doc/granitemoe
         title: GraniteMoe
+      - local: model_doc/granitemoehybrid
+        title: GraniteMoeHybrid
       - local: model_doc/granitemoeshared
         title: GraniteMoeShared
       - local: model_doc/helium
         title: Helium
       - local: model_doc/herbert
         title: HerBERT
+      - local: model_doc/hgnet_v2
+        title: HGNet-V2
       - local: model_doc/ibert
         title: I-BERT
       - local: model_doc/jamba
@@ -691,6 +703,8 @@
         title: ConvNeXTV2
       - local: model_doc/cvt
         title: CvT
+      - local: model_doc/d_fine
+        title: D-FINE
       - local: model_doc/dab-detr
         title: DAB-DETR
       - local: model_doc/deformable_detr
@@ -817,6 +831,8 @@
         title: Bark
       - local: model_doc/clap
         title: CLAP
+      - local: model_doc/csm
+        title: CSM
       - local: model_doc/dac
         title: dac
       - local: model_doc/encodec
@@ -1017,6 +1033,8 @@
         title: Qwen2VL
       - local: model_doc/sam
         title: Segment Anything
+      - local: model_doc/sam_hq
+        title: Segment Anything High Quality
       - local: model_doc/shieldgemma2
         title: ShieldGemma2
       - local: model_doc/siglip

docs/source/en/add_new_model.md

@@ -161,7 +161,7 @@ The downside is that if you aren't used to them, it may take some time to get us
 Run the command below to start and complete the questionnaire with some basic information about the new model. This command jumpstarts the process by automatically generating some model code that you'll need to adapt.
 
 ```bash
-transformers-cli add-new-model-like
+transformers add-new-model-like
 ```
 
 ## Create a pull request
@@ -292,7 +292,7 @@ Once you're able to run the original checkpoint, you're ready to start adapting
 
 ## Adapt the model code
 
-The `transformers-cli add-new-model-like` command should have generated a model and configuration file.
+The `transformers add-new-model-like` command should have generated a model and configuration file.
 
 - `src/transformers/models/brand_new_llama/modeling_brand_new_llama.py`
 - `src/transformers/models/brand_new_llama/configuration_brand_new_llama.py`
@@ -551,10 +551,10 @@ While this example doesn't include an image processor, you may need to implement
 
 If you do need to implement a new image processor, refer to an existing image processor to understand the expected structure. Slow image processors ([`BaseImageProcessor`]) and fast image processors ([`BaseImageProcessorFast`]) are designed differently, so make sure you follow the correct structure based on the processor type you're implementing.
 
-Run the following command (only if you haven't already created the fast image processor with the `transformers-cli add-new-model-like` command) to generate the necessary imports and to create a prefilled template for the fast image processor. Modify the template to fit your model.
+Run the following command (only if you haven't already created the fast image processor with the `transformers add-new-model-like` command) to generate the necessary imports and to create a prefilled template for the fast image processor. Modify the template to fit your model.
 
 ```bash
-transformers-cli add-fast-image-processor --model-name your_model_name
+transformers add-fast-image-processor --model-name your_model_name
 ```
 
 This command will generate the necessary imports and provide a pre-filled template for the fast image processor. You can then modify it to fit your model's needs.

docs/source/en/attention_interface.md

@@ -108,7 +108,7 @@ If in doubt about what args/kwargs a given model sends to the attention function
 ## Accessing current available implementations
 
 Most of the time, you will simply need to `register` a new function. If, however, you need to access an existing one,
-and/or perform a few checks, the prefered way is to use the global `ALL_ATTENTION_FUNCTIONS`. It behaves the same way you
+and/or perform a few checks, the preferred way is to use the global `ALL_ATTENTION_FUNCTIONS`. It behaves the same way you
 would expect from a usual Python dictionary:
 
 ```python

docs/source/en/auto_docstring.md

@@ -0,0 +1,279 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Utilizing the @auto_docstring Decorator
+
+The `@auto_docstring` decorator in the Hugging Face Transformers library helps generate docstrings for model classes and their methods, which will be used to build the documentation for the library. It aims to improve consistency and reduce boilerplate by automatically including standard argument descriptions and allowing for targeted overrides and additions.
+
+---
+
+## 📜 How it Works
+
+The `@auto_docstring` decorator constructs docstrings by:
+
+1.  **Signature Inspection:** It inspects the signature (arguments, types, defaults) of the decorated class's `__init__` method or the decorated function.
+2.  **Centralized Docstring Fetching:** It retrieves predefined docstrings for common arguments (e.g., `input_ids`, `attention_mask`) from internal library sources (like `ModelArgs` or `ImageProcessorArgs` in `utils/args_doc.py`).
+3.  **Overriding or Adding Arguments Descriptions:**
+    * **Direct Docstring Block:** It incorporates custom docstring content from an `r""" """` (or `""" """`) block below the method signature or within the `__init__` docstring. This is for documenting new arguments or overriding standard descriptions.
+    * **Decorator Arguments (`custom_args`):** A `custom_args` docstring block can be passed to the decorator to provide docstrings for specific arguments directly in the decorator call. This can be used to define the docstring block for new arguments once if they are repeated in multiple places in the modeling file.
+4.  **Adding Classes and Functions Introduction:**
+    * **`custom_intro` argument:** Allows prepending a custom introductory paragraph to a class or function docstring.
+    * **Automatic Introduction Generation:** For model classes with standard naming patterns (like `ModelForCausalLM`) or belonging to a pipeline, the decorator automatically generates an appropriate introductory paragraph using `ClassDocstring` in `utils/args_doc.py` as the source.
+5.  **Templating:** The decorator uses a templating system, allowing predefined docstrings to include dynamic information deduced from the `auto_modules` of the library, such as `{{processor_class}}` or `{{config_class}}`.
+6.  **Deducing Relevant Examples:** The decorator attempts to find appropriate usage examples based on the model's task or pipeline compatibility. It extracts checkpoint information from the model's configuration class to provide concrete examples with real model identifiers.
+7.  **Adding Return Value Documentation:** For methods like `forward`, the decorator can automatically generate the "Returns" section based on the method's return type annotation. For example, for a method returning a `ModelOutput` subclass, it will extracts field descriptions from that class's docstring to create a comprehensive return value description. A custom `Returns` section can also be manually specified in the function docstring block.
+8.  **Unrolling Kwargs Typed With Unpack Operator:** For specific methods (defined in `UNROLL_KWARGS_METHODS`) or classes (defined in `UNROLL_KWARGS_CLASSES`), the decorator processes `**kwargs` parameters that are typed with `Unpack[KwargsTypedDict]`. It extracts the documentation from the TypedDict and adds each parameter to the function's docstring. Currently, this functionality is only supported for `FastImageProcessorKwargs`.
+
+
+---
+
+## 🚀 How to Use @auto_docstring
+
+### 1. Importing the Decorator
+Import the decorator into your modeling file:
+
+```python
+from ...utils import auto_docstring
+```
+
+### 2. Applying to Classes
+Place `@auto_docstring` directly above the class definition. It uses the `__init__` method's signature and its docstring for parameter descriptions.
+
+```python
+from transformers.modeling_utils import PreTrainedModel
+from ...utils import auto_docstring
+
+@auto_docstring
+class MyAwesomeModel(PreTrainedModel):
+    def __init__(self, config, custom_parameter: int = 10, another_custom_arg: str = "default"):
+        r"""
+        custom_parameter (`int`, *optional*, defaults to 10):
+            Description of the custom_parameter for MyAwesomeModel.
+        another_custom_arg (`str`, *optional*, defaults to "default"):
+            Documentation for another unique argument.
+        """
+        super().__init__(config)
+        self.custom_parameter = custom_parameter
+        self.another_custom_arg = another_custom_arg
+        # ... rest of your init
+
+    # ... other methods
+```
+
+#### Advanced Class Decoration:
+
+Arguments can be passed directly to `@auto_docstring` for more control:
+
+```python
+@auto_docstring(
+    custom_intro="""This model performs specific synergistic operations.
+    It builds upon the standard Transformer architecture with unique modifications.""",
+    custom_args="""
+    custom_parameter (`type`, *optional*, defaults to `default_value`):
+        A concise description for custom_parameter if not defined or overriding the description in `args_doc.py`.
+    internal_helper_arg (`type`, *optional*, defaults to `default_value`):
+        A concise description for internal_helper_arg if not defined or overriding the description in `args_doc.py`.
+    """
+)
+class MySpecialModel(PreTrainedModel):
+    def __init__(self, config: ConfigType, custom_parameter: "type" = "default_value", internal_helper_arg=None):
+        # ...
+```
+
+Or:
+
+```python
+@auto_docstring(
+    custom_intro="""This model performs specific synergistic operations.
+    It builds upon the standard Transformer architecture with unique modifications.""",
+)
+class MySpecialModel(PreTrainedModel):
+    def __init__(self, config: ConfigType, custom_parameter: "type" = "default_value", internal_helper_arg=None):
+        r"""
+        custom_parameter (`type`, *optional*, defaults to `default_value`):
+            A concise description for custom_parameter if not defined or overriding the description in `args_doc.py`.
+        internal_helper_arg (`type`, *optional*, defaults to `default_value`):
+            A concise description for internal_helper_arg if not defined or overriding the description in `args_doc.py`.
+        """
+        # ...
+```
+
+### 3. Applying to Functions (e.g., `forward` method)
+Apply the decorator above method definitions, such as the `forward` method.
+
+```python
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        new_custom_argument: Optional[torch.Tensor] = None,
+        arg_documented_in_args_doc: Optional[torch.Tensor] = None,
+        # ... other arguments
+    ) -> Union[Tuple, ModelOutput]: # The description of the return value will automatically be generated from the ModelOutput class docstring.
+        r"""
+        new_custom_argument (`torch.Tensor`, *optional*):
+            Description of this new custom argument and its expected shape or type.
+        """
+        # ...
+```
+
+#### Advanced Function Decoration:
+
+Arguments can be passed directly to `@auto_docstring` for more control. `Returns` and `Examples` sections can also be manually specified:
+
+```python
+MODEL_COMMON_CUSTOM_ARGS = r"""
+    common_arg_1 (`torch.Tensor`, *optional*, defaults to `default_value`):
+        Description of common_arg_1
+    common_arg_2 (`torch.Tensor`, *optional*, defaults to `default_value`):
+        Description of common_arg_2
+    ...
+"""
+
+class MyModel(PreTrainedModel):
+    # ...
+    @auto_docstring(
+        custom_intro="""
+        This is a custom introduction for the function.
+        """
+        custom_args=MODEL_COMMON_CUSTOM_ARGS
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        common_arg_1: Optional[torch.Tensor] = None,
+        common_arg_2: Optional[torch.Tensor] = None,
+        #...
+        function_specific_argument: Optional[torch.Tensor] = None,
+        # ... other arguments
+    ) -> torch.Tensor:
+        r"""
+        function_specific_argument (`torch.Tensor`, *optional*):
+            Description of an argument specific to this function
+
+        Returns:
+            `torch.Tensor`: For a function returning a generic type, a custom "Returns" section can be specified.
+
+        Example:
+
+        (To override the default example with a custom one or to add an example for a model class that does not have a pipeline)
+
+        ```python
+        ...
+        ```
+        """
+        # ...
+```
+
+---
+
+### ✍️ Documenting Arguments: Approach & Priority
+
+1.  **Standard Arguments (e.g., `input_ids`, `attention_mask`, `pixel_values`, `encoder_hidden_states` etc.):**
+    * `@auto_docstring` retrieves descriptions from a central source. Do not redefine these locally if their description and shape are the same as in `args_doc.py`.
+
+2.  **New or Custom Arguments:**
+    * **Primary Method:** Document these within an `r""" """` docstring block following the signature (for functions) or in the `__init__` method's docstring (for class parameters).
+    * **Format:**
+        ```
+        argument_name (`type`, *optional*, defaults to `X`):
+            Description of the argument.
+            Explain its purpose, expected shape/type if complex, and default behavior.
+            This can span multiple lines.
+        ```
+    * Include `type` in backticks.
+    * Add "*optional*" if the argument is not required (has a default value).
+    * Add "defaults to `X`" if it has a default value (no need to specify "defaults to `None`" if the default value is `None`).
+
+3.  **Overriding Standard Arguments:**
+    * If a standard argument behaves differently (e.g., different expected shape, model-specific behavior), provide its complete description in the local `r""" """` docstring. This local definition takes precedence.
+    * The `labels` argument is often customized per model and typically requires a specific docstring.
+
+4.  **Using Decorator Arguments for Overrides or New Arguments (`custom_args`):**
+    * New or custom arguments docstrings can also be passed to `@auto_docstring` as a `custom_args` argument. This can be used to define the docstring block for new arguments once if they are repeated in multiple places in the modeling file.
+
+---
+
+### Usage with [modular files](./modular_transformers)
+
+When working with modular files, follow these guidelines for applying the `@auto_docstring` decorator:
+
+- **For standalone models in modular files:**
+  Apply the `@auto_docstring` decorator just as you would in regular modeling files.
+
+- **For models inheriting from other library models:**
+  - When inheriting from a parent model, decorators (including `@auto_docstring`) are automatically carried over to the generated modeling file without needing to add them in your modular file.
+  - If you need to modify the `@auto_docstring` behavior, apply the customized decorator in your modular file, making sure to *include all other decorators* that were present on the original function/class.
+
+  > **Warning**: When overriding any decorator in a modular file, you must include ALL decorators that were applied to that function/class in the parent model. If you only override some decorators, the others won't be included in the generated modeling file.
+
+
+**Note**: The `check_auto_docstrings` tool doesn't check modular files directly, but it will check (and modify when using `--fix_and_overwrite`) the generated modeling files. If issues are found in the generated files, you'll need to update your modular files accordingly.
+
+---
+
+## ✅ Checking Your Docstrings with `check_auto_docstrings`
+
+The library includes a utility script to validate docstrings. This check is typically run during Continuous Integration (CI).
+
+#### What it Checks:
+
+* **Decorator Presence:** Ensures `@auto_docstring` is applied to relevant model classes and public methods. (TODO)
+* **Argument Completeness & Consistency:**
+    * Flags arguments in the signature that are not known standard arguments and lack a local description.
+    * Ensures documented arguments exist in the signature. (TODO)
+    * Verifies that types and default values in the docstring match the signature. (TODO)
+* **Placeholder Detection:** Reminds you to complete placeholders like `<fill_type>` or `<fill_docstring>`.
+* **Formatting:** Adherence to the expected docstring style.
+
+#### Running the Check Locally:
+
+Run this check locally before committing. The common command is:
+
+```bash
+make fix-copies
+```
+
+Alternatively, to only perform docstrings and auto-docstring checks, you can use:
+
+```bash
+python utils/check_docstrings.py # to only check files included in the diff without fixing them
+# Or: python utils/check_docstrings.py --fix_and_overwrite # to fix and overwrite the files in the diff
+# Or: python utils/check_docstrings.py --fix_and_overwrite --check_all # to fix and overwrite all files
+```
+
+#### Workflow with the Checker:
+
+1.  Add `@auto_docstring(...)` to the class or method.
+2.  For new, custom, or overridden arguments, add descriptions in an `r""" """` block.
+3.  Run `make fix-copies` (or the `check_docstrings.py` utility).
+    * For unrecognized arguments lacking documentation, the utility will create placeholder entries.
+4.  Manually edit these placeholders with accurate types and descriptions.
+5.  Re-run the check to ensure all issues are resolved.
+
+---
+
+## 🔑 Key Takeaways & Best Practices
+
+* Use `@auto_docstring` for new PyTorch model classes (`PreTrainedModel` subclasses) and their primary for methods (e.g., `forward`, `get_text_features` etc.).
+* For classes, the `__init__` method's docstring is the main source for parameter descriptions when using `@auto_docstring` on the class.
+* Rely on standard docstrings; do not redefine common arguments unless their behavior is different in your specific model.
+* Document new or custom arguments clearly.
+* Run `check_docstrings` locally and iteratively.
+
+By following these guidelines, you help maintain consistent and informative documentation for the Hugging Face Transformers library 🤗.

docs/source/en/conversations.md

@@ -25,22 +25,28 @@ Check model leaderboards like [OpenLLM](https://hf.co/spaces/HuggingFaceH4/open_
 
 This guide shows you how to quickly start chatting with Transformers from the command line, how build and format a conversation, and how to chat using the [`TextGenerationPipeline`].
 
-## transformers-cli
+## transformers CLI
 
-Chat with a model directly from the command line as shown below. It launches an interactive session with a model. Enter `clear` to reset the conversation, `exit` to terminate the session, and `help` to display all the command options.
+After you've [installed Transformers](./installation.md), chat with a model directly from the command line as shown below. It launches an interactive session with a model, with a few base commands listed at the start of the session.
 
 ```bash
-transformers-cli chat --model_name_or_path Qwen/Qwen2.5-0.5B-Instruct
+transformers chat Qwen/Qwen2.5-0.5B-Instruct
 ```
 
 <div class="flex justify-center">
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers-chat-cli.png"/>
 </div>
 
+You can launch the CLI with arbitrary `generate` flags, with the format `arg_1=value_1 arg_2=value_2 ...`
+
+```bash
+transformers chat Qwen/Qwen2.5-0.5B-Instruct do_sample=False max_new_tokens=10
+```
+
 For a full list of options, run the command below.
 
 ```bash
-transformers-cli chat -h
+transformers chat -h
 ```
 
 The chat is implemented on top of the [AutoClass](./model_doc/auto), using tooling from [text generation](./llm_tutorial) and [chat](./chat_templating).
@@ -76,16 +82,16 @@ print(response[0]["generated_text"][-1]["content"])
 (sigh) Oh boy, you're asking me for advice? You're gonna need a map, pal! Alright,
 alright, I'll give you the lowdown. But don't say I didn't warn you, I'm a robot, not a tour guide!
 
-So, you wanna know what's fun to do in the Big Apple? Well, let me tell you, there's a million 
-things to do, but I'll give you the highlights. First off, you gotta see the sights: the Statue of 
-Liberty, Central Park, Times Square... you know, the usual tourist traps. But if you're lookin' for 
-something a little more... unusual, I'd recommend checkin' out the Museum of Modern Art. It's got 
+So, you wanna know what's fun to do in the Big Apple? Well, let me tell you, there's a million
+things to do, but I'll give you the highlights. First off, you gotta see the sights: the Statue of
+Liberty, Central Park, Times Square... you know, the usual tourist traps. But if you're lookin' for
+something a little more... unusual, I'd recommend checkin' out the Museum of Modern Art. It's got
 some wild stuff, like that Warhol guy's soup cans and all that jazz.
 
-And if you're feelin' adventurous, take a walk across the Brooklyn Bridge. Just watch out for 
+And if you're feelin' adventurous, take a walk across the Brooklyn Bridge. Just watch out for
 those pesky pigeons, they're like little feathered thieves! (laughs) Get it? Thieves? Ah, never mind.
 
-Now, if you're lookin' for some serious fun, hit up the comedy clubs in Greenwich Village. You might 
+Now, if you're lookin' for some serious fun, hit up the comedy clubs in Greenwich Village. You might
 even catch a glimpse of some up-and-coming comedians... or a bunch of wannabes tryin' to make it big. (winks)
 
 And finally, if you're feelin' like a real New Yorker, grab a slice of pizza from one of the many amazing
@@ -107,9 +113,9 @@ print(response[0]["generated_text"][-1]["content"])
 ```
 
 ```txt
-(laughs) Oh, you're killin' me, pal! You don't get it, do you? Warhol's soup cans are like, art, man! 
-It's like, he took something totally mundane, like a can of soup, and turned it into a masterpiece. It's 
-like, "Hey, look at me, I'm a can of soup, but I'm also a work of art!" 
+(laughs) Oh, you're killin' me, pal! You don't get it, do you? Warhol's soup cans are like, art, man!
+It's like, he took something totally mundane, like a can of soup, and turned it into a masterpiece. It's
+like, "Hey, look at me, I'm a can of soup, but I'm also a work of art!"
 (sarcastically) Oh, yeah, real original, Andy.
 
 But, you know, back in the '60s, it was like, a big deal. People were all about challenging the

docs/source/en/generation_strategies.md

@@ -20,11 +20,15 @@ A decoding strategy informs how a model should select the next generated token.
 
 This guide will help you understand the different decoding strategies available in Transformers and how and when to use them.
 
-## Greedy search
+## Basic decoding methods
 
-Greedy search is the default decoding strategy. It selects the next most likely token at each step. Unless specified in [`GenerationConfig`], this strategy generates a maximum of 20 tokens.
+These are well established decoding methods, and should be your starting point for text generation tasks.
 
-Greedy search works well for tasks with relatively short outputs. However, it breaks down when generating longer sequences because it begins to repeat itself.
+### Greedy search
+
+Greedy search is the default decoding strategy. It selects the next most likely token at each step. Unless specified in [`GenerationConfig`], this strategy generates a maximum of 20 new tokens.
+
+Greedy search works well for tasks with relatively short outputs where creativity is not a priority. However, it breaks down when generating longer sequences because it begins to repeat itself.
 
 ```py
 import torch
@@ -40,11 +44,11 @@ tokenizer.batch_decode(outputs, skip_special_tokens=True)
 'Hugging Face is an open-source company that provides a suite of tools and services for building, deploying, and maintaining natural language processing'
 ```
 
-## Contrastive search
+### Sampling
 
-[Contrastive search](https://huggingface.co/papers/2202.06417) is a decoding strategy that aims to reduce repetition even while generating longer sequences. This strategy compares how similar a generated token is against previous tokens, and if they're more similar, a penalty is applied.
+Sampling, or multinomial sampling, randomly selects a token based on the probability distribution over the entire model's vocabulary (as opposed to the most likely token, as in greedy search). This means every token with a non-zero probability has a chance to be selected. Sampling strategies reduce repetition and can generate more creative and diverse outputs.
 
-Enable contrastive search with the `penalty_alpha` and `top_k` parameters. The `penalty_alpha` manages the penalty applied and `top_k` is the number of most likely tokens to return.
+Enable multinomial sampling with `do_sample=True` and `num_beams=1`.
 
 ```py
 import torch
@@ -55,14 +59,14 @@ inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt"
 
 model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.float16).to("cuda")
 # explicitly set to 100 because Llama2 generation length is 4096
-outputs = model.generate(**inputs, max_new_tokens=100, penalty_alpha=0.6, top_k=4)
+outputs = model.generate(**inputs, max_new_tokens=50, do_sample=True, num_beams=1)
 tokenizer.batch_decode(outputs, skip_special_tokens=True)
-'Hugging Face is an open-source company that provides a platform for building and deploying AI models.\nHugging Face is an open-source company that provides a platform for building and deploying AI models. The platform allows developers to build and deploy AI models, as well as collaborate with other developers.\nHugging Face was founded in 2019 by Thibault Wittemberg and Clément Delangue. The company is based in Paris, France.\nHugging Face has'
+'Hugging Face is an open-source company 🤗\nWe are open-source and believe that open-source is the best way to build technology. Our mission is to make AI accessible to everyone, and we believe that open-source is the best way to achieve that.'
 ```
 
-## Beam search
+### Beam search
 
-Beam search keeps track of several generated sequences (beams) at each time step. After a certain number of steps, it selects the sequence with the highest *overall* probability. Unlike greedy search, this strategy can "look ahead" and pick a sequence with a higher probability overall even if the initial tokens have a lower probability.
+Beam search keeps track of several generated sequences (beams) at each time step. After a certain number of steps, it selects the sequence with the highest *overall* probability. Unlike greedy search, this strategy can "look ahead" and pick a sequence with a higher probability overall even if the initial tokens have a lower probability. It is best suited for input-grounded tasks, like describing an image or speech recognition. You can also use `do_sample=True` with beam search to sample at each step, but beam search will still greedily prune out low probability sequences between steps.
 
 > [!TIP]
 > Check out the [beam search visualizer](https://huggingface.co/spaces/m-ric/beam_search_visualizer) to see how beam search works.
@@ -83,66 +87,11 @@ tokenizer.batch_decode(outputs, skip_special_tokens=True)
 "['Hugging Face is an open-source company that develops and maintains the Hugging Face platform, which is a collection of tools and libraries for building and deploying natural language processing (NLP) models. Hugging Face was founded in 2018 by Thomas Wolf']"
 ```
 
-## Diverse beam search
+## Advanced decoding methods
 
-[Diverse beam search](https://hf.co/papers/1610.02424) is a variant of beam search that produces more diverse output candidates to choose from. This strategy measures the dissimilarity of sequences and a penalty is applied if sequences are too similar. To avoid high computation costs, the number of beams is divided into groups.
+Advanced decoding methods aim at either tackling specific generation quality issues (e.g. repetition) or at improving the generation throughput in certain situations. These techniques are more complex, and may not work correctly with all models.
 
-Enable diverse beam search with the `num_beams`, `num_beam_groups` and `diversity_penalty` parameters (the `num_beams` parameter should be divisible by `num_beam_groups`).
-
-```py
-import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
-inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to("cuda")
-
-model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.float16).to("cuda")
-# explicitly set to 100 because Llama2 generation length is 4096
-outputs = model.generate(**inputs, max_new_tokens=50, num_beams=6, num_beam_groups=3, diversity_penalty=1.0, do_sample=False)
-tokenizer.batch_decode(outputs, skip_special_tokens=True)
-'Hugging Face is an open-source company 🤗\nWe are an open-source company. Our mission is to democratize AI and make it accessible to everyone. We believe that AI should be used for the benefit of humanity, not for the benefit of a'
-```
-
-## Multinomial sampling
-
-Search methods selects the most likely tokens. Sampling, or multinomial sampling, randomly selects a token based on the probability distribution over the entire models vocabulary. This means every token with a non-zero probability has a chance to be selected. Sampling strategies reduce repetition and can generate more creative and diverse outputs.
-
-Enable multinomial sampling with `do_sample=True` and `num_beams=1`.
-
-```py
-import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
-inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to("cuda")
-
-model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.float16).to("cuda")
-# explicitly set to 100 because Llama2 generation length is 4096
-outputs = model.generate(**inputs, max_new_tokens=50, do_sample=True, num_beams=1)
-tokenizer.batch_decode(outputs, skip_special_tokens=True)
-'Hugging Face is an open-source company 🤗\nWe are open-source and believe that open-source is the best way to build technology. Our mission is to make AI accessible to everyone, and we believe that open-source is the best way to achieve that.'
-```
-
-## Beam search multinomial sampling
-
-This decoding strategy is a combination of beam search and multinomial sampling. It generates multiple beams and uses a sampling strategy for each beam.
-
-Enable beam search multinomial sampling by setting `num_beams` to a value greater than 1 and `do_sample=True`.
-
-```py
-import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
-inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to("cuda")
-
-model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.float16).to("cuda")
-# explicitly set to 100 because Llama2 generation length is 4096
-outputs = model.generate(**inputs, max_new_tokens=50, do_sample=True, num_beams=4)
-'Hugging Face is an open-source company 100% dedicated to making AI more accessible. We believe that AI should be available to everyone, and we’re working hard to make that a reality.\nWe’re a team of passionate engineers, designers,'
-```
-
-## Speculative decoding
+### Speculative decoding
 
 [Speculative](https://hf.co/papers/2211.17192) or assistive decoding isn't a search or sampling strategy. Instead, speculative decoding adds a second smaller model to generate candidate tokens. The main model verifies the candidate tokens in a single `forward` pass, which speeds up the decoding process overall. This method is especially useful for LLMs where it can be more costly and slower to generate tokens. Refer to the [speculative decoding](./llm_optims#speculative-decoding) guide to learn more.
 
@@ -203,7 +152,7 @@ tokenizer.batch_decode(outputs, skip_special_tokens=True)
 </hfoption>
 </hfoptions>
 
-### Prompt lookup decoding
+#### Prompt lookup decoding
 
 [Prompt lookup decoding](./llm_optims#prompt-lookup-decoding) is a variant of speculative decoding that uses overlapping n-grams as the candidate tokens. It works well for input-grounded tasks such as summarization. Refer to the [prompt lookup decoding](./llm_optims#prompt-lookup-decoding) guide to learn more.
 
@@ -245,7 +194,7 @@ outputs = model.generate(**inputs, assistant_early_exit=4, do_sample=False, max_
 tokenizer.batch_decode(outputs, skip_special_tokens=True)
 ```
 
-### Universal assisted decoding
+#### Universal assisted decoding
 
 Universal assisted decoding (UAD) enables the main and assistant models to use different tokenizers. The main models input tokens are re-encoded into assistant model tokens. Candidate tokens are generated in the assistant encoding which are re-encoded into the main model candidate tokens. The candidate tokens are verified as explained in [speculative decoding](#speculative-decoding).
 
@@ -269,7 +218,27 @@ tokenizer.batch_decode(outputs, skip_special_tokens=True)
 ['Alice and Bob are sitting in a bar. Alice is drinking a beer and Bob is drinking a']
 ```
 
-## DoLa
+### Contrastive search
+
+[Contrastive search](https://huggingface.co/papers/2202.06417) is a decoding strategy that aims to reduce repetition even while generating longer sequences. This strategy compares how similar a generated token is against previous tokens, and if they're more similar, a penalty is applied.
+
+Enable contrastive search with the `penalty_alpha` and `top_k` parameters. The `penalty_alpha` manages the penalty applied and `top_k` is the number of most likely tokens to return.
+
+```py
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
+inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to("cuda")
+
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.float16).to("cuda")
+# explicitly set to 100 because Llama2 generation length is 4096
+outputs = model.generate(**inputs, max_new_tokens=100, penalty_alpha=0.6, top_k=4)
+tokenizer.batch_decode(outputs, skip_special_tokens=True)
+'Hugging Face is an open-source company that provides a platform for building and deploying AI models.\nHugging Face is an open-source company that provides a platform for building and deploying AI models. The platform allows developers to build and deploy AI models, as well as collaborate with other developers.\nHugging Face was founded in 2019 by Thibault Wittemberg and Clément Delangue. The company is based in Paris, France.\nHugging Face has'
+```
+
+### DoLa
 
 [Decoding by Contrasting Layers (DoLa)](https://hf.co/papers/2309.03883) is a contrastive decoding strategy for improving factuality and reducing hallucination. This strategy works by contrasting the logit differences between the final and early layers. As a result, factual knowledge localized to particular layers are amplified. DoLa is not recommended for smaller models like GPT-2.
 
@@ -325,6 +294,210 @@ tokenizer.batch_decode(outputs[:, inputs.input_ids.shape[-1]:], skip_special_tok
 </hfoption>
 </hfoptions>
 
+### Diverse beam search
+
+[Diverse beam search](https://hf.co/papers/1610.02424) is a variant of beam search that produces more diverse output candidates to choose from. This strategy measures the dissimilarity of sequences and a penalty is applied if sequences are too similar. To avoid high computation costs, the number of beams is divided into groups.
+
+Enable diverse beam search with the `num_beams`, `num_beam_groups` and `diversity_penalty` parameters (the `num_beams` parameter should be divisible by `num_beam_groups`).
+
+```py
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
+inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to("cuda")
+
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.float16).to("cuda")
+# explicitly set to 100 because Llama2 generation length is 4096
+outputs = model.generate(**inputs, max_new_tokens=50, num_beams=6, num_beam_groups=3, diversity_penalty=1.0, do_sample=False)
+tokenizer.batch_decode(outputs, skip_special_tokens=True)
+'Hugging Face is an open-source company 🤗\nWe are an open-source company. Our mission is to democratize AI and make it accessible to everyone. We believe that AI should be used for the benefit of humanity, not for the benefit of a'
+```
+
+
+## Custom decoding methods
+
+Custom decoding methods enable specialized generation behavior such as the following:
+- have the model continue thinking if it is uncertain;
+- roll back generation if the model gets stuck;
+- handle special tokens with custom logic;
+- enhanced input preparation for advanced models;
+
+We enable custom decoding methods through model repositories, assuming a specific model tag and file structure (see subsection below). This feature is an extension of [custom modeling code](./models.md#custom-models) and, like such, requires setting `trust_remote_code=True`.
+
+If a model repository holds a custom decoding method, the easiest way to try it out is to load the model and generate with it:
+
+<!-- TODO before merging: 1) better repo name (use a `generate-community` org?) 2) prettify the repo -->
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+# `transformers-community/custom_generate_example` holds a copy of `Qwen/Qwen2.5-0.5B-Instruct`, but
+# with custom generation code -> calling `generate` uses the custom decoding method!
+tokenizer = AutoTokenizer.from_pretrained("transformers-community/custom_generate_example")
+model = AutoModelForCausalLM.from_pretrained(
+    "transformers-community/custom_generate_example", device_map="auto", trust_remote_code=True
+)
+
+inputs = tokenizer(["The quick brown"], return_tensors="pt").to(model.device)
+# The custom decoding method is a minimal greedy decoding implementation. It also prints a custom message at run time.
+gen_out = model.generate(**inputs)
+# you should now see its custom message, "✨ using a custom generation method ✨"
+print(tokenizer.batch_decode(gen_out, skip_special_tokens=True))
+'The quick brown fox jumps over a lazy dog, and the dog is a type of animal. Is'
+```
+
+Model repositories with custom decoding methods have a special property: their decoding method can be loaded from **any** model through [`~GenerationMixin.generate`]'s `custom_generate` argument. This means anyone can create and share their custom generation method to potentially work with any Transformers model, without requiring users to install additional Python packages.
+
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
+model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct", device_map="auto")
+
+inputs = tokenizer(["The quick brown"], return_tensors="pt").to(model.device)
+# `custom_generate` replaces the original `generate` by the custom decoding method defined in
+# `transformers-community/custom_generate_example`
+gen_out = model.generate(**inputs, custom_generate="transformers-community/custom_generate_example", trust_remote_code=True)
+print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
+'The quick brown fox jumps over a lazy dog, and the dog is a type of animal. Is'
+```
+
+You should read the `README.md` file of the repository containing the custom generation strategy to see what the new arguments and output type differences are, if they exist. Otherwise, you can assume it works like the base [`~GenerationMixin.generate`] method.
+
+> [!TIP]
+> You can find all custom decoding methods by [searching for their custom tag.](https://huggingface.co/models?other=custom_generate), `custom_generate`
+
+Consider the Hub repository [transformers-community/custom_generate_example](https://huggingface.co/transformers-community/custom_generate_example) as an example. The `README.md` states that it has an additional input argument, `left_padding`, which adds a number of padding tokens before the prompt.
+
+```py
+gen_out = model.generate(
+    **inputs, custom_generate="transformers-community/custom_generate_example", trust_remote_code=True, left_padding=5
+)
+print(tokenizer.batch_decode(gen_out)[0])
+'<|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|>The quick brown fox jumps over the lazy dog.\n\nThe sentence "The quick'
+```
+
+If the custom method has pinned Python requirements that your environment doesn't meet, you'll get an exception about missing requirements. For instance, [transformers-community/custom_generate_bad_requirements](https://huggingface.co/transformers-community/custom_generate_bad_requirements) has an impossible set of requirements defined in its `custom_generate/requirements.txt` file, and you'll see the error message below if you try to run it.
+
+```
+ImportError: Missing requirements in your local environment for `transformers-community/custom_generate_bad_requirements`:
+foo (installed: None)
+bar==0.0.0 (installed: None)
+torch>=99.0 (installed: 2.6.0)
+```
+
+Updating your Python requirements accordingly will remove this error message.
+
+### Creating a custom decoding method
+
+To create a new decoding method, you need to create a new [**Model**](https://huggingface.co/new) repository and push a few files into it.
+1. The model you've designed your decoding method with.
+2. `custom_generate/generate.py`, which contains all the logic for your custom decoding method.
+3. `custom_generate/requirements.txt`, used to optionally add new Python requirements and/or lock specific versions to correctly use your method.
+4. `README.md`, where you should add the `custom_generate` tag and document any new arguments or output type differences of your custom method here.
+
+After you've added all required files, your repository should look like this
+
+```
+your_repo/
+├── README.md          # include the 'custom_generate' tag
+├── config.json
+├── ...
+└── custom_generate/
+    ├── generate.py
+    └── requirements.txt
+```
+
+#### Adding the base model
+
+The starting point for your custom decoding method is a model repository just like any other. The model to add to this repository should be the model you've designed your method with, and it is meant to be part of a working self-contained model-generate pair. When the model in this repository is loaded, your custom decoding method will override `generate`. Don't worry -- your decoding method can still be loaded with any other Transformers model, as explained in the section above.
+
+If you simply want to copy an existing model, you can do
+
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("source/model_repo")
+model = AutoModelForCausalLM.from_pretrained("source/model_repo")
+tokenizer.save_pretrained("your/decoding_method", push_to_hub=True)
+model.save_pretrained("your/decoding_method", push_to_hub=True)
+```
+
+#### generate.py
+
+This is the core of your decoding method. It *must* contain a method named `generate`, and this method *must* contain a `model` argument as its first argument. `model` is the model instance, which means you have access to all attributes and methods in the model, including the ones defined in [`GenerationMixin`] (like the base `generate` method).
+
+> [!WARNING]
+> `generate.py` must be placed in a folder named `custom_generate`, and not at the root level of the repository. The file paths for this feature are hardcoded.
+
+Under the hood, when the base [`~GenerationMixin.generate`] method is called with a `custom_generate` argument, it first checks its Python requirements (if any), then locates the custom `generate` method in `generate.py`, and finally calls the custom `generate`. All received arguments and `model` are forwarded to your custom `generate` method.
+
+This means your `generate` can have a mix of original and custom arguments (as well as a different output type) as shown below.
+
+```py
+import torch
+
+def generate(model, input_ids, generation_config=None, left_padding=None, **kwargs):
+    generation_config = generation_config or model.generation_config  # default to the model generation config
+    cur_length = input_ids.shape[1]
+    max_length = generation_config.max_length or cur_length + generation_config.max_new_tokens
+
+    # Example of custom argument: add `left_padding` (integer) pad tokens before the prompt
+    if left_padding is not None:
+        if not isinstance(left_padding, int) or left_padding < 0:
+            raise ValueError(f"left_padding must be an integer larger than 0, but is {left_padding}")
+
+        pad_token = kwargs.pop("pad_token", None) or generation_config.pad_token_id or model.config.pad_token_id
+        if pad_token is None:
+            raise ValueError("pad_token is not defined")
+        batch_size = input_ids.shape[0]
+        pad_tensor = torch.full(size=(batch_size, left_padding), fill_value=pad_token).to(input_ids.device)
+        input_ids = torch.cat((pad_tensor, input_ids), dim=1)
+        cur_length = input_ids.shape[1]
+
+    # Simple greedy decoding loop
+    while cur_length < max_length:
+        logits = model(input_ids).logits
+        next_token_logits = logits[:, -1, :]
+        next_tokens = torch.argmax(next_token_logits, dim=-1)
+        input_ids = torch.cat((input_ids, next_tokens[:, None]), dim=-1)
+        cur_length += 1
+
+    return input_ids
+```
+
+Follow the recommended practices below to ensure your custom decoding method works as expected.
+- Feel free to reuse the logic for validation and input preparation in the original [`~GenerationMixin.generate`].
+- Pin the `transformers` version in the requirements if you use any private method/attribute in `model`.
+- You can add other files in the `custom_generate` folder, and use relative imports.
+- Consider adding model validation, input validation, or even a separate test file to help users sanity-check your code in their environment.
+
+#### requirements.txt
+
+You can optionally specify additional Python requirements in a `requirements.txt` file inside the `custom_generate` folder. These are checked at runtime and an exception will be thrown if they're missing, nudging users to update their environment accordingly.
+
+#### README.md
+
+The root level `README.md` in the model repository usually describes the model therein. However, since the focus of the repository is the custom decoding method, we highly recommend to shift its focus towards describing the custom decoding method. In addition to a description of the method, we recommend documenting any input and/or output differences to the original [`~GenerationMixin.generate`]. This way, users can focus on what's new, and rely on Transformers docs for generic implementation details.
+
+For discoverability, we highly recommend you to add the `custom_generate` tag to your repository. To do so, the top of your `README.md` file should look like the example below. After you push the file, you should see the tag in your repository!
+
+```
+---
+library_name: transformers
+tags:
+  - custom_generate
+---
+
+(your markdown content here)
+```
+
+Recommended practices:
+- Document input and output differences in [`~GenerationMixin.generate`].
+- Add self-contained examples to enable quick experimentation.
+- Describe soft-requirements such as if the method only works well with a certain family of models.
+
+
 ## Resources
 
 Read the [How to generate text: using different decoding methods for language generation with Transformers](https://huggingface.co/blog/how-to-generate) blog post for an explanation of how common decoding strategies work.

docs/source/en/how_to_hack_models.md

@@ -90,11 +90,6 @@ class SamVisionAttentionSplit(SamVisionAttention, nn.Module):
 
         attn_weights = (query * self.scale) @ key.transpose(-2, -1)
 
-        if self.use_rel_pos:
-            attn_weights = self.add_decomposed_rel_pos(
-                attn_weights, query, self.rel_pos_h, self.rel_pos_w, (height, width), (height, width)
-            )
-
         attn_weights = torch.nn.functional.softmax(attn_weights, dtype=torch.float32, dim=-1).to(query.dtype)
         attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
         attn_output = (attn_probs @ value).reshape(batch_size, self.num_attention_heads, height, width, -1)
@@ -114,13 +109,14 @@ Load the model with [`~PreTrainedModel.from_pretrained`].
 
 ```py
 from transformers import SamModel
-from transformers.models.sam import modeling_sam
-
-# replace the attention class in the modeling_sam module
-modeling_sam.SamVisionAttention = SamVisionAttentionSplit
 
 # load the pretrained SAM model
 model = SamModel.from_pretrained("facebook/sam-vit-base")
+
+# replace the attention class in the vision_encoder module
+for layer in model.vision_encoder.layers:
+    if hasattr(layer, "attn"):
+        layer.attn = SamVisionAttentionSplit(model.config.vision_config, model.config.vision_config.window_size)
 ```
 
 ## LoRA
@@ -138,7 +134,7 @@ config = LoraConfig(
     # apply LoRA to q and v
     target_modules=["q", "v"],
     lora_dropout=0.1,
-    task_type="mask-generation"
+    task_type="FEATURE_EXTRACTION"
 )
 ```
 
@@ -152,5 +148,5 @@ Call [print_trainable_parameters](https://huggingface.co/docs/peft/package_refer
 
 ```py
 model.print_trainable_parameters()
-"trainable params: 608,256 || all params: 94,343,728 || trainable%: 0.6447"
+"trainable params: 589,824 || all params: 94,274,096 || trainable%: 0.6256"
 ```

docs/source/en/image_processors.md

@@ -16,7 +16,7 @@ rendered properly in your Markdown viewer.
 
 # Image processors
 
-Image processors converts images into pixel values, tensors that represent image colors and size. The pixel values are inputs to a vision or video model. To ensure a pretrained model receives the correct input, an image processor can perform the following operations to make sure an image is exactly like the images a model was pretrained on.
+Image processors converts images into pixel values, tensors that represent image colors and size. The pixel values are inputs to a vision model. To ensure a pretrained model receives the correct input, an image processor can perform the following operations to make sure an image is exactly like the images a model was pretrained on.
 
 - [`~BaseImageProcessor.center_crop`] to resize an image
 - [`~BaseImageProcessor.normalize`] or [`~BaseImageProcessor.rescale`] pixel values

docs/source/en/internal/import_utils.md

@@ -84,6 +84,19 @@ class Trainer:
 
 Backends that can be added here are all the backends that are available in the `import_utils.py` module.
 
+Additionally, specific versions can be specified in each backend. For example, this is how you would specify
+a requirement on torch>=2.6 on the `Trainer` class:
+
+```python
+from .utils.import_utils import requires
+
+@requires(backends=("torch>=2.6", "accelerate"))
+class Trainer:
+    ...
+```
+
+You can specify the following operators: `==`, `>`, `>=`, `<`, `<=`, `!=`.
+
 ## Methods
 
 [[autodoc]] utils.import_utils.define_import_structure

docs/source/en/llm_tutorial.md

@@ -20,9 +20,13 @@ rendered properly in your Markdown viewer.
 
 Text generation is the most popular application for large language models (LLMs). A LLM is trained to generate the next word (token) given some initial text (prompt) along with its own generated outputs up to a predefined length or when it reaches an end-of-sequence (`EOS`) token.
 
-In Transformers, the [`~GenerationMixin.generate`] API handles text generation, and it is available for all models with generative capabilities.
+In Transformers, the [`~GenerationMixin.generate`] API handles text generation, and it is available for all models with generative capabilities. This guide will show you the basics of text generation with [`~GenerationMixin.generate`] and some common pitfalls to avoid.
 
-This guide will show you the basics of text generation with [`~GenerationMixin.generate`] and some common pitfalls to avoid.
+> [!TIP]
+> You can also chat with a model directly from the command line. ([reference](./conversations.md#transformers-cli))
+> ```shell
+> transformers chat Qwen/Qwen2.5-0.5B-Instruct
+> ```
 
 ## Default generate
 
@@ -134,6 +138,20 @@ outputs = model.generate(**inputs, generation_config=generation_config)
 print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
 ```
 
+## Common Options
+
+[`~GenerationMixin.generate`] is a powerful tool that can be heavily customized. This can be daunting for a new users. This section contains a list of popular generation options that you can define in most text generation tools in Transformers: [`~GenerationMixin.generate`], [`GenerationConfig`], `pipelines`, the `chat` CLI, ...
+
+| Option name | Type | Simplified description |
+|---|---|---|
+| `max_new_tokens` | `int` | Controls the maximum generation length. Be sure to define it, as it usually defaults to a small value. |
+| `do_sample` | `bool` | Defines whether generation will sample the next token (`True`), or is greedy instead (`False`). Most use cases should set this flag to `True`. Check [this guide](./generation_strategies.md) for more information. |
+| `temperature` | `float` | How unpredictable the next selected token will be. High values (`>0.8`) are good for creative tasks, low values (e.g. `<0.4`) for tasks that require "thinking". Requires `do_sample=True`. |
+| `num_beams` | `int` | When set to `>1`, activates the beam search algorithm. Beam search is good on input-grounded tasks. Check [this guide](./generation_strategies.md) for more information. |
+| `repetition_penalty` | `float` | Set it to `>1.0` if you're seeing the model repeat itself often. Larger values apply a larger penalty. |
+| `eos_token_id` | `List[int]` | The token(s) that will cause generation to stop. The default value is usually good, but you can specify a different token. |
+
+
 ## Pitfalls
 
 The section below covers some common issues you may encounter during text generation and how to solve them.
@@ -286,4 +304,4 @@ Take a look below for some more specific and specialized text generation librari
 - [SynCode](https://github.com/uiuc-focal-lab/syncode): a library for context-free grammar guided generation (JSON, SQL, Python).
 - [Text Generation Inference](https://github.com/huggingface/text-generation-inference): a production-ready server for LLMs.
 - [Text generation web UI](https://github.com/oobabooga/text-generation-webui): a Gradio web UI for text generation.
-- [logits-processor-zoo](https://github.com/NVIDIA/logits-processor-zoo): additional logits processors for controlling text generation.
+- [logits-processor-zoo](https://github.com/NVIDIA/logits-processor-zoo): additional logits processors for controlling text generation.

docs/source/en/main_classes/video_processor.md

@@ -0,0 +1,55 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+
+# Video Processor
+
+A **Video Processor** is a utility responsible for preparing input features for video models, as well as handling the post-processing of their outputs. It provides transformations such as resizing, normalization, and conversion into PyTorch. 
+
+The video processor extends the functionality of image processors by allowing Vision Large Language Models (VLMs) to handle videos with a distinct set of arguments compared to images. It serves as the bridge between raw video data and the model, ensuring that input features are optimized for the VLM.
+
+When adding a new VLM or updating an existing one to enable distinct video preprocessing, saving and reloading the processor configuration will store the video related arguments in a dedicated file named `video_preprocessing_config.json`. Don't worry if you haven't upadted your VLM, the processor will try to load video related configurations from a file named `preprocessing_config.json`.
+
+
+### Usage Example
+Here's an example of how to load a video processor with [`llava-hf/llava-onevision-qwen2-0.5b-ov-hf`](https://huggingface.co/llava-hf/llava-onevision-qwen2-0.5b-ov-hf) model:
+
+```python
+from transformers import AutoVideoProcessor
+
+processor = AutoVideoProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-0.5b-ov-hf")
+```
+
+Currently, if using base image processor for videos, it processes video data by treating each frame as an individual image and applying transformations frame-by-frame. While functional, this approach is not highly efficient. Using `AutoVideoProcessor` allows us to take advantage of **fast video processors**, leveraging the [torchvision](https://pytorch.org/vision/stable/index.html) library. Fast processors handle the whole batch of videos at once, without iterating over each video or frame. These updates introduce GPU acceleration and significantly enhance processing speed, especially for tasks requiring high throughput.
+
+Fast video processors are available for all models and are loaded by default when an `AutoVideoProcessor` is initialized. When using a fast video processor, you can also set the `device` argument to specify the device on which the processing should be done. By default, the processing is done on the same device as the inputs if the inputs are tensors, or on the CPU otherwise. For even more speed improvement, we can compile the processor when using 'cuda' as device.
+
+```python
+import torch
+from transformers.video_utils import load_video
+from transformers import AutoVideoProcessor
+
+video = load_video("video.mp4")
+processor = AutoVideoProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-0.5b-ov-hf", device="cuda")
+processor = torch.compile(processor)
+processed_video = processor(video, return_tensors="pt")
+```
+
+
+## BaseVideoProcessor
+
+[[autodoc]] video_processing_utils.BaseVideoProcessor
+

docs/source/en/model_doc/albert.md

@@ -57,6 +57,7 @@ This model was contributed by [lysandre](https://huggingface.co/lysandre). This
 - Embedding size E is different from hidden size H justified because the embeddings are context independent (one embedding vector represents one token), whereas hidden states are context dependent (one hidden state represents a sequence of tokens) so it's more logical to have H >> E. Also, the embedding matrix is large since it's V x E (V being the vocab size). If E < H, it has less parameters.
 - Layers are split in groups that share parameters (to save memory).
 Next sentence prediction is replaced by a sentence ordering prediction: in the inputs, we have two sentences A and B (that are consecutive) and we either feed A followed by B or B followed by A. The model must predict if they have been swapped or not.
+- The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`  
 
 ### Using Scaled Dot Product Attention (SDPA)
 

docs/source/en/model_doc/aria.md

@@ -102,6 +102,10 @@ response = processor.decode(output_ids, skip_special_tokens=True)
 
 [[autodoc]] AriaTextModel
 
+## AriaModel
+
+[[autodoc]] AriaModel
+
 ## AriaTextForCausalLM
 
 [[autodoc]] AriaTextForCausalLM

docs/source/en/model_doc/auto.md

@@ -74,6 +74,10 @@ Likewise, if your `NewModel` is a subclass of [`PreTrainedModel`], make sure its
 
 [[autodoc]] AutoImageProcessor
 
+## AutoVideoProcessor
+
+[[autodoc]] AutoVideoProcessor
+
 ## AutoProcessor
 
 [[autodoc]] AutoProcessor

docs/source/en/model_doc/aya_vision.md

@@ -237,6 +237,10 @@ for i, output in enumerate(batch_outputs):
 
 [[autodoc]] AyaVisionConfig
 
+## AyaVisionModel
+
+[[autodoc]] AyaVisionModel
+
 ## AyaVisionForConditionalGeneration
 
 [[autodoc]] AyaVisionForConditionalGeneration

docs/source/en/model_doc/bamba.md

@@ -39,7 +39,7 @@ Checkout all Bamba-9B model checkpoints [here](https://github.com/foundation-mod
 <!---
 ## Usage Tips
 
-Tips: 
+Tips:
 
 - The architecture is based on Mamba-2 models.
 
@@ -63,7 +63,35 @@ response = model.generate(**inputs, max_new_tokens=64)
 print(tokenizer.batch_decode(response, skip_special_tokens=True)[0])
 ```
 
+
+## Padding-Free Training
+
+Bamba supports padding-free training in which distinct training examples can be concatenated
+together while nevertheless processing the inputs as though they belonged to separate batches. When
+the examples are of varying lengths, padding-free training can provide significant speed ups and
+memory savings compared to batching the examples together and using padding, as the unnecessary
+compute and memory due to padding is avoided entirely. The performance gains depend on factors such
+as the model and the data distribution, but throughput gains up to [~2x are commonly
+seen](https://github.com/huggingface/transformers/pull/35861#issue-2807873129).
+
+Using padding-free training with Bamba requires the `flash-attn`, `mamba-ssm`, and `causal-conv1d`
+packages, and the following arguments must be passed to the model in addition to `input_ids` and
+`labels`:
+* `position_ids: torch.LongTensor`: the position index of each token in each sequence.
+* `seq_idx: torch.IntTensor`: the index of each sequence in the batch.
+* Each of the [`FlashAttentionKwargs`]
+    * `cu_seq_lens_q: torch.LongTensor`: The cumulative sequence lengths of all queries.
+    * `cu_seq_lens_k: torch.LongTensor`: The cumulative sequence lengths of all keys.
+    * `max_length_q: int`: the longest query length in the batch.
+    * `max_length_k: int`: the longest key length in the batch.
+
+The `attention_mask` inputs should not be provided. The [`DataCollatorWithFlattening`] can be used
+to programmatically generate the above set of additional arguments using `return_seq_idx=True` and
+`return_flash_attn_kwargs=True`. See [this blog post](https://huggingface.co/blog/packing-with-FA2)
+for additional information.
+
+
 [[autodoc]] BambaForCausalLM
     - forward
 
-This HF implementation is contributed by [ani300](https://github.com/ani300) and [fabianlim](https://github.com/fabianlim). 
+This HF implementation is contributed by [ani300](https://github.com/ani300) and [fabianlim](https://github.com/fabianlim).

docs/source/en/model_doc/bart.md

@@ -55,6 +55,7 @@ This model was contributed by [sshleifer](https://huggingface.co/sshleifer). The
   * mask a span of k tokens with a single mask token (a span of 0 tokens is an insertion of a mask token)
   * permute sentences
   * rotate the document to make it start at a specific token
+- The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`  
 
 ## Implementation Notes
 

docs/source/en/model_doc/beit.md

@@ -151,6 +151,12 @@ If you're interested in submitting a resource to be included here, please feel f
     - preprocess
     - post_process_semantic_segmentation
 
+## BeitImageProcessorFast
+
+[[autodoc]] BeitImageProcessorFast
+    - preprocess
+    - post_process_semantic_segmentation
+
 <frameworkcontent>
 <pt>
 

docs/source/en/model_doc/bert.md

@@ -81,10 +81,10 @@ print(f"The predicted token is: {predicted_token}")
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "Plants create [MASK] through a process known as photosynthesis." | transformers-cli run --task fill-mask --model google-bert/bert-base-uncased --device 0
+echo -e "Plants create [MASK] through a process known as photosynthesis." | transformers run --task fill-mask --model google-bert/bert-base-uncased --device 0
 ```
 
 </hfoption>
@@ -256,4 +256,4 @@ echo -e "Plants create [MASK] through a process known as photosynthesis." | tran
 
 [[autodoc]] models.bert.modeling_tf_bert.TFBertForPreTrainingOutput
 
-[[autodoc]] models.bert.modeling_flax_bert.FlaxBertForPreTrainingOutput
+[[autodoc]] models.bert.modeling_flax_bert.FlaxBertForPreTrainingOutput

docs/source/en/model_doc/biogpt.md

@@ -36,6 +36,7 @@ This model was contributed by [kamalkraj](https://huggingface.co/kamalkraj). The
 - BioGPT is a model with absolute position embeddings so it's usually advised to pad the inputs on the right rather than the left.
 - BioGPT was trained with a causal language modeling (CLM) objective and is therefore powerful at predicting the next token in a sequence. Leveraging this feature allows BioGPT to generate syntactically coherent text as it can be observed in the run_generation.py example script.
 - The model can take the `past_key_values` (for PyTorch) as input, which is the previously computed key/value attention pairs. Using this (past_key_values or past) value prevents the model from re-computing pre-computed values in the context of text generation. For PyTorch, see past_key_values argument of the BioGptForCausalLM.forward() method for more information on its usage.
+- The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`  
 
 ### Using Scaled Dot Product Attention (SDPA)
 

docs/source/en/model_doc/code_llama.md

@@ -35,7 +35,7 @@ The example below demonstrates how to generate code with [`Pipeline`], or the [`
 
 <hfoptions id="usage">
 <hfoption id="Pipeline">
-    
+
 ```py
 import torch
 from transformers import pipeline
@@ -76,7 +76,7 @@ prompt = "# Function to calculate the factorial of a number\ndef factorial(n):"
 input_ids = tokenizer(prompt, return_tensors="pt").to("cuda")
 
 output = model.generate(
-    **input_ids, 
+    **input_ids,
     max_new_tokens=256,
     cache_implementation="static"
 )
@@ -92,10 +92,10 @@ print(filled_text)
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
-    
+<hfoption id="transformers CLI">
+
 ```bash
-echo -e "# Function to calculate the factorial of a number\ndef factorial(n):" | transformers-cli run --task text-generation --model meta-llama/CodeLlama-7b-hf --device 0
+echo -e "# Function to calculate the factorial of a number\ndef factorial(n):" | transformers run --task text-generation --model meta-llama/CodeLlama-7b-hf --device 0
 ```
 
 </hfoption>
@@ -146,7 +146,7 @@ visualizer("""def func(a, b):
 - Use the `<FILL_ME>` token where you want your input to be filled. The tokenizer splits this token to create a formatted input string that follows the [original training pattern](https://github.com/facebookresearch/codellama/blob/cb51c14ec761370ba2e2bc351374a79265d0465e/llama/generation.py#L402). This is more robust than preparing the pattern yourself.
     ```py
     from transformers import LlamaForCausalLM, CodeLlamaTokenizer
-    
+
     tokenizer = CodeLlamaTokenizer.from_pretrained("meta-llama/CodeLlama-7b-hf")
     model = LlamaForCausalLM.from_pretrained("meta-llama/CodeLlama-7b-hf")
     PROMPT = '''def remove_non_ascii(s: str) -> str:
@@ -155,7 +155,7 @@ visualizer("""def func(a, b):
     '''
     input_ids = tokenizer(PROMPT, return_tensors="pt")["input_ids"]
     generated_ids = model.generate(input_ids, max_new_tokens=128)
-    
+
     filling = tokenizer.batch_decode(generated_ids[:, input_ids.shape[1]:], skip_special_tokens = True)[0]
     print(PROMPT.replace("<FILL_ME>", filling))
     ```

docs/source/en/model_doc/cohere.md

@@ -49,9 +49,9 @@ model = AutoModelForCausalLM.from_pretrained("CohereForAI/c4ai-command-r-v01", t
 messages = [{"role": "user", "content": "How do plants make energy?"}]
 input_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
 output = model.generate(
-    input_ids, 
-    max_new_tokens=100, 
-    do_sample=True, 
+    input_ids,
+    max_new_tokens=100,
+    do_sample=True,
     temperature=0.3,
     cache_implementation="static",
 )
@@ -59,11 +59,11 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
 # pip install -U flash-attn --no-build-isolation
-transformers-cli chat --model_name_or_path CohereForAI/c4ai-command-r-v01 --torch_dtype auto --attn_implementation flash_attention_2
+transformers chat CohereForAI/c4ai-command-r-v01 --torch_dtype auto --attn_implementation flash_attention_2
 ```
 
 </hfoption>
@@ -85,9 +85,9 @@ model = AutoModelForCausalLM.from_pretrained("CohereForAI/c4ai-command-r-v01", t
 messages = [{"role": "user", "content": "How do plants make energy?"}]
 input_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
 output = model.generate(
-    input_ids, 
-    max_new_tokens=100, 
-    do_sample=True, 
+    input_ids,
+    max_new_tokens=100,
+    do_sample=True,
     temperature=0.3,
     cache_implementation="static",
 )

docs/source/en/model_doc/csm.md

@@ -0,0 +1,377 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Csm
+
+## Overview
+
+The Conversational Speech Model (CSM) is the first open-source contextual text-to-speech model [released by Sesame](https://www.sesame.com/research/crossing_the_uncanny_valley_of_voice). It is designed to generate natural-sounding speech with or without conversational context. This context typically consists of multi-turn dialogue between speakers, represented as sequences of text and corresponding spoken audio.
+
+**Model Architecture:**
+CSM is composed of two LLaMA-style auto-regressive transformer decoders: a backbone decoder that predicts the first codebook token and a depth decoder that generates the remaining tokens. It uses the pretrained codec model [Mimi](./mimi.md), introduced by Kyutai, to encode speech into discrete codebook tokens and decode them back into audio.
+
+The original csm-1b checkpoint is available under the [Sesame](https://huggingface.co/sesame/csm-1b) organization on Hugging Face.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/eustlb/documentation-images/resolve/main/csm_architecture.png"/>
+</div>
+
+## Usage Tips
+
+### Without Conversational Context
+
+CSM can be used to simply generate speech from a text prompt:
+
+```python
+import torch
+from transformers import CsmForConditionalGeneration, AutoProcessor
+
+model_id = "eustlb/csm-1b"
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+# load the model and the processor
+processor = AutoProcessor.from_pretrained(model_id)
+model = CsmForConditionalGeneration.from_pretrained(model_id, device_map=device)
+
+# prepare the inputs
+text = "[0]The past is just a story we tell ourselves." # `[0]` for speaker id 0
+inputs = processor(text, add_special_tokens=True).to(device)
+
+# another equivalent way to prepare the inputs
+conversation = [
+    {"role": "0", "content": [{"type": "text", "text": "The past is just a story we tell ourselves."}]},
+]
+inputs = processor.apply_chat_template(
+    conversation,
+    tokenize=True,
+    return_dict=True,
+).to(device)
+
+# infer the model
+audio = model.generate(**inputs, output_audio=True)
+processor.save_audio(audio, "example_without_context.wav")
+```
+
+### With Conversational Context
+
+CSM can be used to generate speech given a conversation, allowing consistency in the voices and content-aware generation:
+
+```python
+import torch
+from transformers import CsmForConditionalGeneration, AutoProcessor
+from datasets import load_dataset, Audio
+
+model_id = "eustlb/csm-1b"
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+# load the model and the processor
+processor = AutoProcessor.from_pretrained(model_id)
+model = CsmForConditionalGeneration.from_pretrained(model_id, device_map=device)
+
+# prepare the inputs
+ds = load_dataset("hf-internal-testing/dailytalk-dummy", split="train")
+# ensure the audio is 24kHz
+ds = ds.cast_column("audio", Audio(sampling_rate=24000))
+conversation = []
+
+# 1. context
+for text, audio, speaker_id in zip(ds[:4]["text"], ds[:4]["audio"], ds[:4]["speaker_id"]):
+    conversation.append(
+        {
+            "role": f"{speaker_id}",
+            "content": [{"type": "text", "text": text}, {"type": "audio", "path": audio["array"]}],
+        }
+    )
+
+# 2. text prompt
+conversation.append({"role": f"{ds[4]['speaker_id']}", "content": [{"type": "text", "text": ds[4]["text"]}]})
+
+inputs = processor.apply_chat_template(
+    conversation,
+    tokenize=True,
+    return_dict=True,
+).to(device)
+
+# infer the model
+audio = model.generate(**inputs, output_audio=True)
+processor.save_audio(audio, "example_with_context.wav")
+```
+
+### Batched Inference
+
+CSM supports batched inference!
+
+```python
+import torch
+from transformers import CsmForConditionalGeneration, AutoProcessor
+from datasets import load_dataset, Audio
+
+model_id = "eustlb/csm-1b"
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+# load the model and the processor
+processor = AutoProcessor.from_pretrained(model_id)
+model = CsmForConditionalGeneration.from_pretrained(model_id, device_map=device)
+
+# prepare the inputs 
+ds = load_dataset("hf-internal-testing/dailytalk-dummy", split="train")
+# ensure the audio is 24kHz
+ds = ds.cast_column("audio", Audio(sampling_rate=24000))
+# here a batch with two prompts
+conversation = [
+    [
+        {
+            "role": f"{ds[0]['speaker_id']}",
+            "content": [
+                {"type": "text", "text": ds[0]["text"]},
+                {"type": "audio", "path": ds[0]["audio"]["array"]},
+            ],
+        },
+        {
+            "role": f"{ds[1]['speaker_id']}",
+            "content": [
+                {"type": "text", "text": ds[1]["text"]},
+            ],
+        },
+    ],
+    [
+        {
+            "role": f"{ds[0]['speaker_id']}",
+            "content": [
+                {"type": "text", "text": ds[0]["text"]},
+            ],
+        }
+    ],
+]
+inputs = processor.apply_chat_template(
+    conversation,
+    tokenize=True,
+    return_dict=True,
+).to(device)
+
+audio = model.generate(**inputs, output_audio=True)
+processor.save_audio(audio, [f"speech_batch_idx_{i}.wav" for i in range(len(audio))])
+```
+
+### Making The Model Go Brrr
+
+CSM supports full-graph compilation with CUDA graphs!
+
+```python
+import torch
+import copy
+from transformers import CsmForConditionalGeneration, AutoProcessor
+from datasets import load_dataset
+
+model_id = "eustlb/csm-1b"
+device = "cuda"
+
+# set logs to ensure no recompilation and graph breaks
+torch._logging.set_logs(graph_breaks=True, recompiles=True, cudagraphs=True)
+
+# load the model and the processor
+processor = AutoProcessor.from_pretrained(model_id)
+model = CsmForConditionalGeneration.from_pretrained(model_id, device_map=device)
+
+# use static cache, enabling automatically torch compile with fullgraph and reduce-overhead
+model.generation_config.max_length = 250 # big enough to avoid recompilation
+model.generation_config.max_new_tokens = None # would take precedence over max_length
+model.generation_config.cache_implementation = "static"
+model.depth_decoder.generation_config.cache_implementation = "static"
+
+# generation kwargs
+gen_kwargs = {
+    "do_sample": False,
+    "depth_decoder_do_sample": False,
+    "temperature": 1.0,
+    "depth_decoder_temperature": 1.0,
+}
+
+# Define a timing decorator
+class TimerContext:
+    def __init__(self, name="Execution"):
+        self.name = name
+        self.start_event = None
+        self.end_event = None
+        
+    def __enter__(self):
+        # Use CUDA events for more accurate GPU timing
+        self.start_event = torch.cuda.Event(enable_timing=True)
+        self.end_event = torch.cuda.Event(enable_timing=True)
+        self.start_event.record()
+        return self
+
+    def __exit__(self, *args):
+        self.end_event.record()
+        torch.cuda.synchronize()
+        elapsed_time = self.start_event.elapsed_time(self.end_event) / 1000.0
+        print(f"{self.name} time: {elapsed_time:.4f} seconds")
+
+# prepare the inputs 
+ds = load_dataset("hf-internal-testing/dailytalk-dummy", split="train")
+
+conversation = [
+    {
+        "role": f"{ds[0]['speaker_id']}",
+        "content": [
+            {"type": "text", "text": ds[0]["text"]},
+            {"type": "audio", "path": ds[0]["audio"]["array"]},
+        ],
+    },
+    {
+        "role": f"{ds[1]['speaker_id']}",
+        "content": [
+            {"type": "text", "text": ds[1]["text"]},
+            {"type": "audio", "path": ds[1]["audio"]["array"]},
+        ],
+    },
+    {
+        "role": f"{ds[2]['speaker_id']}",
+        "content": [
+            {"type": "text", "text": ds[2]["text"]},
+        ],
+    },
+]
+
+padded_inputs_1 = processor.apply_chat_template(
+    conversation,
+    tokenize=True,
+    return_dict=True,
+).to(device)
+
+print("\n" + "="*50)
+print("First generation - compiling and recording CUDA graphs...")
+with TimerContext("First generation"):
+    _ = model.generate(**padded_inputs_1, **gen_kwargs)
+print("="*50)
+
+print("\n" + "="*50)
+print("Second generation - fast !!!")
+with TimerContext("Second generation"):
+    _ = model.generate(**padded_inputs_1, **gen_kwargs)
+print("="*50)
+
+# now with different inputs
+conversation = [
+    {
+        "role": f"{ds[0]['speaker_id']}",
+        "content": [
+            {"type": "text", "text": ds[2]["text"]},
+            {"type": "audio", "path": ds[2]["audio"]["array"]},
+        ],
+    },
+    {
+        "role": f"{ds[1]['speaker_id']}",
+        "content": [
+            {"type": "text", "text": ds[3]["text"]},
+            {"type": "audio", "path": ds[3]["audio"]["array"]},
+        ],
+    },
+    {
+        "role": f"{ds[2]['speaker_id']}",
+        "content": [
+            {"type": "text", "text": ds[4]["text"]},
+        ],
+    },
+]
+padded_inputs_2 = processor.apply_chat_template(
+    conversation,
+    tokenize=True,
+    return_dict=True,
+).to(device)
+
+print("\n" + "="*50)
+print("Generation with other inputs!")
+with TimerContext("Generation with different inputs"):
+    _ = model.generate(**padded_inputs_2, **gen_kwargs)
+print("="*50)
+```
+
+### Training
+
+CSM Transformers integration supports training!
+
+```python
+from transformers import CsmForConditionalGeneration, AutoProcessor
+from datasets import load_dataset, Audio
+
+model_id = "eustlb/csm-1b"
+device = "cuda"
+
+# load the model and the processor
+processor = AutoProcessor.from_pretrained(model_id)
+model = CsmForConditionalGeneration.from_pretrained(model_id, device_map=device)
+model.train()
+
+ds = load_dataset("hf-internal-testing/dailytalk-dummy", split="train")
+# ensure the audio is 24kHz
+ds = ds.cast_column("audio", Audio(sampling_rate=24000))
+conversation = []
+
+# context
+for text, audio, speaker_id in zip(ds[:4]["text"], ds[:4]["audio"], ds[:4]["speaker_id"]):
+    conversation.append(
+        {
+            "role": f"{speaker_id}",
+            "content": [{"type": "text", "text": text}, {"type": "audio", "path": audio["array"]}],
+        }
+    )
+
+inputs = processor.apply_chat_template(
+    conversation,
+    tokenize=True,
+    return_dict=True,
+    output_labels=True,
+).to(device)
+
+out = model(**inputs)
+out.loss.backward()
+```
+
+This model was contributed by [Eustache Le Bihan](https://huggingface.co/eustlb).
+The original code can be found [here](https://github.com/SesameAILabs/csm).
+
+
+## CsmConfig
+
+[[autodoc]] CsmConfig
+
+## CsmDepthDecoderConfig
+
+[[autodoc]] CsmDepthDecoderConfig
+
+## CsmProcessor
+
+[[autodoc]] CsmProcessor
+    - __call__
+
+## CsmForConditionalGeneration
+
+[[autodoc]] CsmForConditionalGeneration
+    - forward
+    - generate
+
+## CsmDepthDecoderForCausalLM
+
+[[autodoc]] CsmDepthDecoderForCausalLM
+
+## CsmDepthDecoderModel
+
+[[autodoc]] CsmDepthDecoderModel
+
+## CsmBackboneModel
+
+[[autodoc]] CsmBackboneModel

docs/source/en/model_doc/d_fine.md

@@ -0,0 +1,76 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# D-FINE
+
+## Overview
+
+The D-FINE model was proposed in [D-FINE: Redefine Regression Task in DETRs as Fine-grained Distribution Refinement](https://arxiv.org/abs/2410.13842) by
+Yansong Peng, Hebei Li, Peixi Wu, Yueyi Zhang, Xiaoyan Sun, Feng Wu
+
+The abstract from the paper is the following:
+
+*We introduce D-FINE, a powerful real-time object detector that achieves outstanding localization precision by redefining the bounding box regression task in DETR models. D-FINE comprises two key components: Fine-grained Distribution Refinement (FDR) and Global Optimal Localization Self-Distillation (GO-LSD). 
+FDR transforms the regression process from predicting fixed coordinates to iteratively refining probability distributions, providing a fine-grained intermediate representation that significantly enhances localization accuracy. GO-LSD is a bidirectional optimization strategy that transfers localization knowledge from refined distributions to shallower layers through self-distillation, while also simplifying the residual prediction tasks for deeper layers. Additionally, D-FINE incorporates lightweight optimizations in computationally intensive modules and operations, achieving a better balance between speed and accuracy. Specifically, D-FINE-L / X achieves 54.0% / 55.8% AP on the COCO dataset at 124 / 78 FPS on an NVIDIA T4 GPU. When pretrained on Objects365, D-FINE-L / X attains 57.1% / 59.3% AP, surpassing all existing real-time detectors. Furthermore, our method significantly enhances the performance of a wide range of DETR models by up to 5.3% AP with negligible extra parameters and training costs. Our code and pretrained models: this https URL.*
+
+This model was contributed by [VladOS95-cyber](https://github.com/VladOS95-cyber). 
+The original code can be found [here](https://github.com/Peterande/D-FINE).
+
+## Usage tips 
+
+```python
+>>> import torch
+>>> from transformers.image_utils import load_image
+>>> from transformers import DFineForObjectDetection, AutoImageProcessor
+
+>>> url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
+>>> image = load_image(url)
+
+>>> image_processor = AutoImageProcessor.from_pretrained("ustc-community/dfine_x_coco")
+>>> model = DFineForObjectDetection.from_pretrained("ustc-community/dfine_x_coco")
+
+>>> inputs = image_processor(images=image, return_tensors="pt")
+
+>>> with torch.no_grad():
+...     outputs = model(**inputs)
+
+>>> results = image_processor.post_process_object_detection(outputs, target_sizes=[(image.height, image.width)], threshold=0.5)
+
+>>> for result in results:
+...     for score, label_id, box in zip(result["scores"], result["labels"], result["boxes"]):
+...         score, label = score.item(), label_id.item()
+...         box = [round(i, 2) for i in box.tolist()]
+...         print(f"{model.config.id2label[label]}: {score:.2f} {box}")
+cat: 0.96 [344.49, 23.4, 639.84, 374.27]
+cat: 0.96 [11.71, 53.52, 316.64, 472.33]
+remote: 0.95 [40.46, 73.7, 175.62, 117.57]
+sofa: 0.92 [0.59, 1.88, 640.25, 474.74]
+remote: 0.89 [333.48, 77.04, 370.77, 187.3]
+```
+
+## DFineConfig
+
+[[autodoc]] DFineConfig
+
+## DFineModel
+
+[[autodoc]] DFineModel
+    - forward
+
+## DFineForObjectDetection
+
+[[autodoc]] DFineForObjectDetection
+    - forward

docs/source/en/model_doc/data2vec.md

@@ -53,6 +53,7 @@ The original code for vision can be found [here](https://github.com/facebookrese
 - For Data2VecAudio, preprocessing is identical to [`Wav2Vec2Model`], including feature extraction
 - For Data2VecText, preprocessing is identical to [`RobertaModel`], including tokenization.
 - For Data2VecVision, preprocessing is identical to [`BeitModel`], including feature extraction.
+- The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`  
 
 ### Using Scaled Dot Product Attention (SDPA)
 

docs/source/en/model_doc/dinov2.md

@@ -111,33 +111,68 @@ print("Predicted class:", model.config.id2label[predicted_class_idx])
 
 ## Notes
 
-- Use [torch.jit.trace](https://pytorch.org/docs/stable/generated/torch.jit.trace.html) to speedup inference. However, it will produce some mismatched elements. The difference between the original and traced model is 1e-4.
+- The example below shows how to split the output tensor into:
+  - one embedding for the whole image, commonly referred to as a `CLS` token,
+    useful for classification and retrieval
+  - a set of local embeddings, one for each `14x14` patch of the input image,
+    useful for dense tasks, such as semantic segmentation
 
-    ```py
-    import torch
-    from transformers import AutoImageProcessor, AutoModel
-    from PIL import Image
-    import requests
+  ```py
+  from transformers import AutoImageProcessor, AutoModel
+  from PIL import Image
+  import requests
+  
+  url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
+  image = Image.open(requests.get(url, stream=True).raw)
+  print(image.height, image.width)  # [480, 640]
+  
+  processor = AutoImageProcessor.from_pretrained('facebook/dinov2-base')
+  model = AutoModel.from_pretrained('facebook/dinov2-base')
+  patch_size = model.config.patch_size
+  
+  inputs = processor(images=image, return_tensors="pt")
+  print(inputs.pixel_values.shape)  # [1, 3, 224, 224]
+  batch_size, rgb, img_height, img_width = inputs.pixel_values.shape
+  num_patches_height, num_patches_width = img_height // patch_size, img_width // patch_size
+  num_patches_flat = num_patches_height * num_patches_width
+  
+  outputs = model(**inputs)
+  last_hidden_states = outputs[0]
+  print(last_hidden_states.shape)  # [1, 1 + 256, 768]
+  assert last_hidden_states.shape == (batch_size, 1 + num_patches_flat, model.config.hidden_size)
+  
+  cls_token = last_hidden_states[:, 0, :]
+  patch_features = last_hidden_states[:, 1:, :].unflatten(1, (num_patches_height, num_patches_width))
+  ```
 
-    url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
-    image = Image.open(requests.get(url, stream=True).raw)
+- Use [torch.jit.trace](https://pytorch.org/docs/stable/generated/torch.jit.trace.html) to speedup inference.
+  However, it will produce some mismatched elements. The difference between the original and traced model is 1e-4.
 
-    processor = AutoImageProcessor.from_pretrained('facebook/dinov2-base')
-    model = AutoModel.from_pretrained('facebook/dinov2-base')
-
-    inputs = processor(images=image, return_tensors="pt")
-    outputs = model(**inputs)
-    last_hidden_states = outputs[0]
-
-    # We have to force return_dict=False for tracing
-    model.config.return_dict = False
-
-    with torch.no_grad():
-        traced_model = torch.jit.trace(model, [inputs.pixel_values])
-        traced_outputs = traced_model(inputs.pixel_values)
-
-    print((last_hidden_states - traced_outputs[0]).abs().max())
-    ```
+  ```py
+  import torch
+  from transformers import AutoImageProcessor, AutoModel
+  from PIL import Image
+  import requests
+  
+  url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
+  image = Image.open(requests.get(url, stream=True).raw)
+  
+  processor = AutoImageProcessor.from_pretrained('facebook/dinov2-base')
+  model = AutoModel.from_pretrained('facebook/dinov2-base')
+  
+  inputs = processor(images=image, return_tensors="pt")
+  outputs = model(**inputs)
+  last_hidden_states = outputs[0]
+  
+  # We have to force return_dict=False for tracing
+  model.config.return_dict = False
+  
+  with torch.no_grad():
+      traced_model = torch.jit.trace(model, [inputs.pixel_values])
+      traced_outputs = traced_model(inputs.pixel_values)
+  
+  print((last_hidden_states - traced_outputs[0]).abs().max())
+  ```
 
 ## Dinov2Config
 

docs/source/en/model_doc/distilbert.md

@@ -83,10 +83,10 @@ print(f"Predicted label: {predicted_label}")
 
 </hfoption>
 
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "I love using Hugging Face Transformers!" | transformers-cli run --task text-classification --model distilbert-base-uncased-finetuned-sst-2-english
+echo -e "I love using Hugging Face Transformers!" | transformers run --task text-classification --model distilbert-base-uncased-finetuned-sst-2-english
 ```
 
 </hfoption>
@@ -213,7 +213,3 @@ echo -e "I love using Hugging Face Transformers!" | transformers-cli run --task
 
 </jax>
 </frameworkcontent>
-
-
-
-

docs/source/en/model_doc/electra.md

@@ -45,9 +45,9 @@ import torch
 from transformers import pipeline
 
 classifier = pipeline(
-    task="text-classification", 
-    model="bhadresh-savani/electra-base-emotion", 
-    torch_dtype=torch.float16, 
+    task="text-classification",
+    model="bhadresh-savani/electra-base-emotion",
+    torch_dtype=torch.float16,
     device=0
 )
 classifier("This restaurant has amazing food!")
@@ -64,7 +64,7 @@ tokenizer = AutoTokenizer.from_pretrained(
     "bhadresh-savani/electra-base-emotion",
 )
 model = AutoModelForSequenceClassification.from_pretrained(
-    "bhadresh-savani/electra-base-emotion", 
+    "bhadresh-savani/electra-base-emotion",
     torch_dtype=torch.float16
 )
 inputs = tokenizer("ELECTRA is more efficient than BERT", return_tensors="pt")
@@ -78,10 +78,10 @@ print(f"Predicted label: {predicted_label}")
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "This restaurant has amazing food." | transformers-cli run --task text-classification --model bhadresh-savani/electra-base-emotion --device 0
+echo -e "This restaurant has amazing food." | transformers run --task text-classification --model bhadresh-savani/electra-base-emotion --device 0
 ```
 
 </hfoption>
@@ -96,12 +96,12 @@ echo -e "This restaurant has amazing food." | transformers-cli run --task text-c
 
     ```py
     # Example of properly handling padding with attention masks
-    inputs = tokenizer(["Short text", "This is a much longer text that needs padding"], 
-                    padding=True, 
+    inputs = tokenizer(["Short text", "This is a much longer text that needs padding"],
+                    padding=True,
                     return_tensors="pt")
     outputs = model(**inputs)  # automatically uses the attention_mask
     ```
-    
+
 - When using the discriminator for a downstream task, you can load it into any of the ELECTRA model classes ([`ElectraForSequenceClassification`], [`ElectraForTokenClassification`], etc.).
 
 ## ElectraConfig

docs/source/en/model_doc/emu3.md

@@ -174,6 +174,10 @@ for i, image in enumerate(images['pixel_values']):
 [[autodoc]] Emu3TextModel
     - forward
 
+## Emu3Model
+
+[[autodoc]] Emu3Model
+
 ## Emu3ForCausalLM
 
 [[autodoc]] Emu3ForCausalLM

docs/source/en/model_doc/falcon.md

@@ -41,7 +41,7 @@ import torch
 from transformers import pipeline
 
 pipeline = pipeline(
-    task="text-generation", 
+    task="text-generation",
     model="tiiuae/falcon-7b-instruct",
     torch_dtype=torch.bfloat16,
     device=0
@@ -76,11 +76,11 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
 # pip install -U flash-attn --no-build-isolation
-transformers-cli chat --model_name_or_path tiiuae/falcon-7b-instruct --torch_dtype auto --attn_implementation flash_attention_2 --device 0
+transformers chat tiiuae/falcon-7b-instruct --torch_dtype auto --attn_implementation flash_attention_2 --device 0
 ```
 
 </hfoption>
@@ -150,4 +150,4 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
 ## FalconForQuestionAnswering
 
 [[autodoc]] FalconForQuestionAnswering
-    - forward
+    - forward

docs/source/en/model_doc/falcon_mamba.md

@@ -39,7 +39,7 @@ import torch
 from transformers import pipeline
 
 pipeline = pipeline(
-    "text-generation", 
+    "text-generation",
     model="tiiuae/falcon-mamba-7b-instruct",
     torch_dtype=torch.bfloat16,
     device=0
@@ -73,10 +73,10 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-transformers-cli chat --model_name_or_path tiiuae/falcon-mamba-7b-instruct --torch_dtype auto --device 0
+transformers chat tiiuae/falcon-mamba-7b-instruct --torch_dtype auto --device 0
 ```
 
 </hfoption>

docs/source/en/model_doc/fuyu.md

@@ -103,6 +103,10 @@ The `LlamaTokenizer` is used as it is a standard wrapper around sentencepiece.
 
 [[autodoc]] FuyuConfig
 
+## FuyuModel
+
+[[autodoc]] FuyuModel
+
 ## FuyuForCausalLM
 
 [[autodoc]] FuyuForCausalLM

docs/source/en/model_doc/gemma.md

@@ -80,10 +80,10 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "LLMs generate text through a process known as" | transformers-cli run --task text-generation --model google/gemma-2b --device 0
+echo -e "LLMs generate text through a process known as" | transformers run --task text-generation --model google/gemma-2b --device 0
 ```
 
 </hfoption>
@@ -114,8 +114,8 @@ model = AutoModelForCausalLM.from_pretrained(
 input_text = "LLMs generate text through a process known as."
 input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
 outputs = model.generate(
-    **input_ids, 
-    max_new_tokens=50, 
+    **input_ids,
+    max_new_tokens=50,
     cache_implementation="static"
 )
 print(tokenizer.decode(outputs[0], skip_special_tokens=True))
@@ -127,7 +127,7 @@ Use the [AttentionMaskVisualizer](https://github.com/huggingface/transformers/bl
 from transformers.utils.attention_visualizer import AttentionMaskVisualizer
 
 visualizer = AttentionMaskVisualizer("google/gemma-2b")
-visualizer("LLMs generate text through a process known as") 
+visualizer("LLMs generate text through a process known as")
 ```
 
 <div class="flex justify-center">

docs/source/en/model_doc/gemma2.md

@@ -58,7 +58,7 @@ pipe("Explain quantum computing simply. ", max_new_tokens=50)
 
 </hfoption>
 <hfoption id="AutoModel">
-    
+
 ```python
 import torch
 from transformers import AutoTokenizer, AutoModelForCausalLM
@@ -80,16 +80,16 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```
-echo -e "Explain quantum computing simply." | transformers-cli run --task text-generation --model google/gemma-2-2b --device 0
+echo -e "Explain quantum computing simply." | transformers run --task text-generation --model google/gemma-2-2b --device 0
 ```
 </hfoption>
 </hfoptions>
 
 Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
-	
+
 The example below uses [bitsandbytes](../quantization/bitsandbytes) to only quantize the weights to int4.
 
 ```python
@@ -118,7 +118,7 @@ Use the [AttentionMaskVisualizer](https://github.com/huggingface/transformers/bl
 ```python
 from transformers.utils.attention_visualizer import AttentionMaskVisualizer
 visualizer = AttentionMaskVisualizer("google/gemma-2b")
-visualizer("You are an assistant. Make sure you print me") 
+visualizer("You are an assistant. Make sure you print me")
 ```
 
 <div class="flex justify-center">
@@ -137,7 +137,7 @@ visualizer("You are an assistant. Make sure you print me")
 
     inputs = tokenizer(text="My name is Gemma", return_tensors="pt")
     max_generated_length = inputs.input_ids.shape[1] + 10
-    past_key_values = HybridCache(config=model.config, max_batch_size=1, 
+    past_key_values = HybridCache(config=model.config, max_batch_size=1,
     max_cache_len=max_generated_length, device=model.device, dtype=model.dtype)
     outputs = model(**inputs, past_key_values=past_key_values, use_cache=True)
     ```

docs/source/en/model_doc/gemma3.md

@@ -28,7 +28,7 @@ rendered properly in your Markdown viewer.
 
 The instruction-tuned variant was post-trained with knowledge distillation and reinforcement learning.
 
-You can find all the original Gemma 3 checkpoints under the [Gemma 3](https://huggingface.co/collections/meta-llama/llama-2-family-661da1f90a9d678b6f55773b) release.
+You can find all the original Gemma 3 checkpoints under the [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) release.
 
 > [!TIP]
 > Click on the Gemma 3 models in the right sidebar for more examples of how to apply Gemma to different vision and language tasks.
@@ -99,10 +99,10 @@ print(processor.decode(output[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "Plants create energy through a process known as" | transformers-cli run --task text-generation --model google/gemma-3-1b-pt --device 0
+echo -e "Plants create energy through a process known as" | transformers run --task text-generation --model google/gemma-3-1b-pt --device 0
 ```
 
 </hfoption>
@@ -254,6 +254,10 @@ visualizer("<img>What is shown in this image?")
 [[autodoc]] Gemma3TextModel
     - forward
 
+## Gemma3Model
+
+[[autodoc]] Gemma3Model
+
 ## Gemma3ForCausalLM
 
 [[autodoc]] Gemma3ForCausalLM

docs/source/en/model_doc/got_ocr2.md

@@ -277,6 +277,10 @@ alt="drawing" width="600"/>
 
 [[autodoc]] GotOcr2Processor
 
+## GotOcr2Model
+
+[[autodoc]] GotOcr2Model
+
 ## GotOcr2ForConditionalGeneration
 
 [[autodoc]] GotOcr2ForConditionalGeneration

docs/source/en/model_doc/gpt2.md

@@ -64,15 +64,21 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "Hello, I'm a language model" | transformers-cli run --task text-generation --model openai-community/gpt2 --device 0
+echo -e "Hello, I'm a language model" | transformers run --task text-generation --model openai-community/gpt2 --device 0
 ```
 
 </hfoption>
 </hfoptions>
 
+One can also serve the model using vLLM with the `transformers backend`.
+
+```
+vllm serve openai-community/gpt2 --model-imp transformers
+```
+
 Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
 
 The example below uses [bitsandbytes](../quantization/bitsandbytes) to only quantize the weights to 4-bits.
@@ -82,16 +88,16 @@ import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, pipeline
 
 quantization_config = BitsAndBytesConfig(
-    load_in_4bit=True,  
-    bnb_4bit_quant_type="nf4",  
-    bnb_4bit_compute_dtype="float16",  
-    bnb_4bit_use_double_quant=True 
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype="float16",
+    bnb_4bit_use_double_quant=True
 )
 
 model = AutoModelForCausalLM.from_pretrained(
     "openai-community/gpt2-xl",
     quantization_config=quantization_config,
-    device_map="auto"  
+    device_map="auto"
 )
 
 tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2-xl")

docs/source/en/model_doc/gpt_bigcode.md

@@ -46,8 +46,12 @@ The main differences compared to GPT2.
 - Merge the key and value caches into one (this changes the format of layer_past/ present, does it risk creating problems?)
 - Use the memory layout (self.num_heads, 3, self.head_dim) instead of `(3, self.num_heads, self.head_dim)` for the QKV tensor with MHA. (prevents an overhead with the merged key and values, but makes the checkpoints incompatible with the original openai-community/gpt2 model).
 
+
 You can read more about the optimizations in the [original pull request](https://github.com/huggingface/transformers/pull/22575)
 
+> [!NOTE]
+> The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`
+
 ## Combining Starcoder and Flash Attention 2
 
 First, make sure to install the latest version of Flash Attention 2 to include the sliding window attention feature.

docs/source/en/model_doc/granitemoehybrid.md

@@ -0,0 +1,64 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# GraniteMoeHybrid
+
+## Overview
+
+
+The `GraniteMoeHybrid` model builds on top of `GraniteMoeSharedModel` and `Bamba`. Its decoding layers consist of state space layers or MoE attention layers with shared experts. By default, the attention layers do not use positional encoding.
+
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_path = "ibm-granite/granite-4.0-tiny-preview"
+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 HF implementation is contributed by [Sukriti Sharma](https://huggingface.co/SukritiSharma) and [Alexander Brooks](https://huggingface.co/abrooks9944).
+
+
+## GraniteMoeHybridConfig
+
+[[autodoc]] GraniteMoeHybridConfig
+
+## GraniteMoeHybridModel
+
+[[autodoc]] GraniteMoeHybridModel
+    - forward
+
+## GraniteMoeHybridForCausalLM
+
+[[autodoc]] GraniteMoeHybridForCausalLM
+    - forward

docs/source/en/model_doc/hgnet_v2.md

@@ -0,0 +1,46 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# HGNet-V2
+
+## Overview
+
+A HGNet-V2 (High Performance GPU Net) image classification model.
+HGNet arhtictecture was proposed in [HGNET: A Hierarchical Feature Guided Network for Occupancy Flow Field Prediction](https://arxiv.org/abs/2407.01097) by
+Zhan Chen, Chen Tang, Lu Xiong
+
+The abstract from the HGNET paper is the following:
+
+*Predicting the motion of multiple traffic participants has always been one of the most challenging tasks in autonomous driving. The recently proposed occupancy flow field prediction method has shown to be a more effective and scalable representation compared to general trajectory prediction methods. However, in complex multi-agent traffic scenarios, it remains difficult to model the interactions among various factors and the dependencies among prediction outputs at different time steps. In view of this, we propose a transformer-based hierarchical feature guided network (HGNET), which can efficiently extract features of agents and map information from visual and vectorized inputs, modeling multimodal interaction relationships. Second, we design the Feature-Guided Attention (FGAT) module to leverage the potential guiding effects between different prediction targets, thereby improving prediction accuracy. Additionally, to enhance the temporal consistency and causal relationships of the predictions, we propose a Time Series Memory framework to learn the conditional distribution models of the prediction outputs at future time steps from multivariate time series. The results demonstrate that our model exhibits competitive performance, which ranks 3rd in the 2024 Waymo Occupancy and Flow Prediction Challenge.*
+
+This model was contributed by [VladOS95-cyber](https://github.com/VladOS95-cyber). 
+The original code can be found [here](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/ppdet/modeling/backbones/hgnet_v2.py).
+
+## HGNetV2Config
+
+[[autodoc]] HGNetV2Config
+
+
+## HGNetV2Backbone
+
+[[autodoc]] HGNetV2Backbone
+    - forward
+
+
+## HGNetV2ForImageClassification
+
+[[autodoc]] HGNetV2ForImageClassification
+    - forward

docs/source/en/model_doc/hubert.md

@@ -50,7 +50,7 @@ This model was contributed by [patrickvonplaten](https://huggingface.co/patrickv
 - Hubert is a speech model that accepts a float array corresponding to the raw waveform of the speech signal.
 - Hubert model was fine-tuned using connectionist temporal classification (CTC) so the model output has to be decoded
   using [`Wav2Vec2CTCTokenizer`].
-
+- The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`  
 
 ## Using Flash Attention 2
 

docs/source/en/model_doc/instructblip.md

@@ -69,6 +69,10 @@ The attributes can be obtained from model config, as `model.config.num_query_tok
 [[autodoc]] InstructBlipQFormerModel
     - forward
 
+## InstructBlipModel
+
+[[autodoc]] InstructBlipModel
+
 ## InstructBlipForConditionalGeneration
 
 [[autodoc]] InstructBlipForConditionalGeneration

docs/source/en/model_doc/instructblipvideo.md

@@ -58,6 +58,12 @@ The attributes can be obtained from model config, as `model.config.num_query_tok
 
 [[autodoc]] InstructBlipVideoProcessor
 
+
+## InstructBlipVideoVideoProcessor
+
+[[autodoc]] InstructBlipVideoVideoProcessor
+    - preprocess
+
 ## InstructBlipVideoImageProcessor
 
 [[autodoc]] InstructBlipVideoImageProcessor
@@ -73,6 +79,10 @@ The attributes can be obtained from model config, as `model.config.num_query_tok
 [[autodoc]] InstructBlipVideoQFormerModel
     - forward
 
+## InstructBlipVideoModel
+[[autodoc]] InstructBlipVideoModel
+    - forward
+
 ## InstructBlipVideoForConditionalGeneration
 
 [[autodoc]] InstructBlipVideoForConditionalGeneration

docs/source/en/model_doc/internvl.md

@@ -340,6 +340,11 @@ This example showcases how to handle a batch of chat conversations with interlea
 [[autodoc]] InternVLVisionModel
     - forward
 
+## InternVLModel
+
+[[autodoc]] InternVLModel
+    - forward
+
 ## InternVLForConditionalGeneration
 
 [[autodoc]] InternVLForConditionalGeneration
@@ -348,3 +353,7 @@ This example showcases how to handle a batch of chat conversations with interlea
 ## InternVLProcessor
 
 [[autodoc]] InternVLProcessor
+
+## InternVLVideoProcessor
+
+[[autodoc]] InternVLVideoProcessor

docs/source/en/model_doc/jamba.md

@@ -75,10 +75,10 @@ output = model.generate(**input_ids, cache_implementation="static")
 print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "Plants create energy through a process known as" | transformers-cli run --task text-generation --model ai21labs/AI21-Jamba-Mini-1.6 --device 0
+echo -e "Plants create energy through a process known as" | transformers run --task text-generation --model ai21labs/AI21-Jamba-Mini-1.6 --device 0
 ```
 
 </hfoption>

docs/source/en/model_doc/llama.md

@@ -74,10 +74,10 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "Plants create energy through a process known as" | transformers-cli run --task text-generation --model huggyllama/llama-7b --device 0
+echo -e "Plants create energy through a process known as" | transformers run --task text-generation --model huggyllama/llama-7b --device 0
 ```
 
 </hfoption>

docs/source/en/model_doc/llama2.md

@@ -74,10 +74,10 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-transformers-cli chat --model_name_or_path meta-llama/Llama-2-7b-chat-hf --torch_dtype auto --attn_implementation flash_attention_2
+transformers chat meta-llama/Llama-2-7b-chat-hf --torch_dtype auto --attn_implementation flash_attention_2
 ```
 
 </hfoption>
@@ -175,4 +175,3 @@ visualizer("Plants create energy through a process known as")
 
 [[autodoc]] LlamaForSequenceClassification
     - forward
-

docs/source/en/model_doc/llava.md

@@ -256,6 +256,10 @@ A list of official Hugging Face and community (indicated by 🌎) resources to h
 
 [[autodoc]] LlavaProcessor
 
+## LlavaModel
+
+[[autodoc]] LlavaModel
+
 ## LlavaForConditionalGeneration
 
 [[autodoc]] LlavaForConditionalGeneration

docs/source/en/model_doc/llava_next.md

@@ -315,6 +315,10 @@ model = AutoModelForImageTextToText.from_pretrained(
 
 [[autodoc]] LlavaNextProcessor
 
+## LlavaNextModel
+
+[[autodoc]] LlavaNextModel
+
 ## LlavaNextForConditionalGeneration
 
 [[autodoc]] LlavaNextForConditionalGeneration

docs/source/en/model_doc/llava_next_video.md

@@ -262,6 +262,14 @@ model = LlavaNextVideoForConditionalGeneration.from_pretrained(
 
 [[autodoc]] LlavaNextVideoImageProcessor
 
+## LlavaNextVideoVideoProcessor
+
+[[autodoc]] LlavaNextVideoVideoProcessor
+
+## LlavaNextVideoModel
+
+[[autodoc]] LlavaNextVideoModel
+
 ## LlavaNextVideoForConditionalGeneration
 
 [[autodoc]] LlavaNextVideoForConditionalGeneration

docs/source/en/model_doc/llava_onevision.md

@@ -303,6 +303,7 @@ model = LlavaOnevisionForConditionalGeneration.from_pretrained(
 ## LlavaOnevisionImageProcessor
 
 [[autodoc]] LlavaOnevisionImageProcessor
+    - preprocess
 
 ## LlavaOnevisionImageProcessorFast
 
@@ -313,6 +314,14 @@ model = LlavaOnevisionForConditionalGeneration.from_pretrained(
 
 [[autodoc]] LlavaOnevisionVideoProcessor
 
+## LlavaOnevisionVideoProcessor
+
+[[autodoc]] LlavaOnevisionVideoProcessor
+
+## LlavaOnevisionModel
+
+[[autodoc]] LlavaOnevisionModel
+
 ## LlavaOnevisionForConditionalGeneration
 
 [[autodoc]] LlavaOnevisionForConditionalGeneration

docs/source/en/model_doc/longformer.md

@@ -76,10 +76,10 @@ tokenizer.decode(predictions).split()
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "San Francisco 49ers cornerback Shawntae Spencer will miss the rest of the <mask> with a torn ligament in his left knee." | transformers-cli run --task fill-mask --model allenai/longformer-base-4096 --device 0
+echo -e "San Francisco 49ers cornerback Shawntae Spencer will miss the rest of the <mask> with a torn ligament in his left knee." | transformers run --task fill-mask --model allenai/longformer-base-4096 --device 0
 ```
 
 </hfoption>
@@ -147,42 +147,42 @@ echo -e "San Francisco 49ers cornerback Shawntae Spencer will miss the rest of t
 
 ## LongformerForMaskedLM
 
-[[autodoc]] LongformerForMaskedLM 
+[[autodoc]] LongformerForMaskedLM
     - forward
 
 ## LongformerForSequenceClassification
 
-[[autodoc]] LongformerForSequenceClassification 
+[[autodoc]] LongformerForSequenceClassification
     - forward
 
 ## LongformerForMultipleChoice
 
-[[autodoc]] LongformerForMultipleChoice 
+[[autodoc]] LongformerForMultipleChoice
     - forward
 
 ## LongformerForTokenClassification
 
-[[autodoc]] LongformerForTokenClassification 
+[[autodoc]] LongformerForTokenClassification
     - forward
 
 ## LongformerForQuestionAnswering
 
-[[autodoc]] LongformerForQuestionAnswering 
+[[autodoc]] LongformerForQuestionAnswering
     - forward
 
 ## TFLongformerModel
 
-[[autodoc]] TFLongformerModel    
+[[autodoc]] TFLongformerModel
     - call
 
 ## TFLongformerForMaskedLM
 
-[[autodoc]] TFLongformerForMaskedLM 
+[[autodoc]] TFLongformerForMaskedLM
     - call
 
 ## TFLongformerForQuestionAnswering
 
-[[autodoc]] TFLongformerForQuestionAnswering 
+[[autodoc]] TFLongformerForQuestionAnswering
     - call
 
 ## TFLongformerForSequenceClassification
@@ -192,10 +192,10 @@ echo -e "San Francisco 49ers cornerback Shawntae Spencer will miss the rest of t
 
 ## TFLongformerForTokenClassification
 
-[[autodoc]] TFLongformerForTokenClassification 
+[[autodoc]] TFLongformerForTokenClassification
     - call
 
 ## TFLongformerForMultipleChoice
 
-[[autodoc]] TFLongformerForMultipleChoice 
+[[autodoc]] TFLongformerForMultipleChoice
     - call

docs/source/en/model_doc/m2m_100.md

@@ -51,6 +51,9 @@ multilingual it expects the sequences in a certain format: A special language id
 source and target text. The source text format is `[lang_code] X [eos]`, where `lang_code` is source language
 id for source text and target language id for target text, with `X` being the source or target text.
 
+> [!NOTE]
+> The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`
+
 The [`M2M100Tokenizer`] depends on `sentencepiece` so be sure to install it before running the
 examples. To install `sentencepiece` run `pip install sentencepiece`.
 

docs/source/en/model_doc/mbart.md

@@ -35,6 +35,9 @@ You can find all the original mBART checkpoints under the [AI at Meta](https://h
 > [!TIP]
 > Click on the mBART models in the right sidebar for more examples of applying mBART to different language tasks.
 
+> [!NOTE]
+> The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`
+
 The example below demonstrates how to translate text with [`Pipeline`] or the [`AutoModel`] class.
 
 <hfoptions id="usage">

docs/source/en/model_doc/mistral.md

@@ -27,7 +27,7 @@ rendered properly in your Markdown viewer.
 
 # Mistral
 
-[Mistral](https://huggingface.co/papers/2310.06825) is a 7B parameter language model, available as a pretrained and instruction-tuned variant, focused on balancing 
+[Mistral](https://huggingface.co/papers/2310.06825) is a 7B parameter language model, available as a pretrained and instruction-tuned variant, focused on balancing
 the scaling costs of large models with performance and efficient inference. This model uses sliding window attention (SWA) trained with a 8K context length and a fixed cache size to handle longer sequences more effectively. Grouped-query attention (GQA) speeds up inference and reduces memory requirements. Mistral also features a byte-fallback BPE tokenizer to improve token handling and efficiency by ensuring characters are never mapped to out-of-vocabulary tokens.
 
 You can find all the original Mistral checkpoints under the [Mistral AI_](https://huggingface.co/mistralai) organization.
@@ -78,10 +78,10 @@ The example below demonstrates how to chat with [`Pipeline`] or the [`AutoModel`
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```python
-echo -e "My favorite condiment is" | transformers-cli chat --model_name_or_path mistralai/Mistral-7B-v0.3 --torch_dtype auto --device 0 --attn_implementation flash_attention_2
+echo -e "My favorite condiment is" | transformers chat mistralai/Mistral-7B-v0.3 --torch_dtype auto --device 0 --attn_implementation flash_attention_2
 ```
 
 </hfoption>

docs/source/en/model_doc/mistral3.md

@@ -227,6 +227,9 @@ This example also how to use `BitsAndBytes` to load the model in 4bit quantizati
 
 [[autodoc]] Mistral3Config
 
+## Mistral3Model
+
+[[autodoc]] Mistral3Model
 
 ## Mistral3ForConditionalGeneration
 

docs/source/en/model_doc/mllama.md

@@ -130,6 +130,10 @@ print(processor.decode(output[0], skip_special_tokens=True))
 [[autodoc]] MllamaTextModel
     - forward
 
+## MllamaModel
+
+[[autodoc]] MllamaModel
+
 ## MllamaForCausalLM
 
 [[autodoc]] MllamaForCausalLM

docs/source/en/model_doc/mobilebert.md

@@ -76,10 +76,10 @@ print(f"The predicted token is: {predicted_token}")
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "The capital of France is [MASK]." | transformers-cli run --task fill-mask --model google/mobilebert-uncased --device 0
+echo -e "The capital of France is [MASK]." | transformers run --task fill-mask --model google/mobilebert-uncased --device 0
 ```
 
 </hfoption>

docs/source/en/model_doc/modernbert.md

@@ -79,10 +79,10 @@ print(f"The predicted token is: {predicted_token}")
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "Plants create [MASK] through a process known as photosynthesis." | transformers-cli run --task fill-mask --model answerdotai/ModernBERT-base --device 0
+echo -e "Plants create [MASK] through a process known as photosynthesis." | transformers run --task fill-mask --model answerdotai/ModernBERT-base --device 0
 ```
 
 </hfoption>

docs/source/en/model_doc/musicgen.md

@@ -62,6 +62,9 @@ python src/transformers/models/musicgen/convert_musicgen_transformers.py \
     --checkpoint small --pytorch_dump_folder /output/path --safe_serialization 
 ```
 
+> [!NOTE]
+> The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`
+
 ## Generation
 
 MusicGen is compatible with two generation modes: greedy and sampling. In practice, sampling leads to significantly

docs/source/en/model_doc/musicgen_melody.md

@@ -44,6 +44,9 @@ There are two key differences with MusicGen:
 1. The audio prompt is used here as a conditional signal for the generated audio sample, whereas it's used for audio continuation in [MusicGen](https://huggingface.co/docs/transformers/main/en/model_doc/musicgen).
 2. Conditional text and audio signals are concatenated to the decoder's hidden states instead of being used as a cross-attention signal, as in MusicGen.
 
+> [!NOTE]
+> The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`
+
 ## Generation
 
 MusicGen Melody is compatible with two generation modes: greedy and sampling. In practice, sampling leads to significantly better results than greedy, thus we encourage sampling mode to be used where possible. Sampling is enabled by default, and can be explicitly specified by setting `do_sample=True` in the call to [`MusicgenMelodyForConditionalGeneration.generate`], or by overriding the model's generation config (see below).

docs/source/en/model_doc/openai-gpt.md

@@ -70,10 +70,10 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "The future of AI is" | transformers-cli run --task text-generation --model openai-community/openai-gpt --device 0
+echo -e "The future of AI is" | transformers run --task text-generation --model openai-community/openai-gpt --device 0
 
 ```
 </hfoption>

docs/source/en/model_doc/opt.md

@@ -41,6 +41,9 @@ Tips:
 - OPT has the same architecture as [`BartDecoder`].
 - Contrary to GPT2, OPT adds the EOS token `</s>` to the beginning of every prompt.
 
+> [!NOTE]
+> The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`
+
 ## Resources
 
 A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with OPT. If you're

docs/source/en/model_doc/paligemma.md

@@ -174,6 +174,10 @@ visualizer("<img> What is in this image?")
 
 [[autodoc]] PaliGemmaProcessor
 
+## PaliGemmaModel
+
+[[autodoc]] PaliGemmaModel
+
 ## PaliGemmaForConditionalGeneration
 
 [[autodoc]] PaliGemmaForConditionalGeneration

docs/source/en/model_doc/phi.md

@@ -65,10 +65,10 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "'''def print_prime(n): """ Print all primes between 1 and n"""'''" | transformers-cli run --task text-classification --model microsoft/phi-1.5 --device 0
+echo -e "'''def print_prime(n): """ Print all primes between 1 and n"""'''" | transformers run --task text-classification --model microsoft/phi-1.5 --device 0
 ```
 
 </hfoption>
@@ -102,7 +102,7 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
     ```py
     import torch
     from transformers import AutoTokenizer, AutoModelForCausalLM
-    
+
     tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-1")
     model = AutoModelForCausalLM.from_pretrained(
         "microsoft/phi-1",
@@ -110,12 +110,12 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
         device_map="auto",
         trust_remote_code=True,
         attn_implementation="sdpa")
-    
+
     input_ids = tokenizer('''def print_prime(n):
        """
        Print all primes between 1 and n
        """''', return_tensors="pt").to("cuda")
-    
+
     output = model.generate(**input_ids, cache_implementation="static")
     print(tokenizer.decode(output[0], skip_special_tokens=True))
     ```

docs/source/en/model_doc/qwen2.md

@@ -64,7 +64,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 
 model = AutoModelForCausalLM.from_pretrained(
     "Qwen/Qwen2-1.5B-Instruct",
-    torch_dtype=torch.bfloat16, 
+    torch_dtype=torch.bfloat16,
     device_map="auto",
     attn_implementation="sdpa"
 )
@@ -86,10 +86,10 @@ generated_ids = model.generate(
     model_inputs.input_ids,
     cache_implementation="static",
     max_new_tokens=512,
-    do_sample=True, 
-    temperature=0.7, 
-    top_k=50,        
-    top_p=0.95       
+    do_sample=True,
+    temperature=0.7,
+    top_k=50,
+    top_p=0.95
 )
 generated_ids = [
     output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
@@ -100,11 +100,11 @@ print(response)
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
 # pip install -U flash-attn --no-build-isolation
-transformers-cli chat --model_name_or_path Qwen/Qwen2-7B-Instruct --torch_dtype auto --attn_implementation flash_attention_2 --device 0
+transformers chat Qwen/Qwen2-7B-Instruct --torch_dtype auto --attn_implementation flash_attention_2 --device 0
 ```
 
 </hfoption>
@@ -121,21 +121,21 @@ from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
 
 quantization_config = BitsAndBytesConfig(
     load_in_4bit=True,
-    bnb_4bit_compute_dtype=torch.bfloat16, 
-    bnb_4bit_quant_type="nf4",             
-    bnb_4bit_use_double_quant=True,       
+    bnb_4bit_compute_dtype=torch.bfloat16,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_use_double_quant=True,
 )
 
-tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-7B") 
+tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-7B")
 model = AutoModelForCausalLM.from_pretrained(
     "Qwen/Qwen2-7B",
     torch_dtype=torch.bfloat16,
     device_map="auto",
     quantization_config=quantization_config,
-    attn_implementation="flash_attention_2" 
+    attn_implementation="flash_attention_2"
 )
 
-inputs = tokenizer("The Qwen2 model family is", return_tensors="pt").to("cuda") 
+inputs = tokenizer("The Qwen2 model family is", return_tensors="pt").to("cuda")
 outputs = model.generate(**inputs, max_new_tokens=100)
 print(tokenizer.decode(outputs[0], skip_special_tokens=True))
 ```

docs/source/en/model_doc/qwen2_5_vl.md

@@ -118,7 +118,7 @@ The example below uses [torchao](../quantization/torchao) to only quantize the w
 
 ```python
 import torch
-from transformers import TorchAoConfig, Gemma3ForConditionalGeneration, AutoProcessor
+from transformers import TorchAoConfig, Qwen2_5_VLForConditionalGeneration, AutoProcessor
 
 quantization_config = TorchAoConfig("int4_weight_only", group_size=128)
 model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
@@ -240,6 +240,10 @@ model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
 
 [[autodoc]] Qwen2_5_VLProcessor
 
+## Qwen2_5_VLTextModel
+
+[[autodoc]] Qwen2_5_VLTextModel
+    - forward
 
 ## Qwen2_5_VLModel
 

docs/source/en/model_doc/qwen2_audio.md

@@ -40,6 +40,9 @@ The abstract from the paper is the following:
 
 `Qwen2-Audio-7B` and `Qwen2-Audio-7B-Instruct` can be found on the [Huggingface Hub](https://huggingface.co/Qwen)
 
+> [!NOTE]
+> The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`
+
 ### Inference
 
 ```python

docs/source/en/model_doc/qwen2_vl.md

@@ -287,6 +287,11 @@ model = Qwen2VLForConditionalGeneration.from_pretrained(
 [[autodoc]] Qwen2VLImageProcessor
     - preprocess
 
+## Qwen2VLVideoProcessor
+
+[[autodoc]] Qwen2VLVideoProcessor
+    - preprocess
+
 ## Qwen2VLImageProcessorFast
 
 [[autodoc]] Qwen2VLImageProcessorFast
@@ -296,6 +301,11 @@ model = Qwen2VLForConditionalGeneration.from_pretrained(
 
 [[autodoc]] Qwen2VLProcessor
 
+## Qwen2VLTextModel
+
+[[autodoc]] Qwen2VLTextModel
+    - forward
+    
 ## Qwen2VLModel
 
 [[autodoc]] Qwen2VLModel

docs/source/en/model_doc/roberta.md

@@ -23,6 +23,7 @@ rendered properly in your Markdown viewer.
 ">
 <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
 </div>
+
 ## Overview
 
 The RoBERTa model was proposed in [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, [Myle Ott](https://huggingface.co/myleott), Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer

docs/source/en/model_doc/sam_hq.md

@@ -0,0 +1,127 @@
+# SAM-HQ
+
+## Overview
+
+SAM-HQ (High-Quality Segment Anything Model) was proposed in [Segment Anything in High Quality](https://arxiv.org/pdf/2306.01567.pdf) by Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu.
+
+The model is an enhancement to the original SAM model that produces significantly higher quality segmentation masks while maintaining SAM's original promptable design, efficiency, and zero-shot generalizability.
+
+![example image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/sam-output.png)
+
+
+SAM-HQ introduces several key improvements over the original SAM model:
+
+1. High-Quality Output Token: A learnable token injected into SAM's mask decoder for higher quality mask prediction
+2. Global-local Feature Fusion: Combines features from different stages of the model for improved mask details
+3. Training Data: Uses a carefully curated dataset of 44K high-quality masks instead of SA-1B
+4. Efficiency: Adds only 0.5% additional parameters while significantly improving mask quality
+5. Zero-shot Capability: Maintains SAM's strong zero-shot performance while improving accuracy
+
+The abstract from the paper is the following:
+
+*The recent Segment Anything Model (SAM) represents a big leap in scaling up segmentation models, allowing for powerful zero-shot capabilities and flexible prompting. Despite being trained with 1.1 billion masks, SAM's mask prediction quality falls short in many cases, particularly when dealing with objects that have intricate structures. We propose HQ-SAM, equipping SAM with the ability to accurately segment any object, while maintaining SAM's original promptable design, efficiency, and zero-shot generalizability. Our careful design reuses and preserves the pre-trained model weights of SAM, while only introducing minimal additional parameters and computation. We design a learnable High-Quality Output Token, which is injected into SAM's mask decoder and is responsible for predicting the high-quality mask. Instead of only applying it on mask-decoder features, we first fuse them with early and final ViT features for improved mask details. To train our introduced learnable parameters, we compose a dataset of 44K fine-grained masks from several sources. HQ-SAM is only trained on the introduced dataset of 44k masks, which takes only 4 hours on 8 GPUs.*
+
+Tips:
+
+- SAM-HQ produces higher quality masks than the original SAM model, particularly for objects with intricate structures and fine details
+- The model predicts binary masks with more accurate boundaries and better handling of thin structures
+- Like SAM, the model performs better with input 2D points and/or input bounding boxes
+- You can prompt multiple points for the same image and predict a single high-quality mask
+- The model maintains SAM's zero-shot generalization capabilities
+- SAM-HQ only adds ~0.5% additional parameters compared to SAM
+- Fine-tuning the model is not supported yet
+
+This model was contributed by [sushmanth](https://huggingface.co/sushmanth).
+The original code can be found [here](https://github.com/SysCV/SAM-HQ).
+
+Below is an example on how to run mask generation given an image and a 2D point:
+
+```python
+import torch
+from PIL import Image
+import requests
+from transformers import SamHQModel, SamHQProcessor
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+model = SamHQModel.from_pretrained("syscv-community/sam-hq-vit-base").to(device)
+processor = SamHQProcessor.from_pretrained("syscv-community/sam-hq-vit-base")
+
+img_url = "https://huggingface.co/ybelkada/segment-anything/resolve/main/assets/car.png"
+raw_image = Image.open(requests.get(img_url, stream=True).raw).convert("RGB")
+input_points = [[[450, 600]]]  # 2D location of a window in the image
+
+inputs = processor(raw_image, input_points=input_points, return_tensors="pt").to(device)
+with torch.no_grad():
+    outputs = model(**inputs)
+
+masks = processor.image_processor.post_process_masks(
+    outputs.pred_masks.cpu(), inputs["original_sizes"].cpu(), inputs["reshaped_input_sizes"].cpu()
+)
+scores = outputs.iou_scores
+```
+
+You can also process your own masks alongside the input images in the processor to be passed to the model:
+
+```python
+import torch
+from PIL import Image
+import requests
+from transformers import SamHQModel, SamHQProcessor
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+model = SamHQModel.from_pretrained("syscv-community/sam-hq-vit-base").to(device)
+processor = SamHQProcessor.from_pretrained("syscv-community/sam-hq-vit-base")
+
+img_url = "https://huggingface.co/ybelkada/segment-anything/resolve/main/assets/car.png"
+raw_image = Image.open(requests.get(img_url, stream=True).raw).convert("RGB")
+mask_url = "https://huggingface.co/ybelkada/segment-anything/resolve/main/assets/car.png"
+segmentation_map = Image.open(requests.get(mask_url, stream=True).raw).convert("1")
+input_points = [[[450, 600]]]  # 2D location of a window in the image
+
+inputs = processor(raw_image, input_points=input_points, segmentation_maps=segmentation_map, return_tensors="pt").to(device)
+with torch.no_grad():
+    outputs = model(**inputs)
+
+masks = processor.image_processor.post_process_masks(
+    outputs.pred_masks.cpu(), inputs["original_sizes"].cpu(), inputs["reshaped_input_sizes"].cpu()
+)
+scores = outputs.iou_scores
+```
+
+
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with SAM-HQ:
+
+- Demo notebook for using the model (coming soon)
+- Paper implementation and code: [SAM-HQ GitHub Repository](https://github.com/SysCV/SAM-HQ)
+
+## SamHQConfig
+
+[[autodoc]] SamHQConfig
+
+## SamHQVisionConfig
+
+[[autodoc]] SamHQVisionConfig
+
+## SamHQMaskDecoderConfig
+
+[[autodoc]] SamHQMaskDecoderConfig
+
+## SamHQPromptEncoderConfig
+
+[[autodoc]] SamHQPromptEncoderConfig
+
+## SamHQProcessor
+
+[[autodoc]] SamHQProcessor
+
+## SamHQVisionModel
+
+[[autodoc]] SamHQVisionModel
+
+
+## SamHQModel
+
+[[autodoc]] SamHQModel
+    - forward

docs/source/en/model_doc/sew.md

@@ -46,6 +46,9 @@ This model was contributed by [anton-l](https://huggingface.co/anton-l).
 - SEWForCTC is fine-tuned using connectionist temporal classification (CTC) so the model output has to be decoded using
   [`Wav2Vec2CTCTokenizer`].
 
+> [!NOTE]
+> The `head_mask` argument is ignored when using all attention implementation other than "eager". If you have a `head_mask` and want it to have effect, load the model with `XXXModel.from_pretrained(model_id, attn_implementation="eager")`
+
 ## Resources
 
 - [Audio classification task guide](../tasks/audio_classification)

docs/source/en/model_doc/smolvlm.md

@@ -197,6 +197,9 @@ print(generated_texts[0])
 [[autodoc]] SmolVLMImageProcessor
     - preprocess
 
+## SmolVLMVideoProcessor
+[[autodoc]] SmolVLMVideoProcessor
+    - preprocess
 
 ## SmolVLMProcessor
 [[autodoc]] SmolVLMProcessor

docs/source/en/model_doc/swin2sr.md

@@ -50,6 +50,11 @@ A demo Space for image super-resolution with SwinSR can be found [here](https://
 [[autodoc]] Swin2SRImageProcessor
     - preprocess
 
+## Swin2SRImageProcessorFast
+
+[[autodoc]] Swin2SRImageProcessorFast
+    - preprocess
+
 ## Swin2SRConfig
 
 [[autodoc]] Swin2SRConfig

docs/source/en/model_doc/t5.md

@@ -75,10 +75,10 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
 </hfoption>
-<hfoption id="transformers-cli">
+<hfoption id="transformers CLI">
 
 ```bash
-echo -e "translate English to French: The weather is nice today." | transformers-cli run --task text2text-generation --model google-t5/t5-base --device 0
+echo -e "translate English to French: The weather is nice today." | transformers run --task text2text-generation --model google-t5/t5-base --device 0
 ```
 
 </hfoption>