Release: v4.34.0

* fix issues with PEFT

* logger warning futurewarning issues

* fixup

* adapt from suggestions

* oops

* rm test

.circleci/create_circleci_config.py

@@ -151,10 +151,13 @@ class CircleCIJob:
         pytest_flags.append(
             f"--make-reports={self.name}" if "examples" in self.name else f"--make-reports=tests_{self.name}"
         )
+
+        steps.append({"run": {"name": "Create `test-results` directory", "command": "mkdir test-results"}})
+
         test_command = ""
         if self.command_timeout:
             test_command = f"timeout {self.command_timeout} "
-        test_command += f"python -m pytest -n {self.pytest_num_workers} " + " ".join(pytest_flags)
+        test_command += f"python -m pytest --junitxml=test-results/junit.xml -n {self.pytest_num_workers} " + " ".join(pytest_flags)
 
         if self.parallelism == 1:
             if self.tests_to_run is None:
@@ -223,18 +226,40 @@ class CircleCIJob:
             # failure.
             test_command = f"({test_command}) || true"
         else:
-            test_command += " | tee tests_output.txt"
+            test_command += " || true"
         steps.append({"run": {"name": "Run tests", "command": test_command}})
 
+        # Deal with errors
+        check_test_command = f'if [ -s reports/{self.job_name}/errors.txt ]; '
+        check_test_command += 'then echo "Some tests errored out!"; echo ""; '
+        check_test_command += f'cat reports/{self.job_name}/errors.txt; '
+        check_test_command += 'echo ""; echo ""; '
+
+        py_command = f'import os; fp = open("reports/{self.job_name}/summary_short.txt"); failed = os.linesep.join([x for x in fp.read().split(os.linesep) if x.startswith("ERROR ")]); fp.close(); fp = open("summary_short.txt", "w"); fp.write(failed); fp.close()'
+        check_test_command += f"$(python3 -c '{py_command}'); "
+        check_test_command += f'cat summary_short.txt; echo ""; exit -1; '
+
+        # Deeal with failed tests
+        check_test_command += f'elif [ -s reports/{self.job_name}/failures_short.txt ]; '
+        check_test_command += 'then echo "Some tests failed!"; echo ""; '
+        check_test_command += f'cat reports/{self.job_name}/failures_short.txt; '
+        check_test_command += 'echo ""; echo ""; '
+
+        py_command = f'import os; fp = open("reports/{self.job_name}/summary_short.txt"); failed = os.linesep.join([x for x in fp.read().split(os.linesep) if x.startswith("FAILED ")]); fp.close(); fp = open("summary_short.txt", "w"); fp.write(failed); fp.close()'
+        check_test_command += f"$(python3 -c '{py_command}'); "
+        check_test_command += f'cat summary_short.txt; echo ""; exit -1; '
+
+        check_test_command += f'elif [ -s reports/{self.job_name}/stats.txt ]; then echo "All tests pass!"; '
+
         # return code `124` means the previous (pytest run) step is timeout
         if self.name == "pr_documentation_tests":
-            checkout_doctest_command = 'if [ -s reports/tests_pr_documentation_tests/failures_short.txt ]; '
-            checkout_doctest_command += 'then echo "some test failed"; '
-            checkout_doctest_command += 'cat reports/tests_pr_documentation_tests/failures_short.txt; '
-            checkout_doctest_command += 'cat reports/tests_pr_documentation_tests/summary_short.txt; exit -1; '
-            checkout_doctest_command += 'elif [ -s reports/tests_pr_documentation_tests/stats.txt ]; then echo "All tests pass!"; '
-            checkout_doctest_command += 'elif [ -f 124.txt ]; then echo "doctest timeout!"; else echo "other fatal error)"; exit -1; fi;'
-            steps.append({"run": {"name": "Check doctest results", "command": checkout_doctest_command}})
+            check_test_command += 'elif [ -f 124.txt ]; then echo "doctest timeout!"; '
+
+        check_test_command += 'else echo "other fatal error"; echo ""; exit -1; fi;'
+
+        steps.append({"run": {"name": "Check test results", "command": check_test_command}})
+
+        steps.append({"store_test_results": {"path": "test-results"}})
 
         steps.append({"store_artifacts": {"path": "~/transformers/tests_output.txt"}})
         steps.append({"store_artifacts": {"path": "~/transformers/reports"}})
@@ -441,13 +466,15 @@ exotic_models_job = CircleCIJob(
         "sudo apt install tesseract-ocr",
         "pip install -U --upgrade-strategy eager pytesseract",
         "pip install -U --upgrade-strategy eager natten",
-        # TODO (ydshieh): Remove this line once `https://github.com/facebookresearch/detectron2/issues/5010` is resolved
-        'pip install -U --upgrade-strategy eager "Pillow<10.0.0"',
+        "pip install -U --upgrade-strategy eager python-Levenshtein",
+        "pip install -U --upgrade-strategy eager opencv-python",
+        "pip install -U --upgrade-strategy eager nltk",
     ],
     tests_to_run=[
         "tests/models/*layoutlmv*",
         "tests/models/*nat",
         "tests/models/deta",
+        "tests/models/nougat",
     ],
     pytest_num_workers=1,
     pytest_options={"durations": 100},
@@ -594,7 +621,7 @@ def create_circleci_config(folder=None):
                 job.tests_to_run = [f"examples/{framework}"]
             else:
                 job.tests_to_run = [f for f in example_tests.split(" ") if f.startswith(f"examples/{framework}")]
-            
+
             if len(job.tests_to_run) > 0:
                 jobs.append(job)
 

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -37,7 +37,7 @@ body:
           - pipelines: @Narsil
           - tensorflow: @gante and @Rocketknight1
           - tokenizers: @ArthurZucker
-          - trainer: @muellerz and @pacman100
+          - trainer: @muellerzr and @pacman100
         
         Integrations:
         

.github/PULL_REQUEST_TEMPLATE.md

@@ -51,7 +51,7 @@ Library:
 - pipelines: @Narsil
 - tensorflow: @gante and @Rocketknight1
 - tokenizers: @ArthurZucker
-- trainer: @muellerz and @pacman100
+- trainer: @muellerzr and @pacman100
 
 Integrations:
 

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

@@ -34,19 +34,19 @@ jobs:
           sudo du -sh /usr/share/
       -
         name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
       -
         name: Check out code
         uses: actions/checkout@v3
       -
         name: Login to DockerHub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
       -
         name: Build and push
-        uses: docker/build-push-action@v3
+        uses: docker/build-push-action@v5
         with:
           context: ./docker/transformers-all-latest-gpu
           build-args: |
@@ -59,7 +59,7 @@ jobs:
         # This condition allows `schedule` events, or `push` events that trigger this workflow NOT via `workflow_call`.
         # The later case is useful for manual image building for debugging purpose. Use another tag in this case!
         if: inputs.image_postfix != '-push-ci'
-        uses: docker/build-push-action@v3
+        uses: docker/build-push-action@v5
         with:
           context: ./docker/transformers-all-latest-gpu
           build-args: |
@@ -83,19 +83,19 @@ jobs:
           sudo du -sh /usr/share/
       -
         name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
       -
         name: Check out code
         uses: actions/checkout@v3
       -
         name: Login to DockerHub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
       -
         name: Build and push
-        uses: docker/build-push-action@v3
+        uses: docker/build-push-action@v5
         with:
           context: ./docker/transformers-pytorch-deepspeed-latest-gpu
           build-args: |
@@ -120,13 +120,13 @@ jobs:
           sudo du -sh /usr/share/
       -
         name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
       -
         name: Check out code
         uses: actions/checkout@v3
       -
         name: Login to DockerHub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
@@ -136,7 +136,7 @@ jobs:
         # This condition allows `schedule` events, or `push` events that trigger this workflow NOT via `workflow_call`.
         # The later case is useful for manual image building for debugging purpose. Use another tag in this case!
         if: inputs.image_postfix != '-push-ci'
-        uses: docker/build-push-action@v3
+        uses: docker/build-push-action@v5
         with:
           context: ./docker/transformers-pytorch-deepspeed-latest-gpu
           build-args: |
@@ -152,19 +152,19 @@ jobs:
     steps:
       -
         name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
       -
         name: Check out code
         uses: actions/checkout@v3
       -
         name: Login to DockerHub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
       -
         name: Build and push
-        uses: docker/build-push-action@v3
+        uses: docker/build-push-action@v5
         with:
           context: ./docker/transformers-doc-builder
           push: true
@@ -188,19 +188,19 @@ jobs:
           sudo du -sh /usr/share/
       -
         name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
       -
         name: Check out code
         uses: actions/checkout@v3
       -
         name: Login to DockerHub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
       -
         name: Build and push
-        uses: docker/build-push-action@v3
+        uses: docker/build-push-action@v5
         with:
           context: ./docker/transformers-pytorch-gpu
           build-args: |
@@ -208,6 +208,41 @@ jobs:
           push: true
           tags: huggingface/transformers-pytorch-gpu
 
+  latest-pytorch-amd:
+    name: "Latest PyTorch (AMD) [dev]"
+    runs-on: [self-hosted, docker-gpu, amd-gpu, single-gpu, mi210]
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Check out code
+        uses: actions/checkout@v3
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+      - name: Build and push
+        uses: docker/build-push-action@v5
+        with:
+          context: ./docker/transformers-pytorch-amd-gpu
+          build-args: |
+            REF=main
+          push: true
+          tags: huggingface/transformers-pytorch-amd-gpu${{ inputs.image_postfix }}
+      # Push CI images still need to be re-built daily
+      -
+        name: Build and push (for Push CI) in a daily basis
+        # This condition allows `schedule` events, or `push` events that trigger this workflow NOT via `workflow_call`.
+        # The later case is useful for manual image building for debugging purpose. Use another tag in this case!
+        if: inputs.image_postfix != '-push-ci'
+        uses: docker/build-push-action@v5
+        with:
+          context: ./docker/transformers-pytorch-amd-gpu
+          build-args: |
+            REF=main
+          push: true
+          tags: huggingface/transformers-pytorch-amd-gpu-push-ci
+
   latest-tensorflow:
     name: "Latest TensorFlow [dev]"
     # Push CI doesn't need this image
@@ -216,19 +251,19 @@ jobs:
     steps:
       -
         name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
       -
         name: Check out code
         uses: actions/checkout@v3
       -
         name: Login to DockerHub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
       -
         name: Build and push
-        uses: docker/build-push-action@v3
+        uses: docker/build-push-action@v5
         with:
           context: ./docker/transformers-tensorflow-gpu
           build-args: |

.github/workflows/doctests.yml

@@ -20,7 +20,7 @@ env:
 
 jobs:
   run_doctests:
-    runs-on: [self-hosted, doc-tests-gpu]
+    runs-on: [single-gpu, nvidia-gpu, t4, doctest-ci]
     container:
       image: huggingface/transformers-all-latest-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/

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

@@ -39,7 +39,7 @@ jobs:
     strategy:
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, past-ci]
     container:
       image: huggingface/transformers-all-latest-torch-nightly-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -54,7 +54,7 @@ jobs:
     strategy:
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, past-ci]
     container:
       image: huggingface/transformers-all-latest-torch-nightly-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -94,7 +94,7 @@ jobs:
       matrix:
         folders: ${{ fromJson(needs.setup.outputs.matrix) }}
         machine_type: [single-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, past-ci]
     container:
       image: huggingface/transformers-all-latest-torch-nightly-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -155,7 +155,7 @@ jobs:
       matrix:
         folders: ${{ fromJson(needs.setup.outputs.matrix) }}
         machine_type: [multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, past-ci]
     container:
       image: huggingface/transformers-all-latest-torch-nightly-gpu
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -215,7 +215,7 @@ jobs:
       fail-fast: false
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, past-ci]
     needs: setup
     container:
       image: huggingface/transformers-pytorch-deepspeed-nightly-gpu

.github/workflows/self-past.yml

@@ -50,7 +50,7 @@ jobs:
     strategy:
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, past-ci]
     container:
       image: huggingface/transformers-${{ inputs.framework }}-past-${{ inputs.version }}-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -65,7 +65,7 @@ jobs:
     strategy:
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, past-ci]
     container:
       image: huggingface/transformers-${{ inputs.framework }}-past-${{ inputs.version }}-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -101,7 +101,7 @@ jobs:
       matrix:
         folders: ${{ fromJson(needs.setup.outputs.matrix) }}
         machine_type: [single-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, past-ci]
     container:
       image: huggingface/transformers-${{ inputs.framework }}-past-${{ inputs.version }}-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -177,7 +177,7 @@ jobs:
       matrix:
         folders: ${{ fromJson(needs.setup.outputs.matrix) }}
         machine_type: [multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, past-ci]
     container:
       image: huggingface/transformers-${{ inputs.framework }}-past-${{ inputs.version }}-gpu
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -253,7 +253,7 @@ jobs:
       fail-fast: false
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, past-ci]
     needs: setup
     container:
       image: huggingface/transformers-${{ inputs.framework }}-past-${{ inputs.version }}-gpu

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

@@ -0,0 +1,337 @@
+name: Self-hosted runner AMD GPU (push)
+
+on:
+  workflow_run:
+    workflows: ["Self-hosted runner (push-caller)"]
+    branches: ["main"]
+    types: [completed]
+  push:
+    branches:
+      - ci_*
+      - ci-*
+    paths:
+      - "src/**"
+      - "tests/**"
+      - ".github/**"
+      - "templates/**"
+      - "utils/**"
+  repository_dispatch:
+
+env:
+  HF_HOME: /mnt/cache
+  TRANSFORMERS_IS_CI: yes
+  OMP_NUM_THREADS: 8
+  MKL_NUM_THREADS: 8
+  PYTEST_TIMEOUT: 60
+  TF_FORCE_GPU_ALLOW_GROWTH: true
+  RUN_PT_TF_CROSS_TESTS: 1
+
+jobs:
+  check_runner_status:
+    name: Check Runner Status
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout transformers
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 2
+
+      - name: Check Runner Status
+        run: python utils/check_self_hosted_runner.py --target_runners amd-mi210-single-gpu-ci-runner-docker --token ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
+
+  check_runners:
+    name: Check Runners
+    needs: check_runner_status
+    strategy:
+      matrix:
+        machine_type: [single-gpu, multi-gpu]
+        gpu_flavor: [mi210]
+    runs-on: [self-hosted, docker-gpu, amd-gpu, '${{ matrix.machine_type }}', '${{ matrix.gpu_flavor }}']
+    container:
+      image: huggingface/transformers-pytorch-amd-gpu-push-ci  # <--- We test only for PyTorch for now
+      options: --device /dev/kfd --device /dev/dri --env HIP_VISIBLE_DEVICES --env ROCR_VISIBLE_DEVICES --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    steps:
+      - name: ROCM-SMI
+        run: |
+          rocminfo  | grep "Agent" -A 14
+      - name: Show HIP environment
+        run: |
+          echo "HIP: $HIP_VISIBLE_DEVICES"
+          echo "ROCR: $ROCR_VISIBLE_DEVICES"
+
+  setup_gpu:
+    name: Setup
+    needs: check_runners
+    strategy:
+      matrix:
+        machine_type: [single-gpu, multi-gpu]
+        gpu_flavor: [mi210]
+    runs-on: [self-hosted, docker-gpu, amd-gpu, '${{ matrix.machine_type }}', '${{ matrix.gpu_flavor }}']
+    container:
+      image: huggingface/transformers-pytorch-amd-gpu-push-ci  # <--- We test only for PyTorch for now
+      options: --device /dev/kfd --device /dev/dri --env HIP_VISIBLE_DEVICES --env ROCR_VISIBLE_DEVICES --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    outputs:
+      matrix: ${{ steps.set-matrix.outputs.matrix }}
+      test_map: ${{ steps.set-matrix.outputs.test_map }}
+    steps:
+      # Necessary to get the correct branch name and commit SHA for `workflow_run` event
+      # We also take into account the `push` event (we might want to test some changes in a branch)
+      - name: Prepare custom environment variables
+        shell: bash
+        # `CI_BRANCH_PUSH`: The branch name from the push event
+        # `CI_BRANCH_WORKFLOW_RUN`: The name of the branch on which this workflow is triggered by `workflow_run` event
+        # `CI_BRANCH`: The non-empty branch name from the above two (one and only one of them is empty)
+        # `CI_SHA_PUSH`: The commit SHA from the push event
+        # `CI_SHA_WORKFLOW_RUN`: The commit SHA that triggers this workflow by `workflow_run` event
+        # `CI_SHA`: The non-empty commit SHA from the above two (one and only one of them is empty)
+        run: |
+          CI_BRANCH_PUSH=${{ github.event.ref }}
+          CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
+          CI_BRANCH_WORKFLOW_RUN=${{ github.event.workflow_run.head_branch }}
+          CI_SHA_PUSH=${{ github.event.head_commit.id }}
+          CI_SHA_WORKFLOW_RUN=${{ github.event.workflow_run.head_sha }}
+          echo $CI_BRANCH_PUSH
+          echo $CI_BRANCH_WORKFLOW_RUN
+          echo $CI_SHA_PUSH
+          echo $CI_SHA_WORKFLOW_RUN
+          [[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
+          [[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
+
+      - name: print environment variables
+        run: |
+          echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
+          echo "env.CI_SHA = ${{ env.CI_SHA }}"
+
+      - name: Update clone using environment variables
+        working-directory: /transformers
+        run: |
+          echo "original branch = $(git branch --show-current)"
+          git fetch && git checkout ${{ env.CI_BRANCH }}
+          echo "updated branch = $(git branch --show-current)"
+          git checkout ${{ env.CI_SHA }}
+          echo "log = $(git log -n 1)"
+
+      - name: Cleanup
+        working-directory: /transformers
+        run: |
+          rm -rf tests/__pycache__
+          rm -rf tests/models/__pycache__
+          rm -rf reports
+
+      - name: Show installed libraries and their versions
+        working-directory: /transformers
+        run: pip freeze
+
+      - name: Fetch the tests to run
+        working-directory: /transformers
+        # TODO: add `git-python` in the docker images
+        run: |
+          pip install --upgrade git-python
+          python3 utils/tests_fetcher.py --diff_with_last_commit | tee test_preparation.txt
+
+      - name: Report fetched tests
+        uses: actions/upload-artifact@v3
+        with:
+          name: test_fetched
+          path: /transformers/test_preparation.txt
+
+      - id: set-matrix
+        name: Organize tests into models
+        working-directory: /transformers
+        # The `keys` is used as GitHub actions matrix for jobs, i.e. `models/bert`, `tokenization`, `pipeline`, etc.
+        # The `test_map` is used to get the actual identified test files under each key.
+        # If no test to run (so no `test_map.json` file), create a dummy map (empty matrix will fail)
+        run: |
+          if [ -f test_map.json ]; then
+              keys=$(python3 -c 'import json; fp = open("test_map.json"); test_map = json.load(fp); fp.close(); d = list(test_map.keys()); print(d)')
+              test_map=$(python3 -c 'import json; fp = open("test_map.json"); test_map = json.load(fp); fp.close(); print(test_map)')
+          else
+              keys=$(python3 -c 'keys = ["dummy"]; print(keys)')
+              test_map=$(python3 -c 'test_map = {"dummy": []}; print(test_map)')
+          fi
+          echo $keys
+          echo $test_map
+          echo "matrix=$keys" >> $GITHUB_OUTPUT
+          echo "test_map=$test_map" >> $GITHUB_OUTPUT
+
+  run_tests_amdgpu:
+    name: Model tests
+    needs: setup_gpu
+    # `dummy` means there is no test to run
+    if: contains(fromJson(needs.setup_gpu.outputs.matrix), 'dummy') != true
+    strategy:
+      fail-fast: false
+      matrix:
+        folders: ${{ fromJson(needs.setup_gpu.outputs.matrix) }}
+        machine_type: [single-gpu, multi-gpu]
+        gpu_flavor: [mi210]
+    runs-on: [self-hosted, docker-gpu, amd-gpu, '${{ matrix.machine_type }}', '${{ matrix.gpu_flavor }}']
+    container:
+      image: huggingface/transformers-pytorch-amd-gpu-push-ci  # <--- We test only for PyTorch for now
+      options: --device /dev/kfd --device /dev/dri --env HIP_VISIBLE_DEVICES --env ROCR_VISIBLE_DEVICES --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    steps:
+      # Necessary to get the correct branch name and commit SHA for `workflow_run` event
+      # We also take into account the `push` event (we might want to test some changes in a branch)
+      - name: Prepare custom environment variables
+        shell: bash
+        # For the meaning of these environment variables, see the job `Setup`
+        run: |
+          CI_BRANCH_PUSH=${{ github.event.ref }}
+          CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
+          CI_BRANCH_WORKFLOW_RUN=${{ github.event.workflow_run.head_branch }}
+          CI_SHA_PUSH=${{ github.event.head_commit.id }}
+          CI_SHA_WORKFLOW_RUN=${{ github.event.workflow_run.head_sha }}
+          echo $CI_BRANCH_PUSH
+          echo $CI_BRANCH_WORKFLOW_RUN
+          echo $CI_SHA_PUSH
+          echo $CI_SHA_WORKFLOW_RUN
+          [[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
+          [[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
+
+      - name: print environment variables
+        run: |
+          echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
+          echo "env.CI_SHA = ${{ env.CI_SHA }}"
+
+      - name: Update clone using environment variables
+        working-directory: /transformers
+        run: |
+          echo "original branch = $(git branch --show-current)"
+          git fetch && git checkout ${{ env.CI_BRANCH }}
+          echo "updated branch = $(git branch --show-current)"
+          git checkout ${{ env.CI_SHA }}
+          echo "log = $(git log -n 1)"
+
+      - name: Reinstall transformers in edit mode (remove the one installed during docker image build)
+        working-directory: /transformers
+        run: python3 -m pip uninstall -y transformers && python3 -m pip install -e .
+
+      - name: Echo folder ${{ matrix.folders }}
+        shell: bash
+        # For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
+        # set the artifact folder names (because the character `/` is not allowed).
+        run: |
+          echo "${{ matrix.folders }}"
+          echo "${{ fromJson(needs.setup_gpu.outputs.test_map)[matrix.folders] }}"
+          matrix_folders=${{ matrix.folders }}
+          matrix_folders=${matrix_folders/'models/'/'models_'}
+          echo "$matrix_folders"
+          echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
+
+      - name: ROCM-SMI
+        run: |
+          rocminfo  | grep "Agent" -A 14
+      - name: Show HIP environment
+        run: |
+          echo "HIP: $HIP_VISIBLE_DEVICES"
+          echo "ROCR: $ROCR_VISIBLE_DEVICES"
+
+      - name: Environment
+        working-directory: /transformers
+        run: |
+          python3 utils/print_env.py
+
+      - name: Show installed libraries and their versions
+        working-directory: /transformers
+        run: pip freeze
+
+      - name: Run all non-slow selected tests on GPU
+        working-directory: /transformers
+        run: |
+          python3 -m pytest -n 2 --dist=loadfile -v --make-reports=${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }} ${{ fromJson(needs.setup_gpu.outputs.test_map)[matrix.folders] }}
+
+      - name: Failure short reports
+        if: ${{ failure() }}
+        continue-on-error: true
+        run: cat /transformers/reports/${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }}/failures_short.txt
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v3
+        with:
+          name: ${{ matrix.machine_type }}_run_all_tests_gpu_${{ env.matrix_folders }}_test_reports
+          path: /transformers/reports/${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }}
+
+  send_results:
+    name: Send results to webhook
+    runs-on: ubuntu-latest
+    if: always()
+    needs: [
+        check_runner_status,
+        check_runners,
+        setup_gpu,
+        run_tests_amdgpu,
+#        run_tests_torch_cuda_extensions_single_gpu,
+#        run_tests_torch_cuda_extensions_multi_gpu
+    ]
+    steps:
+      - name: Preliminary job status
+        shell: bash
+        # For the meaning of these environment variables, see the job `Setup`
+        run: |
+          echo "Runner availability: ${{ needs.check_runner_status.result }}"
+          echo "Setup status: ${{ needs.setup_gpu.result }}"
+          echo "Runner status: ${{ needs.check_runners.result }}"
+
+      # Necessary to get the correct branch name and commit SHA for `workflow_run` event
+      # We also take into account the `push` event (we might want to test some changes in a branch)
+      - name: Prepare custom environment variables
+        shell: bash
+        # For the meaning of these environment variables, see the job `Setup`
+        run: |
+          CI_BRANCH_PUSH=${{ github.event.ref }}
+          CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
+          CI_BRANCH_WORKFLOW_RUN=${{ github.event.workflow_run.head_branch }}
+          CI_SHA_PUSH=${{ github.event.head_commit.id }}
+          CI_SHA_WORKFLOW_RUN=${{ github.event.workflow_run.head_sha }}
+          echo $CI_BRANCH_PUSH
+          echo $CI_BRANCH_WORKFLOW_RUN
+          echo $CI_SHA_PUSH
+          echo $CI_SHA_WORKFLOW_RUN
+          [[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
+          [[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
+
+      - name: print environment variables
+        run: |
+          echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
+          echo "env.CI_SHA = ${{ env.CI_SHA }}"
+
+      - uses: actions/checkout@v3
+        # To avoid failure when multiple commits are merged into `main` in a short period of time.
+        # Checking out to an old commit beyond the fetch depth will get an error `fatal: reference is not a tree: ...
+        # (Only required for `workflow_run` event, where we get the latest HEAD on `main` instead of the event commit)
+        with:
+          fetch-depth: 20
+
+      - name: Update clone using environment variables
+        run: |
+          echo "original branch = $(git branch --show-current)"
+          git fetch && git checkout ${{ env.CI_BRANCH }}
+          echo "updated branch = $(git branch --show-current)"
+          git checkout ${{ env.CI_SHA }}
+          echo "log = $(git log -n 1)"
+
+      - uses: actions/download-artifact@v3
+      - name: Send message to Slack
+        env:
+          CI_SLACK_BOT_TOKEN: ${{ secrets.CI_SLACK_BOT_TOKEN }}
+          CI_SLACK_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID }}
+          CI_SLACK_CHANNEL_ID_DAILY: ${{ secrets.CI_SLACK_CHANNEL_ID_DAILY }}
+          CI_SLACK_CHANNEL_ID_AMD: ${{ secrets.CI_SLACK_CHANNEL_ID_AMD }}
+          CI_SLACK_CHANNEL_DUMMY_TESTS: ${{ secrets.CI_SLACK_CHANNEL_DUMMY_TESTS }}
+          CI_SLACK_REPORT_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID_AMD }}
+          ACCESS_REPO_INFO_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
+          CI_EVENT: push
+          CI_TITLE_PUSH: ${{ github.event.head_commit.message }}
+          CI_TITLE_WORKFLOW_RUN: ${{ github.event.workflow_run.head_commit.message }}
+          CI_SHA: ${{ env.CI_SHA }}
+          RUNNER_STATUS: ${{ needs.check_runner_status.result }}
+          RUNNER_ENV_STATUS: ${{ needs.check_runners.result }}
+          SETUP_STATUS: ${{ needs.setup_gpu.result }}
+
+        # We pass `needs.setup_gpu.outputs.matrix` as the argument. A processing in `notification_service.py` to change
+        # `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
+        run: |
+          pip install slack_sdk
+          pip show slack_sdk
+          python utils/notification_service.py "${{ needs.setup_gpu.outputs.matrix }}"

.github/workflows/self-push.yml

@@ -45,7 +45,7 @@ jobs:
     strategy:
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: [self-hosted, docker-gpu, '${{ matrix.machine_type }}']
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, push-ci]
     container:
       image: huggingface/transformers-all-latest-gpu-push-ci
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -60,7 +60,7 @@ jobs:
     strategy:
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: [self-hosted, docker-gpu, '${{ matrix.machine_type }}']
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, push-ci]
     container:
       image: huggingface/transformers-all-latest-gpu-push-ci
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -158,7 +158,7 @@ jobs:
       matrix:
         folders: ${{ fromJson(needs.setup.outputs.matrix) }}
         machine_type: [single-gpu]
-    runs-on: [self-hosted, docker-gpu, '${{ matrix.machine_type }}']
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, push-ci]
     container:
       image: huggingface/transformers-all-latest-gpu-push-ci
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -251,7 +251,7 @@ jobs:
       matrix:
         folders: ${{ fromJson(needs.setup.outputs.matrix) }}
         machine_type: [multi-gpu]
-    runs-on: [self-hosted, docker-gpu, '${{ matrix.machine_type }}']
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, push-ci]
     container:
       image: huggingface/transformers-all-latest-gpu-push-ci
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -344,7 +344,7 @@ jobs:
       fail-fast: false
       matrix:
         machine_type: [single-gpu]
-    runs-on: [self-hosted, docker-gpu, '${{ matrix.machine_type }}']
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, push-ci]
     container:
       image: huggingface/transformers-pytorch-deepspeed-latest-gpu-push-ci
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -434,7 +434,7 @@ jobs:
       fail-fast: false
       matrix:
         machine_type: [multi-gpu]
-    runs-on: [self-hosted, docker-gpu, '${{ matrix.machine_type }}']
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, push-ci]
     container:
       image: huggingface/transformers-pytorch-deepspeed-latest-gpu-push-ci
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/

.github/workflows/self-scheduled.yml

@@ -43,7 +43,7 @@ jobs:
     strategy:
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, daily-ci]
     container:
       image: huggingface/transformers-all-latest-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -58,7 +58,7 @@ jobs:
     strategy:
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, daily-ci]
     container:
       image: huggingface/transformers-all-latest-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -98,7 +98,7 @@ jobs:
       matrix:
         folders: ${{ fromJson(needs.setup.outputs.matrix) }}
         machine_type: [single-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, daily-ci]
     container:
       image: huggingface/transformers-all-latest-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -159,7 +159,7 @@ jobs:
       matrix:
         folders: ${{ fromJson(needs.setup.outputs.matrix) }}
         machine_type: [multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, daily-ci]
     container:
       image: huggingface/transformers-all-latest-gpu
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -219,7 +219,7 @@ jobs:
       fail-fast: false
       matrix:
         machine_type: [single-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, daily-ci]
     container:
       image: huggingface/transformers-all-latest-gpu
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -270,7 +270,7 @@ jobs:
       fail-fast: false
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, daily-ci]
     container:
       image: huggingface/transformers-pytorch-gpu
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -320,7 +320,7 @@ jobs:
       fail-fast: false
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, daily-ci]
     container:
       image: huggingface/transformers-tensorflow-gpu
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -371,7 +371,7 @@ jobs:
       fail-fast: false
       matrix:
         machine_type: [single-gpu, multi-gpu]
-    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    runs-on: ['${{ matrix.machine_type }}', nvidia-gpu, t4, daily-ci]
     needs: setup
     container:
       image: huggingface/transformers-pytorch-deepspeed-latest-gpu

.gitignore

@@ -166,4 +166,4 @@ tags
 .DS_Store
 
 # ruff
-.ruff_cache
+.ruff_cache

README.md

@@ -51,8 +51,9 @@ limitations under the License.
         <a href="https://github.com/huggingface/transformers/blob/main/README_ko.md">한국어</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_es.md">Español</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_ja.md">日本語</a> |
-        <a href="https://github.com/huggingface/transformers/blob/main/README_hd.md">हिन्दी</a>
-    <p>
+        <a href="https://github.com/huggingface/transformers/blob/main/README_hd.md">हिन्दी</a> |
+        <a href="https://github.com/huggingface/transformers/blob/main/README_ru.md">Русский</a>
+    </p>
 </h4>
 
 <h3 align="center">
@@ -310,6 +311,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (from NAVER CLOVA) released with the paper [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
@@ -373,6 +375,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (from Allegro.pl, AGH University of Science and Technology) released with the paper [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) by Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
 1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh.
@@ -406,6 +409,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Megatron-BERT](https://huggingface.co/docs/transformers/model_doc/megatron-bert)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
 1. **[Megatron-GPT2](https://huggingface.co/docs/transformers/model_doc/megatron_gpt2)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
 1. **[MGP-STR](https://huggingface.co/docs/transformers/model_doc/mgp-str)** (from Alibaba Research) released with the paper [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592) by Peng Wang, Cheng Da, and Cong Yao.
+1. **[Mistral](https://huggingface.co/docs/transformers/model_doc/mistral)** (from Mistral AI) by The [Mistral AI](https://mistral.ai) team: Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.
 1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
 1. **[MMS](https://huggingface.co/docs/transformers/model_doc/mms)** (from Facebook) released with the paper [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516) by Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli.
 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
@@ -423,6 +427,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Nezha](https://huggingface.co/docs/transformers/model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu.
 1. **[NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
 1. **[NLLB-MOE](https://huggingface.co/docs/transformers/model_doc/nllb-moe)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
+1. **[Nougat](https://huggingface.co/docs/transformers/model_doc/nougat)** (from Meta AI) released with the paper [Nougat: Neural Optical Understanding for Academic Documents](https://arxiv.org/abs/2308.13418) by Lukas Blecher, Guillem Cucurull, Thomas Scialom, Robert Stojnic.
 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
 1. **[OneFormer](https://huggingface.co/docs/transformers/model_doc/oneformer)** (from SHI Labs) released with the paper [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) by Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi.
 1. **[OpenLlama](https://huggingface.co/docs/transformers/model_doc/open-llama)** (from [s-JoL](https://huggingface.co/s-JoL)) released in [Open-Llama](https://github.com/s-JoL/Open-Llama).
@@ -431,11 +436,12 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, and Peter J. Liu.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
+1. **[Persimmon](https://huggingface.co/docs/transformers/model_doc/persimmon)** (from ADEPT) released in a [blog post](https://www.adept.ai/blog/persimmon-8b) by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
-1. **[Pop2Piano](https://huggingface.co/docs/transformers/main/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi and Kyogu Lee.
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi and Kyogu Lee.
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
 1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
@@ -486,10 +492,11 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
-1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
+1. **[ViTMatte](https://huggingface.co/docs/transformers/model_doc/vitmatte)** (from HUST-VL) rreleased with the paper [ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers](https://arxiv.org/abs/2305.15272) by Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
-1. **[VITS](https://huggingface.co/docs/transformers/main/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.

README_es.md

@@ -18,7 +18,7 @@ limitations under the License.
     <br>
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers_logo_name.png" width="400"/>
     <br>
-<p>
+</p>
 <p align="center">
     <a href="https://circleci.com/gh/huggingface/transformers">
         <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/main">
@@ -47,7 +47,7 @@ limitations under the License.
         <b>Español</b> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_ja.md">日本語</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_hd.md">हिन्दी</a>
-    <p>
+    </p>
 </h4>
 
 <h3 align="center">
@@ -287,6 +287,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (from NAVER CLOVA) released with the paper [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
@@ -350,6 +351,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (from Allegro.pl, AGH University of Science and Technology) released with the paper [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) by Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
 1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
@@ -383,6 +385,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[Megatron-BERT](https://huggingface.co/docs/transformers/model_doc/megatron-bert)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
 1. **[Megatron-GPT2](https://huggingface.co/docs/transformers/model_doc/megatron_gpt2)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
 1. **[MGP-STR](https://huggingface.co/docs/transformers/model_doc/mgp-str)** (from Alibaba Research) released with the paper [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592) by Peng Wang, Cheng Da, and Cong Yao.
+1. **[Mistral](https://huggingface.co/docs/transformers/model_doc/mistral)** (from Mistral AI) by The Mistral AI team: Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.. 
 1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
 1. **[MMS](https://huggingface.co/docs/transformers/model_doc/mms)** (from Facebook) released with the paper [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516) by Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli.
 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
@@ -400,6 +403,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[Nezha](https://huggingface.co/docs/transformers/model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu.
 1. **[NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
 1. **[NLLB-MOE](https://huggingface.co/docs/transformers/model_doc/nllb-moe)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
+1. **[Nougat](https://huggingface.co/docs/transformers/model_doc/nougat)** (from Meta AI) released with the paper [Nougat: Neural Optical Understanding for Academic Documents](https://arxiv.org/abs/2308.13418) by Lukas Blecher, Guillem Cucurull, Thomas Scialom, Robert Stojnic.
 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
 1. **[OneFormer](https://huggingface.co/docs/transformers/model_doc/oneformer)** (from SHI Labs) released with the paper [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) by Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi.
 1. **[OpenLlama](https://huggingface.co/docs/transformers/model_doc/open-llama)** (from [s-JoL](https://huggingface.co/s-JoL)) released in [Open-Llama](https://github.com/s-JoL/Open-Llama).
@@ -408,11 +412,12 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, and Peter J. Liu.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
+1. **[Persimmon](https://huggingface.co/docs/transformers/model_doc/persimmon)** (from ADEPT) released with the paper [blog post](https://www.adept.ai/blog/persimmon-8b) by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
-1. **[Pop2Piano](https://huggingface.co/docs/transformers/main/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
 1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
@@ -463,10 +468,11 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
-1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
+1. **[ViTMatte](https://huggingface.co/docs/transformers/model_doc/vitmatte)** (from HUST-VL) released with the paper [ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers](https://arxiv.org/abs/2305.15272) by Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
-1. **[VITS](https://huggingface.co/docs/transformers/main/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.

README_hd.md

@@ -43,7 +43,7 @@ checkpoint: जाँच बिंदु
     <br>
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers_logo_name.png" width="400"/>
     <br>
-<p>
+</p>
 <p align="center">
     <a href="https://circleci.com/gh/huggingface/transformers">
         <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/main">
@@ -72,7 +72,7 @@ checkpoint: जाँच बिंदु
         <a href="https://github.com/huggingface/transformers/blob/main/README_es.md">Español</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_ja.md">日本語</a> |
         <b>हिन्दी</b> |
-    <p>
+    </p>
 </h4>
 
 <h3 align="center">
@@ -259,6 +259,7 @@ conda install -c huggingface transformers
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigSicence Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (एलेक्सा से) कागज के साथ [बीईआरटी के लिए ऑप्टिमल सबआर्किटेक्चर एक्सट्रैक्शन](https://arxiv.org/abs/ 2010.10499) एड्रियन डी विंटर और डैनियल जे पेरी द्वारा।
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (हरबिन इंस्टिट्यूट ऑफ़ टेक्नोलॉजी/माइक्रोसॉफ्ट रिसर्च एशिया/इंटेल लैब्स से) कागज के साथ [ब्रिजटॉवर: विजन-लैंग्वेज रिप्रेजेंटेशन लर्निंग में एनकोडर्स के बीच ब्रिज बनाना](<https://arxiv.org/abs/2206.08657>) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (NAVER CLOVA से) Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park. द्वाराअनुसंधान पत्र [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) के साथ जारी किया गया
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (Google अनुसंधान से) साथ में कागज [ByT5: पूर्व-प्रशिक्षित बाइट-टू-बाइट मॉडल के साथ एक टोकन-मुक्त भविष्य की ओर] (https://arxiv.org/abs/2105.13626) Linting Xue, Aditya Barua, Noah Constant, रामी अल-रफू, शरण नारंग, मिहिर काले, एडम रॉबर्ट्स, कॉलिन रैफेल द्वारा पोस्ट किया गया।
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (इनरिया/फेसबुक/सोरबोन से) साथ में कागज [CamemBERT: एक टेस्टी फ्रेंच लैंग्वेज मॉडल](https:// arxiv.org/abs/1911.03894) लुई मार्टिन*, बेंजामिन मुलर*, पेड्रो जेवियर ऑर्टिज़ सुआरेज़*, योआन ड्यूपॉन्ट, लॉरेंट रोमरी, एरिक विलेमोन्टे डे ला क्लर्जरी, जैमे सेडाह और बेनोइट सगोट द्वारा।
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (Google रिसर्च से) साथ में दिया गया पेपर [कैनाइन: प्री-ट्रेनिंग ए एफिशिएंट टोकनाइजेशन-फ्री एनकोडर फॉर लैंग्वेज रिप्रेजेंटेशन]( https://arxiv.org/abs/2103.06874) जोनाथन एच क्लार्क, डैन गैरेट, यूलिया टर्क, जॉन विएटिंग द्वारा।
@@ -322,6 +323,7 @@ conda install -c huggingface transformers
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA से) साथ में कागज [GroupViT: टेक्स्ट सुपरविजन से सिमेंटिक सेगमेंटेशन इमर्जेस](https://arxiv .org/abs/2202.11094) जियारुई जू, शालिनी डी मेलो, सिफ़ी लियू, वोनमिन बायन, थॉमस ब्रेउएल, जान कौट्ज़, ज़ियाओलोंग वांग द्वारा।
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (Allegro.pl, AGH University of Science and Technology से) Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik. द्वाराअनुसंधान पत्र [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) के साथ जारी किया गया
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (फेसबुक से) साथ में पेपर [ह्यूबर्ट: सेल्फ सुपरवाइज्ड स्पीच रिप्रेजेंटेशन लर्निंग बाय मास्क्ड प्रेडिक्शन ऑफ हिडन यूनिट्स](https ://arxiv.org/abs/2106.07447) वेई-निंग सू, बेंजामिन बोल्टे, याओ-हंग ह्यूबर्ट त्साई, कुशाल लखोटिया, रुस्लान सालाखुतदीनोव, अब्देलरहमान मोहम्मद द्वारा।
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (बर्कले से) साथ में कागज [I-BERT: Integer-only BERT Quantization](https:// arxiv.org/abs/2101.01321) सेहून किम, अमीर घोलमी, ज़ेवेई याओ, माइकल डब्ल्यू महोनी, कर्ट केटज़र द्वारा।
 1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
@@ -355,6 +357,7 @@ conda install -c huggingface transformers
 1. **[Megatron-BERT](https://huggingface.co/docs/transformers/model_doc/megatron-bert)** (NVIDIA से) कागज के साथ [Megatron-LM: मॉडल का उपयोग करके बहु-अरब पैरामीटर भाषा मॉडल का प्रशिक्षण Parallelism](https://arxiv.org/abs/1909.08053) मोहम्मद शोएबी, मोस्टोफा पटवारी, राउल पुरी, पैट्रिक लेग्रेस्ले, जेरेड कैस्पर और ब्रायन कैटानज़ारो द्वारा।
 1. **[Megatron-GPT2](https://huggingface.co/docs/transformers/model_doc/megatron_gpt2)** (NVIDIA से) साथ वाला पेपर [Megatron-LM: ट्रेनिंग मल्टी-बिलियन पैरामीटर लैंग्वेज मॉडल्स यूजिंग मॉडल पैरेललिज़्म] (https://arxiv.org/abs/1909.08053) मोहम्मद शोएबी, मोस्टोफा पटवारी, राउल पुरी, पैट्रिक लेग्रेस्ले, जेरेड कैस्पर और ब्रायन कैटानज़ारो द्वारा पोस्ट किया गया।
 1. **[MGP-STR](https://huggingface.co/docs/transformers/model_doc/mgp-str)** (Alibaba Research से) Peng Wang, Cheng Da, and Cong Yao. द्वाराअनुसंधान पत्र [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592) के साथ जारी किया गया
+1. **[Mistral](https://huggingface.co/docs/transformers/model_doc/mistral)** (from Mistral AI) by The Mistral AI team: Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.. 
 1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (फ्रॉम Studio Ousia) साथ में पेपर [mLUKE: द पावर ऑफ एंटिटी रिप्रेजेंटेशन इन मल्टीलिंगुअल प्रीट्रेन्ड लैंग्वेज मॉडल्स](https://arxiv.org/abs/2110.08151) रयोकन री, इकुया यामाडा, और योशिमासा त्सुरोका द्वारा।
 1. **[MMS](https://huggingface.co/docs/transformers/model_doc/mms)** (Facebook से) Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli. द्वाराअनुसंधान पत्र [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516) के साथ जारी किया गया
 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (सीएमयू/गूगल ब्रेन से) साथ में कागज [मोबाइलबर्ट: संसाधन-सीमित उपकरणों के लिए एक कॉम्पैक्ट टास्क-अज्ञेय बीईआरटी] (https://arxiv.org/abs/2004.02984) Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, और Denny Zhou द्वारा पोस्ट किया गया।
@@ -372,6 +375,7 @@ conda install -c huggingface transformers
 1. **[Nezha](https://huggingface.co/docs/transformers/model_doc/nezha)** (हुआवेई नूह के आर्क लैब से) साथ में कागज़ [NEZHA: चीनी भाषा समझ के लिए तंत्रिका प्रासंगिक प्रतिनिधित्व](https :/ /arxiv.org/abs/1909.00204) जुन्किउ वेई, ज़ियाओज़े रेन, ज़िआओगुआंग ली, वेनयोंग हुआंग, यी लियाओ, याशेंग वांग, जियाशू लिन, शिन जियांग, जिओ चेन और कुन लियू द्वारा।
 1. **[NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)** (फ्रॉम मेटा) साथ में पेपर [नो लैंग्वेज लेफ्ट बिहाइंड: स्केलिंग ह्यूमन-सेंटेड मशीन ट्रांसलेशन] (https://arxiv.org/abs/2207.04672) एनएलएलबी टीम द्वारा प्रकाशित।
 1. **[NLLB-MOE](https://huggingface.co/docs/transformers/model_doc/nllb-moe)** (Meta से) the NLLB team. द्वाराअनुसंधान पत्र [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) के साथ जारी किया गया
+1. **[Nougat](https://huggingface.co/docs/transformers/model_doc/nougat)** (Meta AI से) Lukas Blecher, Guillem Cucurull, Thomas Scialom, Robert Stojnic. द्वाराअनुसंधान पत्र [Nougat: Neural Optical Understanding for Academic Documents](https://arxiv.org/abs/2308.13418) के साथ जारी किया गया
 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (विस्कॉन्सिन विश्वविद्यालय - मैडिसन से) साथ में कागज [Nyströmformer: A Nyström- आधारित एल्गोरिथम आत्म-ध्यान का अनुमान लगाने के लिए ](https://arxiv.org/abs/2102.03902) युनयांग ज़िओंग, झानपेंग ज़ेंग, रुद्रसिस चक्रवर्ती, मिंगक्सिंग टैन, ग्लेन फंग, यिन ली, विकास सिंह द्वारा पोस्ट किया गया।
 1. **[OneFormer](https://huggingface.co/docs/transformers/model_doc/oneformer)** (SHI Labs से) पेपर [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) जितेश जैन, जिआचेन ली, मांगटिक चिउ, अली हसनी, निकिता ओरलोव, हम्फ्री शि के द्वारा जारी किया गया है।
 1. **[OpenLlama](https://huggingface.co/docs/transformers/model_doc/open-llama)** (from [s-JoL](https://huggingface.co/s-JoL)) released in [Open-Llama](https://github.com/s-JoL/Open-Llama).
@@ -380,11 +384,12 @@ conda install -c huggingface transformers
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (Google की ओर से) साथ में दिया गया पेपर [लंबे इनपुट सारांश के लिए ट्रांसफ़ॉर्मरों को बेहतर तरीके से एक्सटेंड करना](https://arxiv .org/abs/2208.04347) जेसन फांग, याओ झाओ, पीटर जे लियू द्वारा।
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (दीपमाइंड से) साथ में पेपर [पर्सीवर आईओ: संरचित इनपुट और आउटपुट के लिए एक सामान्य वास्तुकला] (https://arxiv.org/abs/2107.14795) एंड्रयू जेगल, सेबेस्टियन बोरग्यूड, जीन-बैप्टिस्ट अलायराक, कार्ल डोर्श, कैटलिन इओनेस्कु, डेविड द्वारा डिंग, स्कंद कोप्पुला, डैनियल ज़ोरान, एंड्रयू ब्रॉक, इवान शेलहैमर, ओलिवियर हेनाफ, मैथ्यू एम। बोट्विनिक, एंड्रयू ज़िसरमैन, ओरिओल विनियल्स, जोआओ कैरेरा द्वारा पोस्ट किया गया।
+1. **[Persimmon](https://huggingface.co/docs/transformers/model_doc/persimmon)** (ADEPT से) Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani. द्वाराअनुसंधान पत्र [blog post](https://www.adept.ai/blog/persimmon-8b) के साथ जारी किया गया
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (VinAI Research से) कागज के साथ [PhoBERT: वियतनामी के लिए पूर्व-प्रशिक्षित भाषा मॉडल](https://www .aclweb.org/anthology/2020.findings-emnlp.92/) डैट क्वोक गुयेन और अन्ह तुआन गुयेन द्वारा पोस्ट किया गया।
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (Google से) Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova. द्वाराअनुसंधान पत्र [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) के साथ जारी किया गया
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (UCLA NLP से) साथ वाला पेपर [प्रोग्राम अंडरस्टैंडिंग एंड जेनरेशन के लिए यूनिफाइड प्री-ट्रेनिंग](https://arxiv .org/abs/2103.06333) वसी उद्दीन अहमद, सैकत चक्रवर्ती, बैशाखी रे, काई-वेई चांग द्वारा।
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
-1. **[Pop2Piano](https://huggingface.co/docs/transformers/main/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (माइक्रोसॉफ्ट रिसर्च से) साथ में पेपर [ProphetNet: प्रेडिक्टिंग फ्यूचर एन-ग्राम फॉर सीक्वेंस-टू-सीक्वेंस प्री-ट्रेनिंग ](https://arxiv.org/abs/2001.04063) यू यान, वीज़ेन क्यूई, येयुन गोंग, दयाहेंग लियू, नान डुआन, जिउशेंग चेन, रुओफ़ेई झांग और मिंग झोउ द्वारा पोस्ट किया गया।
 1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. से) Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao. द्वाराअनुसंधान पत्र [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) के साथ जारी किया गया
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (NVIDIA से) साथ वाला पेपर [डीप लर्निंग इंफ़ेक्शन के लिए इंटीजर क्वांटिज़ेशन: प्रिंसिपल्स एंड एम्पिरिकल इवैल्यूएशन](https:// arxiv.org/abs/2004.09602) हाओ वू, पैट्रिक जुड, जिआओजी झांग, मिखाइल इसेव और पॉलियस माइकेविसियस द्वारा।
@@ -435,10 +440,11 @@ conda install -c huggingface transformers
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (गूगल एआई से) कागज के साथ [एक इमेज इज़ वर्थ 16x16 वर्ड्स: ट्रांसफॉर्मर्स फॉर इमेज रिकॉग्निशन एट स्केल](https://arxiv.org/abs/2010.11929) एलेक्सी डोसोवित्स्की, लुकास बेयर, अलेक्जेंडर कोलेसनिकोव, डिर्क वीसेनबोर्न, शियाओहुआ झाई, थॉमस अनटरथिनर, मुस्तफा देहघानी, मैथियास मिंडरर, जॉर्ज हेगोल्ड, सिल्वेन गेली, जैकब उस्ज़कोरेइट द्वारा हॉल्सबी द्वारा पोस्ट किया गया।
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (UCLA NLP से) साथ वाला पेपर [VisualBERT: A Simple and Performant Baseline for Vision and Language](https:/ /arxiv.org/pdf/1908.03557) लियुनियन हेरोल्ड ली, मार्क यात्स्कर, दा यिन, चो-जुई हसीह, काई-वेई चांग द्वारा।
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
-1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (Meta AI से) Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He. द्वाराअनुसंधान पत्र [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) के साथ जारी किया गया
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (Meta AI से) Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He. द्वाराअनुसंधान पत्र [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) के साथ जारी किया गया
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (मेटा एआई से) साथ में कागज [मास्कड ऑटोएन्कोडर स्केलेबल विजन लर्नर्स हैं](https://arxiv.org/ एब्स/2111.06377) कैमिंग हे, ज़िनेली चेन, सेनिंग ज़ी, यांगहो ली, पिओट्र डॉलर, रॉस गिर्शिक द्वारा।
+1. **[ViTMatte](https://huggingface.co/docs/transformers/model_doc/vitmatte)** (HUST-VL से) Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang. द्वाराअनुसंधान पत्र [ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers](https://arxiv.org/abs/2305.15272) के साथ जारी किया गया
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (मेटा एआई से) साथ में कागज [लेबल-कुशल सीखने के लिए मास्क्ड स्याम देश के नेटवर्क](https://arxiv. org/abs/2204.07141) महमूद असरान, मथिल्डे कैरन, ईशान मिश्रा, पियोट्र बोजानोवस्की, फ्लोरियन बोर्डेस, पास्कल विंसेंट, आर्मंड जौलिन, माइकल रब्बत, निकोलस बल्लास द्वारा।
-1. **[VITS](https://huggingface.co/docs/transformers/main/model_doc/vits)** (Kakao Enterprise से) Jaehyeon Kim, Jungil Kong, Juhee Son. द्वाराअनुसंधान पत्र [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) के साथ जारी किया गया
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (Kakao Enterprise से) Jaehyeon Kim, Jungil Kong, Juhee Son. द्वाराअनुसंधान पत्र [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) के साथ जारी किया गया
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (फेसबुक एआई से) साथ में पेपर [wav2vec 2.0: ए फ्रेमवर्क फॉर सेल्फ-सुपरवाइज्ड लर्निंग ऑफ स्पीच रिप्रेजेंटेशन] (https://arxiv.org/abs/2006.11477) एलेक्सी बेवस्की, हेनरी झोउ, अब्देलरहमान मोहम्मद, माइकल औली द्वारा।
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (Facebook AI से) साथ वाला पेपर [FAIRSEQ S2T: FAIRSEQ के साथ फास्ट स्पीच-टू-टेक्स्ट मॉडलिंग ](https://arxiv.org/abs/2010.05171) चांगहान वांग, यूं तांग, जुताई मा, ऐनी वू, सरव्या पोपुरी, दिमित्रो ओखोनको, जुआन पिनो द्वारा पोस्ट किया गया।

README_ja.md

@@ -53,7 +53,7 @@ user: ユーザ
     <br>
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers_logo_name.png" width="400"/>
     <br>
-<p>
+</p>
 <p align="center">
     <a href="https://circleci.com/gh/huggingface/transformers">
         <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/main">
@@ -82,7 +82,7 @@ user: ユーザ
         <a href="https://github.com/huggingface/transformers/blob/main/README_es.md">Español</a> |
         <b>日本語</b> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_hd.md">हिन्दी</a>
-    <p>
+    </p>
 </h4>
 
 <h3 align="center">
@@ -321,6 +321,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (BigScience workshop から) [BigScience Workshop](https://bigscience.huggingface.co/) から公開されました.
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (Alexa から) Adrian de Wynter and Daniel J. Perry から公開された研究論文: [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499)
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (Harbin Institute of Technology/Microsoft Research Asia/Intel Labs から) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (NAVER CLOVA から) Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park. から公開された研究論文 [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539)
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (Google Research から) Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel から公開された研究論文: [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626)
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (Inria/Facebook/Sorbonne から) Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot から公開された研究論文: [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894)
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (Google Research から) Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting から公開された研究論文: [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874)
@@ -384,6 +385,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) 坂本俊之(tanreinama)からリリースされました.
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (Microsoft から) Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu から公開された研究論文: [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234).
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA から) Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang から公開された研究論文: [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094)
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (Allegro.pl, AGH University of Science and Technology から) Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik. から公開された研究論文 [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf)
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (Facebook から) Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed から公開された研究論文: [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447)
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (Berkeley から) Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer から公開された研究論文: [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321)
 1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
@@ -417,6 +419,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[Megatron-BERT](https://huggingface.co/docs/transformers/model_doc/megatron-bert)** (NVIDIA から) Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro から公開された研究論文: [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053)
 1. **[Megatron-GPT2](https://huggingface.co/docs/transformers/model_doc/megatron_gpt2)** (NVIDIA から) Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro から公開された研究論文: [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053)
 1. **[MGP-STR](https://huggingface.co/docs/transformers/model_doc/mgp-str)** (Alibaba Research から) Peng Wang, Cheng Da, and Cong Yao. から公開された研究論文 [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592)
+1. **[Mistral](https://huggingface.co/docs/transformers/model_doc/mistral)** (from Mistral AI) by The Mistral AI team: Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.. 
 1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (Studio Ousia から) Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka から公開された研究論文: [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151)
 1. **[MMS](https://huggingface.co/docs/transformers/model_doc/mms)** (Facebook から) Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli. から公開された研究論文 [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516)
 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (CMU/Google Brain から) Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou から公開された研究論文: [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984)
@@ -434,6 +437,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[Nezha](https://huggingface.co/docs/transformers/model_doc/nezha)** (Huawei Noah’s Ark Lab から) Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu から公開された研究論文: [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204)
 1. **[NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)** (Meta から) the NLLB team から公開された研究論文: [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672)
 1. **[NLLB-MOE](https://huggingface.co/docs/transformers/model_doc/nllb-moe)** (Meta から) the NLLB team. から公開された研究論文 [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672)
+1. **[Nougat](https://huggingface.co/docs/transformers/model_doc/nougat)** (Meta AI から) Lukas Blecher, Guillem Cucurull, Thomas Scialom, Robert Stojnic. から公開された研究論文 [Nougat: Neural Optical Understanding for Academic Documents](https://arxiv.org/abs/2308.13418)
 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (the University of Wisconsin - Madison から) Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh から公開された研究論文: [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902)
 1. **[OneFormer](https://huggingface.co/docs/transformers/model_doc/oneformer)** (SHI Labs から) Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi から公開された研究論文: [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220)
 1. **[OpenLlama](https://huggingface.co/docs/transformers/model_doc/open-llama)** (from [s-JoL](https://huggingface.co/s-JoL)) released in [Open-Llama](https://github.com/s-JoL/Open-Llama).
@@ -442,11 +446,12 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (Google から) Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu から公開された研究論文: [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777)
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (Google から) Jason Phang, Yao Zhao, and Peter J. Liu から公開された研究論文: [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347)
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (Deepmind から) Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira から公開された研究論文: [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795)
+1. **[Persimmon](https://huggingface.co/docs/transformers/model_doc/persimmon)** (ADEPT から) Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani. から公開された研究論文 [blog post](https://www.adept.ai/blog/persimmon-8b)
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (VinAI Research から) Dat Quoc Nguyen and Anh Tuan Nguyen から公開された研究論文: [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/)
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (Google から) Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova. から公開された研究論文 [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347)
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (UCLA NLP から) Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang から公開された研究論文: [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333)
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (Sea AI Labs から) Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng から公開された研究論文: [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418)
-1. **[Pop2Piano](https://huggingface.co/docs/transformers/main/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (Microsoft Research から) Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou から公開された研究論文: [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063)
 1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. から) Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao. から公開された研究論文 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf)
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (NVIDIA から) Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius から公開された研究論文: [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602)
@@ -497,10 +502,11 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (Google AI から) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby から公開された研究論文: [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929)
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (UCLA NLP から) Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang から公開された研究論文: [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557)
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (Google AI から) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby から公開された研究論文: [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929)
-1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (Meta AI から) Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He. から公開された研究論文 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527)
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (Meta AI から) Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He. から公開された研究論文 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527)
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (Meta AI から) Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick から公開された研究論文: [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377)
+1. **[ViTMatte](https://huggingface.co/docs/transformers/model_doc/vitmatte)** (HUST-VL から) Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang. から公開された研究論文 [ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers](https://arxiv.org/abs/2305.15272)
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (Meta AI から) Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas から公開された研究論文: [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141)
-1. **[VITS](https://huggingface.co/docs/transformers/main/model_doc/vits)** (Kakao Enterprise から) Jaehyeon Kim, Jungil Kong, Juhee Son. から公開された研究論文 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103)
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (Kakao Enterprise から) Jaehyeon Kim, Jungil Kong, Juhee Son. から公開された研究論文 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103)
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (Facebook AI から) Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli から公開された研究論文: [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477)
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (Facebook AI から) Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino から公開された研究論文: [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171)

README_ko.md

@@ -18,7 +18,7 @@ limitations under the License.
     <br>
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers_logo_name.png" width="400"/>
     <br>
-<p>
+</p>
 <p align="center">
     <a href="https://circleci.com/gh/huggingface/transformers">
         <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/main">
@@ -47,7 +47,7 @@ limitations under the License.
         <a href="https://github.com/huggingface/transformers/blob/main/README_es.md">Español</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_ja.md">日本語</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_hd.md">हिन्दी</a>
-    <p>
+    </p>
 </h4>
 
 <h3 align="center">
@@ -236,6 +236,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (Alexa 에서) Adrian de Wynter and Daniel J. Perry 의 [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) 논문과 함께 발표했습니다.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (NAVER CLOVA 에서 제공)은 Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.의 [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539)논문과 함께 발표했습니다.
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (Google Research 에서) Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel 의 [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) 논문과 함께 발표했습니다.
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (Inria/Facebook/Sorbonne 에서) Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot 의 [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) 논문과 함께 발표했습니다.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (Google Research 에서) Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting 의 [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) 논문과 함께 발표했습니다.
@@ -299,6 +300,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu  의 [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234)  논문과 함께 발표했습니다.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA 에서) Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang 의 [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) 논문과 함께 발표했습니다.
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (Allegro.pl, AGH University of Science and Technology 에서 제공)은 Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.의 [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf)논문과 함께 발표했습니다.
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (Facebook 에서) Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed 의 [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) 논문과 함께 발표했습니다.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (Berkeley 에서) Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer 의 [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) 논문과 함께 발표했습니다.
 1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
@@ -332,6 +334,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Megatron-BERT](https://huggingface.co/docs/transformers/model_doc/megatron-bert)** (NVIDIA 에서) Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro 의 [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) 논문과 함께 발표했습니다.
 1. **[Megatron-GPT2](https://huggingface.co/docs/transformers/model_doc/megatron_gpt2)** (NVIDIA 에서) Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro 의 [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) 논문과 함께 발표했습니다.
 1. **[MGP-STR](https://huggingface.co/docs/transformers/model_doc/mgp-str)** (Alibaba Research 에서 제공)은 Peng Wang, Cheng Da, and Cong Yao.의 [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592)논문과 함께 발표했습니다.
+1. **[Mistral](https://huggingface.co/docs/transformers/model_doc/mistral)** (from Mistral AI) by The Mistral AI team: Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.. 
 1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (Studio Ousia 에서) Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka 의 [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) 논문과 함께 발표했습니다.
 1. **[MMS](https://huggingface.co/docs/transformers/model_doc/mms)** (Facebook 에서 제공)은 Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli.의 [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516)논문과 함께 발표했습니다.
 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (CMU/Google Brain 에서) Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou 의 [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) 논문과 함께 발표했습니다.
@@ -349,6 +352,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Nezha](https://huggingface.co/docs/transformers/model_doc/nezha)** (Huawei Noah’s Ark Lab 에서) Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu 의 [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) 논문과 함께 발표했습니다.
 1. **[NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)** (Meta 에서) the NLLB team 의 [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) 논문과 함께 발표했습니다.
 1. **[NLLB-MOE](https://huggingface.co/docs/transformers/model_doc/nllb-moe)** (Meta 에서 제공)은 the NLLB team.의 [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672)논문과 함께 발표했습니다.
+1. **[Nougat](https://huggingface.co/docs/transformers/model_doc/nougat)** (Meta AI 에서 제공)은 Lukas Blecher, Guillem Cucurull, Thomas Scialom, Robert Stojnic.의 [Nougat: Neural Optical Understanding for Academic Documents](https://arxiv.org/abs/2308.13418)논문과 함께 발표했습니다.
 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (the University of Wisconsin - Madison 에서) Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh 의 [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) 논문과 함께 발표했습니다.
 1. **[OneFormer](https://huggingface.co/docs/transformers/model_doc/oneformer)** (SHI Labs 에서) Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi 의 [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) 논문과 함께 발표했습니다.
 1. **[OpenLlama](https://huggingface.co/docs/transformers/model_doc/open-llama)** (from [s-JoL](https://huggingface.co/s-JoL)) released in [Open-Llama](https://github.com/s-JoL/Open-Llama).
@@ -357,11 +361,12 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (Google 에서) Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu 의 [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) 논문과 함께 발표했습니다.
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (Google 에서) Jason Phang, Yao Zhao, Peter J. Liu 의 [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) 논문과 함께 발표했습니다.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (Deepmind 에서) Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira 의 [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) 논문과 함께 발표했습니다.
+1. **[Persimmon](https://huggingface.co/docs/transformers/model_doc/persimmon)** (ADEPT 에서 제공)은 Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.의 [blog post](https://www.adept.ai/blog/persimmon-8b)논문과 함께 발표했습니다.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (VinAI Research 에서) Dat Quoc Nguyen and Anh Tuan Nguyen 의 [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) 논문과 함께 발표했습니다.
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (Google 에서 제공)은 Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.의 [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347)논문과 함께 발표했습니다.
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (UCLA NLP 에서) Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang 의 [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) 논문과 함께 발표했습니다.
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (Sea AI Labs 에서) Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng 의 [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) 논문과 함께 발표했습니다.
-1. **[Pop2Piano](https://huggingface.co/docs/transformers/main/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (Microsoft Research 에서) Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou 의 [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) 논문과 함께 발표했습니다.
 1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. 에서 제공)은 Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.의 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf)논문과 함께 발표했습니다.
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (NVIDIA 에서) Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius 의 [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) 논문과 함께 발표했습니다.
@@ -412,10 +417,11 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (Google AI 에서) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 의 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 논문과 함께 발표했습니다.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (UCLA NLP 에서) Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang 의 [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) 논문과 함께 발표했습니다.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (Google AI 에서) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 의 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 논문과 함께 발표했습니다.
-1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (Meta AI 에서 제공)은 Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.의 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527)논문과 함께 발표했습니다.
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (Meta AI 에서 제공)은 Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.의 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527)논문과 함께 발표했습니다.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (Meta AI 에서) Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick 의 [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) 논문과 함께 발표했습니다.
+1. **[ViTMatte](https://huggingface.co/docs/transformers/model_doc/vitmatte)** (HUST-VL 에서 제공)은 Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang.의 [ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers](https://arxiv.org/abs/2305.15272)논문과 함께 발표했습니다.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (Meta AI 에서) Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas 의 [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) 논문과 함께 발표했습니다.
-1. **[VITS](https://huggingface.co/docs/transformers/main/model_doc/vits)** (Kakao Enterprise 에서 제공)은 Jaehyeon Kim, Jungil Kong, Juhee Son.의 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103)논문과 함께 발표했습니다.
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (Kakao Enterprise 에서 제공)은 Jaehyeon Kim, Jungil Kong, Juhee Son.의 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103)논문과 함께 발표했습니다.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (Facebook AI 에서) Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli 의 [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) 논문과 함께 발표했습니다.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (Facebook AI 에서) Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino 의 [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) 논문과 함께 발표했습니다.
@@ -434,7 +440,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[XLS-R](https://huggingface.co/docs/transformers/model_doc/xls_r)** (Facebook AI 에서) Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli 의 [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) 논문과 함께 발표했습니다.
 1. **[XLSR-Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/xlsr_wav2vec2)** (Facebook AI 에서) Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli 의 [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://arxiv.org/abs/2006.13979) 논문과 함께 발표했습니다.
 1. **[YOLOS](https://huggingface.co/docs/transformers/model_doc/yolos)** (Huazhong University of Science & Technology 에서) Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu 의 [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) 논문과 함께 발표했습니다.
-1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (the University of Wisconsin - Madison 에서) Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh 의 [You Only Sample (Almost) 논문과 함께 발표했습니다.
+1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (the University of Wisconsin - Madison 에서) Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh 의 [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://arxiv.org/abs/2111.09714) 논문과 함께 발표했습니다.
 1. 새로운 모델을 올리고 싶나요? 우리가 **상세한 가이드와 템플릿** 으로 새로운 모델을 올리도록 도와드릴게요. 가이드와 템플릿은 이 저장소의 [`templates`](./templates) 폴더에서 확인하실 수 있습니다. [컨트리뷰션 가이드라인](./CONTRIBUTING.md)을 꼭 확인해주시고, PR을 올리기 전에 메인테이너에게 연락하거나 이슈를 오픈해 피드백을 받으시길 바랍니다.
 
 각 모델이 Flax, PyTorch, TensorFlow으로 구현되었는지 또는 🤗 Tokenizers 라이브러리가 지원하는 토크나이저를 사용하는지 확인하려면, [이 표](https://huggingface.co/docs/transformers/index#supported-frameworks)를 확인하세요.

README_ru.md

@@ -0,0 +1,549 @@
+<!---
+Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/datasets/huggingface/documentation-images/raw/main/transformers-logo-dark.svg">
+    <source media="(prefers-color-scheme: light)" srcset="https://huggingface.co/datasets/huggingface/documentation-images/raw/main/transformers-logo-light.svg">
+    <img alt="Hugging Face Transformers Library" src="https://huggingface.co/datasets/huggingface/documentation-images/raw/main/transformers-logo-light.svg" width="352" height="59" style="max-width: 100%;">
+  </picture>
+  <br/>
+  <br/>
+</p>
+
+<p align="center">
+    <a href="https://circleci.com/gh/huggingface/transformers">
+        <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/main">
+    </a>
+    <a href="https://github.com/huggingface/transformers/blob/main/LICENSE">
+        <img alt="GitHub" src="https://img.shields.io/github/license/huggingface/transformers.svg?color=blue">
+    </a>
+    <a href="https://huggingface.co/docs/transformers/index">
+        <img alt="Documentation" src="https://img.shields.io/website/http/huggingface.co/docs/transformers/index.svg?down_color=red&down_message=offline&up_message=online">
+    </a>
+    <a href="https://github.com/huggingface/transformers/releases">
+        <img alt="GitHub release" src="https://img.shields.io/github/release/huggingface/transformers.svg">
+    </a>
+    <a href="https://github.com/huggingface/transformers/blob/main/CODE_OF_CONDUCT.md">
+        <img alt="Contributor Covenant" src="https://img.shields.io/badge/Contributor%20Covenant-v2.0%20adopted-ff69b4.svg">
+    </a>
+    <a href="https://zenodo.org/badge/latestdoi/155220641"><img src="https://zenodo.org/badge/155220641.svg" alt="DOI"></a>
+</p>
+
+<h4 align="center">
+    <p>
+        <a href="https://github.com/huggingface/transformers/blob/main/README.md">English</a> |
+        <a href="https://github.com/huggingface/transformers/blob/main/README_zh-hans.md">简体中文</a> |
+        <a href="https://github.com/huggingface/transformers/blob/main/README_zh-hant.md">繁體中文</a> |
+        <a href="https://github.com/huggingface/transformers/blob/main/README_ko.md">한국어</a> |
+        <a href="https://github.com/huggingface/transformers/blob/main/README_es.md">Español</a> |
+        <a href="https://github.com/huggingface/transformers/blob/main/README_ja.md">日本語</a> |
+        <a href="https://github.com/huggingface/transformers/blob/main/README_hd.md">हिन्दी</a> |
+        <b>Русский</b>
+    <p>
+</h4>
+
+<h3 align="center">
+    <p>Современное машинное обучение для JAX, PyTorch и TensorFlow</p>
+</h3>
+
+<h3 align="center">
+    <a href="https://hf.co/course"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/course_banner.png"></a>
+</h3>
+
+🤗 Transformers предоставляет тысячи предварительно обученных моделей для выполнения различных задач, таких как текст, зрение и аудио.
+
+Эти модели могут быть применены на:
+
+* 📝 Текст, для таких задач, как классификация текстов, извлечение информации, ответы на вопросы, обобщение, перевод, генерация текстов, на более чем 100 языках.
+* 🖼️ Изображения - для задач классификации изображений, обнаружения объектов и сегментации.
+* 🗣️ Аудио - для задач распознавания речи и классификации аудио.
+
+Модели transformers также могут выполнять несколько задач, такие как ответы на табличные вопросы, распознавание оптических символов, извлечение информации из отсканированных документов, классификация видео и ответы на визуальные вопросы.
+
+🤗 Transformers предоставляет API для быстрой загрузки и использования предварительно обученных моделей, их тонкой настройки на собственных датасетах и последующего взаимодействия ими с сообществом на нашем [сайте](https://huggingface.co/models). В то же время каждый python модуль, определяющий архитектуру, полностью автономен и может быть модифицирован для проведения быстрых исследовательских экспериментов.
+
+🤗 Transformers опирается на три самые популярные библиотеки глубокого обучения - [Jax](https://jax.readthedocs.io/en/latest/), [PyTorch](https://pytorch.org/) и [TensorFlow](https://www.tensorflow.org/) - и легко интегрируется между ними. Это позволяет легко обучать модели с помощью одной из них, а затем загружать их для выводов с помощью другой.
+
+## Онлайн демонстрация
+
+Большинство наших моделей можно протестировать непосредственно на их страницах с [сайта](https://huggingface.co/models). Мы также предлагаем [привтаный хостинг моделей, контроль версий и API для выводов](https://huggingface.co/pricing) для публичных и частных моделей.
+
+Вот несколько примеров:
+
+В области NLP ( Обработка текстов на естественном языке ):
+- [Маскированное заполнение слов с помощью BERT](https://huggingface.co/bert-base-uncased?text=Paris+is+the+%5BMASK%5D+of+France)
+- [Распознавание сущностей с помощью Electra](https://huggingface.co/dbmdz/electra-large-discriminator-finetuned-conll03-english?text=My+name+is+Sarah+and+I+live+in+London+city)
+- [Генерация текста с помощью GPT-2](https://huggingface.co/gpt2?text=A+long+time+ago%2C+)
+- [Выводы на естественном языке с помощью RoBERTa](https://huggingface.co/roberta-large-mnli?text=The+dog+was+lost.+Nobody+lost+any+animal)
+- [Обобщение с помощью BART](https://huggingface.co/facebook/bart-large-cnn?text=The+tower+is+324+metres+%281%2C063+ft%29+tall%2C+about+the+same+height+as+an+81-storey+building%2C+and+the+tallest+structure+in+Paris.+Its+base+is+square%2C+measuring+125+metres+%28410+ft%29+on+each+side.+During+its+construction%2C+the+Eiffel+Tower+surpassed+the+Washington+Monument+to+become+the+tallest+man-made+structure+in+the+world%2C+a+title+it+held+for+41+years+until+the+Chrysler+Building+in+New+York+City+was+finished+in+1930.+It+was+the+first+structure+to+reach+a+height+of+300+metres.+Due+to+the+addition+of+a+broadcasting+aerial+at+the+top+of+the+tower+in+1957%2C+it+is+now+taller+than+the+Chrysler+Building+by+5.2+metres+%2817+ft%29.+Excluding+transmitters%2C+the+Eiffel+Tower+is+the+second+tallest+free-standing+structure+in+France+after+the+Millau+Viaduct)
+- [Ответы на вопросы с помощью DistilBERT](https://huggingface.co/distilbert-base-uncased-distilled-squad?text=Which+name+is+also+used+to+describe+the+Amazon+rainforest+in+English%3F&context=The+Amazon+rainforest+%28Portuguese%3A+Floresta+Amaz%C3%B4nica+or+Amaz%C3%B4nia%3B+Spanish%3A+Selva+Amaz%C3%B3nica%2C+Amazon%C3%ADa+or+usually+Amazonia%3B+French%3A+For%C3%AAt+amazonienne%3B+Dutch%3A+Amazoneregenwoud%29%2C+also+known+in+English+as+Amazonia+or+the+Amazon+Jungle%2C+is+a+moist+broadleaf+forest+that+covers+most+of+the+Amazon+basin+of+South+America.+This+basin+encompasses+7%2C000%2C000+square+kilometres+%282%2C700%2C000+sq+mi%29%2C+of+which+5%2C500%2C000+square+kilometres+%282%2C100%2C000+sq+mi%29+are+covered+by+the+rainforest.+This+region+includes+territory+belonging+to+nine+nations.+The+majority+of+the+forest+is+contained+within+Brazil%2C+with+60%25+of+the+rainforest%2C+followed+by+Peru+with+13%25%2C+Colombia+with+10%25%2C+and+with+minor+amounts+in+Venezuela%2C+Ecuador%2C+Bolivia%2C+Guyana%2C+Suriname+and+French+Guiana.+States+or+departments+in+four+nations+contain+%22Amazonas%22+in+their+names.+The+Amazon+represents+over+half+of+the+planet%27s+remaining+rainforests%2C+and+comprises+the+largest+and+most+biodiverse+tract+of+tropical+rainforest+in+the+world%2C+with+an+estimated+390+billion+individual+trees+divided+into+16%2C000+species)
+- [Перевод с помощью T5](https://huggingface.co/t5-base?text=My+name+is+Wolfgang+and+I+live+in+Berlin)
+
+В области компьютерного зрения:
+- [Классификация изображений с помощью ViT](https://huggingface.co/google/vit-base-patch16-224)
+- [Обнаружение объектов с помощью DETR](https://huggingface.co/facebook/detr-resnet-50)
+- [Семантическая сегментация с помощью SegFormer](https://huggingface.co/nvidia/segformer-b0-finetuned-ade-512-512)
+- [Сегментация паноптикума с помощью MaskFormer](https://huggingface.co/facebook/maskformer-swin-small-coco)
+- [Оценка глубины с помощью DPT](https://huggingface.co/docs/transformers/model_doc/dpt)
+- [Классификация видео с помощью VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae)
+- [Универсальная сегментация с помощью OneFormer](https://huggingface.co/shi-labs/oneformer_ade20k_dinat_large)
+
+В области звука:
+- [Автоматическое распознавание речи с помощью Wav2Vec2](https://huggingface.co/facebook/wav2vec2-base-960h)
+- [Поиск ключевых слов с помощью Wav2Vec2](https://huggingface.co/superb/wav2vec2-base-superb-ks)
+- [Классификация аудиоданных с помощью траснформера аудиоспектрограмм](https://huggingface.co/MIT/ast-finetuned-audioset-10-10-0.4593)
+
+В мультимодальных задачах:
+- [Ответы на вопросы по таблице с помощью TAPAS](https://huggingface.co/google/tapas-base-finetuned-wtq)
+- [Визуальные ответы на вопросы с помощью ViLT](https://huggingface.co/dandelin/vilt-b32-finetuned-vqa)
+- [Zero-shot классификация изображений с помощью CLIP](https://huggingface.co/openai/clip-vit-large-patch14)
+- [Ответы на вопросы по документам с помощью LayoutLM](https://huggingface.co/impira/layoutlm-document-qa)
+- [Zero-shot классификация видео с помощью X-CLIP](https://huggingface.co/docs/transformers/model_doc/xclip)
+
+
+## 100 проектов, использующих Transformers
+
+Transformers - это не просто набор инструментов для использования предварительно обученных моделей: это сообщество проектов, созданное на его основе, и
+Hugging Face Hub. Мы хотим, чтобы Transformers позволил разработчикам, исследователям, студентам, профессорам, инженерам и всем желающим
+создавать проекты своей мечты.
+
+Чтобы отпраздновать 100 тысяч звезд Transformers, мы решили сделать акцент на сообществе, и создали страницу [awesome-transformers](./awesome-transformers.md), на которой перечислены 100
+невероятных проектов, созданных с помощью transformers.
+
+Если вы являетесь владельцем или пользователем проекта, который, по вашему мнению, должен быть включен в этот список, пожалуйста, откройте PR для его добавления!
+
+## Если вы хотите получить индивидуальную поддержку от команды Hugging Face
+
+<a target="_blank" href="https://huggingface.co/support">
+    <img alt="HuggingFace Expert Acceleration Program" src="https://cdn-media.huggingface.co/marketing/transformers/new-support-improved.png" style="max-width: 600px; border: 1px solid #eee; border-radius: 4px; box-shadow: 0 1px 2px 0 rgba(0, 0, 0, 0.05);">
+</a><br>
+
+## Быстрый гайд
+
+Для использования модели на заданном входе (текст, изображение, звук, ...) мы предоставляем API `pipeline`. Конвейеры объединяют предварительно обученную модель с препроцессингом, который использовался при ее обучении. Вот как можно быстро использовать конвейер для классификации положительных и отрицательных текстов:
+
+```python
+>>> from transformers import pipeline
+
+# Выделение конвейера для анализа настроений
+>>> classifier = pipeline('sentiment-analysis')
+>>> classifier('Мы очень рады представить конвейер в transformers.')
+[{'label': 'POSITIVE', 'score': 0.9996980428695679}]
+```
+
+Вторая строка кода загружает и кэширует предварительно обученную модель, используемую конвейером, а третья оценивает ее на заданном тексте. Здесь ответ "POSITIVE" с уверенностью 99,97%.
+
+Во многих задачах, как в НЛП, так и в компьютерном зрении и речи, уже есть готовый `pipeline`. Например, мы можем легко извлечь обнаруженные объекты на изображении:
+
+``` python
+>>> import requests
+>>> from PIL import Image
+>>> from transformers import pipeline
+
+# Скачиваем изображение с милыми котиками
+>>> url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/coco_sample.png"
+>>> image_data = requests.get(url, stream=True).raw
+>>> image = Image.open(image_data)
+
+# Выделение конвейера для обнаружения объектов
+>>> object_detector = pipeline('object-detection')
+>>> object_detector(image)
+[{'score': 0.9982201457023621,
+  'label': 'remote',
+  'box': {'xmin': 40, 'ymin': 70, 'xmax': 175, 'ymax': 117}},
+ {'score': 0.9960021376609802,
+  'label': 'remote',
+  'box': {'xmin': 333, 'ymin': 72, 'xmax': 368, 'ymax': 187}},
+ {'score': 0.9954745173454285,
+  'label': 'couch',
+  'box': {'xmin': 0, 'ymin': 1, 'xmax': 639, 'ymax': 473}},
+ {'score': 0.9988006353378296,
+  'label': 'cat',
+  'box': {'xmin': 13, 'ymin': 52, 'xmax': 314, 'ymax': 470}},
+ {'score': 0.9986783862113953,
+  'label': 'cat',
+  'box': {'xmin': 345, 'ymin': 23, 'xmax': 640, 'ymax': 368}}]
+```
+
+Здесь мы получаем список объектов, обнаруженных на изображении, с рамкой вокруг объекта и оценкой достоверности. Слева - исходное изображение, справа прогнозы:
+
+<h3 align="center">
+    <a><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/coco_sample.png" width="400"></a>
+    <a><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/coco_sample_post_processed.png" width="400"></a>
+</h3>
+
+Подробнее о задачах, поддерживаемых API `pipeline`, можно узнать в [этом учебном пособии](https://huggingface.co/docs/transformers/task_sum)
+
+В дополнение к `pipeline`, для загрузки и использования любой из предварительно обученных моделей в заданной задаче достаточно трех строк кода. Вот версия для PyTorch:
+```python
+>>> from transformers import AutoTokenizer, AutoModel
+
+>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+>>> model = AutoModel.from_pretrained("bert-base-uncased")
+
+>>> inputs = tokenizer("Привет мир!", return_tensors="pt")
+>>> outputs = model(**inputs)
+```
+
+А вот эквивалентный код для TensorFlow:
+```python
+>>> from transformers import AutoTokenizer, TFAutoModel
+
+>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+>>> model = TFAutoModel.from_pretrained("bert-base-uncased")
+
+>>> inputs = tokenizer("Привет мир!", return_tensors="tf")
+>>> outputs = model(**inputs)
+```
+
+Токенизатор отвечает за всю предварительную обработку, которую ожидает предварительно обученная модель, и может быть вызван непосредственно с помощью одной строки (как в приведенных выше примерах) или на списке. В результате будет получен словарь, который можно использовать в последующем коде или просто напрямую передать в модель с помощью оператора распаковки аргументов **.
+
+Сама модель представляет собой обычный [Pytorch `nn.Module`](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) или [TensorFlow `tf.keras.Model`](https://www.tensorflow.org/api_docs/python/tf/keras/Model) (в зависимости от используемого бэкенда), который можно использовать как обычно. [В этом руководстве](https://huggingface.co/docs/transformers/training) рассказывается, как интегрировать такую модель в классический цикл обучения PyTorch или TensorFlow, или как использовать наш API `Trainer` для быстрой тонкой настройки на новом датасете.
+
+## Почему необходимо использовать transformers?
+
+1. Простые в использовании современные модели:
+    - Высокая производительность в задачах понимания и генерации естественного языка, компьютерного зрения и аудио.
+    - Низкий входной барьер для преподавателей и практиков.
+    - Небольшое количество абстракций для пользователя и всего три класса для изучения.
+    - Единый API для использования всех наших предварительно обученных моделей.
+
+1. Более низкие вычислительные затраты, меньший "углеродный след":
+    - Исследователи могут обмениваться обученными моделями вместо того, чтобы постоянно их переобучать.
+    - Практики могут сократить время вычислений и производственные затраты.
+    - Десятки архитектур с более чем 60 000 предварительно обученных моделей для всех модальностей.
+
+1. Выбор подходящего фреймворка для каждого этапа жизни модели:
+    - Обучение самых современных моделей за 3 строки кода.
+    - Перемещайте одну модель между фреймворками TF2.0/PyTorch/JAX по своему усмотрению.
+    - Беспрепятственный выбор подходящего фреймворка для обучения, оценки и производства.
+
+1. Легко настроить модель или пример под свои нужды:
+    - Мы предоставляем примеры для каждой архитектуры, чтобы воспроизвести результаты, опубликованные их авторами.
+    - Внутренние компоненты модели раскрываются максимально последовательно.
+    - Файлы моделей можно использовать независимо от библиотеки для проведения быстрых экспериментов.
+
+## Почему я не должен использовать transformers?
+
+- Данная библиотека не является модульным набором строительных блоков для нейронных сетей. Код в файлах моделей специально не рефакторится дополнительными абстракциями, чтобы исследователи могли быстро итеративно работать с каждой из моделей, не погружаясь в дополнительные абстракции/файлы.
+- API обучения не предназначен для работы с любой моделью, а оптимизирован для работы с моделями, предоставляемыми библиотекой. Для работы с общими циклами машинного обучения следует использовать другую библиотеку (возможно, [Accelerate](https://huggingface.co/docs/accelerate)).
+- Несмотря на то, что мы стремимся представить как можно больше примеров использования, скрипты в нашей папке [примеров](https://github.com/huggingface/transformers/tree/main/examples) являются именно примерами. Предполагается, что они не будут работать "из коробки" для решения вашей конкретной задачи, и вам придется изменить несколько строк кода, чтобы адаптировать их под свои нужды.
+
+## Установка
+
+### С помощью pip
+
+Данный репозиторий протестирован на Python 3.8+, Flax 0.4.1+, PyTorch 1.10+ и TensorFlow 2.6+.
+
+Устанавливать 🤗 Transformers следует в [виртуальной среде](https://docs.python.org/3/library/venv.html). Если вы не знакомы с виртуальными средами Python, ознакомьтесь с [руководством пользователя](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/).
+
+Сначала создайте виртуальную среду с той версией Python, которую вы собираетесь использовать, и активируйте ее.
+
+Затем необходимо установить хотя бы один бекенд из Flax, PyTorch или TensorFlow.
+Пожалуйста, обратитесь к страницам [TensorFlow установочная страница](https://www.tensorflow.org/install/), [PyTorch установочная страница](https://pytorch.org/get-started/locally/#start-locally) и/или [Flax](https://github.com/google/flax#quick-install) и [Jax](https://github.com/google/jax#installation), где описаны команды установки для вашей платформы.
+
+После установки одного из этих бэкендов 🤗 Transformers может быть установлен с помощью pip следующим образом:
+
+```bash
+pip install transformers
+```
+
+Если вы хотите поиграть с примерами или вам нужен самый современный код и вы не можете ждать нового релиза, вы должны [установить библиотеку из исходного кода](https://huggingface.co/docs/transformers/installation#installing-from-source).
+
+### С помощью conda
+
+Начиная с версии Transformers v4.0.0, у нас появилсась поддержка conda: `huggingface`.
+
+Установить Transformers с помощью conda можно следующим образом:
+
+```bash
+conda install -c huggingface transformers
+```
+
+О том, как установить Flax, PyTorch или TensorFlow с помощью conda, читайте на страницах, посвященных их установке.
+
+> **_ЗАМЕТКА:_** В операционной системе Windows вам может быть предложено активировать режим разработчика, чтобы воспользоваться преимуществами кэширования. Если для вас это невозможно, сообщите нам об этом [здесь](https://github.com/huggingface/huggingface_hub/issues/1062).
+
+## Модельные архитектуры
+
+**[Все контрольные точки моделей](https://huggingface.co/models)**, предоставляемые 🤗 Transformers, беспрепятственно интегрируются с huggingface.co [model hub](https://huggingface.co/models), куда они загружаются непосредственно [пользователями](https://huggingface.co/users) и [организациями](https://huggingface.co/organizations).
+
+Текущее количество контрольных точек: ![](https://img.shields.io/endpoint?url=https://huggingface.co/api/shields/models&color=brightgreen)
+
+🤗 В настоящее время Transformers предоставляет следующие архитектуры (подробное описание каждой из них см. [здесь](https://huggingface.co/docs/transformers/model_summary)):
+
+1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](https://huggingface.co/docs/transformers/model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
+1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
+1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
+1. **[Autoformer](https://huggingface.co/docs/transformers/model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
+1. **[Bark](https://huggingface.co/docs/transformers/model_doc/bark)** (from Suno) released in the repository [suno-ai/bark](https://github.com/suno-ai/bark) by Suno AI team.
+1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
+1. **[BARThez](https://huggingface.co/docs/transformers/model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
+1. **[BARTpho](https://huggingface.co/docs/transformers/model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.
+1. **[BEiT](https://huggingface.co/docs/transformers/model_doc/beit)** (from Microsoft) released with the paper [BEiT: BERT Pre-Training of Image Transformers](https://arxiv.org/abs/2106.08254) by Hangbo Bao, Li Dong, Furu Wei.
+1. **[BERT](https://huggingface.co/docs/transformers/model_doc/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
+1. **[BERT For Sequence Generation](https://huggingface.co/docs/transformers/model_doc/bert-generation)** (from Google) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
+1. **[BERTweet](https://huggingface.co/docs/transformers/model_doc/bertweet)** (from VinAI Research) released with the paper [BERTweet: A pre-trained language model for English Tweets](https://aclanthology.org/2020.emnlp-demos.2/) by Dat Quoc Nguyen, Thanh Vu and Anh Tuan Nguyen.
+1. **[BigBird-Pegasus](https://huggingface.co/docs/transformers/model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
+1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
+1. **[BioGpt](https://huggingface.co/docs/transformers/model_doc/biogpt)** (from Microsoft Research AI4Science) released with the paper [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) by Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu.
+1. **[BiT](https://huggingface.co/docs/transformers/model_doc/bit)** (from Google AI) released with the paper [Big Transfer (BiT): General Visual Representation Learning](https://arxiv.org/abs/1912.11370) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
+1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
+1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
+1. **[BLIP](https://huggingface.co/docs/transformers/model_doc/blip)** (from Salesforce) released with the paper [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086) by Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi.
+1. **[BLIP-2](https://huggingface.co/docs/transformers/model_doc/blip-2)** (from Salesforce) released with the paper [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597) by Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi.
+1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
+1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
+1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (from NAVER CLOVA) released with the paper [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
+1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
+1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
+1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
+1. **[Chinese-CLIP](https://huggingface.co/docs/transformers/model_doc/chinese_clip)** (from OFA-Sys) released with the paper [Chinese CLIP: Contrastive Vision-Language Pretraining in Chinese](https://arxiv.org/abs/2211.01335) by An Yang, Junshu Pan, Junyang Lin, Rui Men, Yichang Zhang, Jingren Zhou, Chang Zhou.
+1. **[CLAP](https://huggingface.co/docs/transformers/model_doc/clap)** (from LAION-AI) released with the paper [Large-scale Contrastive Language-Audio Pretraining with Feature Fusion and Keyword-to-Caption Augmentation](https://arxiv.org/abs/2211.06687) by Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg-Kirkpatrick, Shlomo Dubnov.
+1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
+1. **[CLIPSeg](https://huggingface.co/docs/transformers/model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
+1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
+1. **[CodeLlama](https://huggingface.co/docs/transformers/model_doc/llama_code)** (from MetaAI) released with the paper [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/) by Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve.
+1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
+1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
+1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
+1. **[ConvNeXTV2](https://huggingface.co/docs/transformers/model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie.
+1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun.
+1. **[CPM-Ant](https://huggingface.co/docs/transformers/model_doc/cpmant)** (from OpenBMB) released by the [OpenBMB](https://www.openbmb.org/).
+1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
+1. **[CvT](https://huggingface.co/docs/transformers/model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang.
+1. **[Data2Vec](https://huggingface.co/docs/transformers/model_doc/data2vec)** (from Facebook) released with the paper [Data2Vec:  A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) by Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli.
+1. **[DeBERTa](https://huggingface.co/docs/transformers/model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
+1. **[DeBERTa-v2](https://huggingface.co/docs/transformers/model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
+1. **[Decision Transformer](https://huggingface.co/docs/transformers/model_doc/decision_transformer)** (from Berkeley/Facebook/Google) released with the paper [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://arxiv.org/abs/2106.01345) by Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch.
+1. **[Deformable DETR](https://huggingface.co/docs/transformers/model_doc/deformable_detr)** (from SenseTime Research) released with the paper [Deformable DETR: Deformable Transformers for End-to-End Object Detection](https://arxiv.org/abs/2010.04159) by Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, Jifeng Dai.
+1. **[DeiT](https://huggingface.co/docs/transformers/model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
+1. **[DePlot](https://huggingface.co/docs/transformers/model_doc/deplot)** (from Google AI) released with the paper [DePlot: One-shot visual language reasoning by plot-to-table translation](https://arxiv.org/abs/2212.10505) by Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, Yasemin Altun.
+1. **[DETA](https://huggingface.co/docs/transformers/model_doc/deta)** (from The University of Texas at Austin) released with the paper [NMS Strikes Back](https://arxiv.org/abs/2212.06137) by Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, Philipp Krähenbühl.
+1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
+1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
+1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
+1. **[DINOv2](https://huggingface.co/docs/transformers/model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
+1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) and a German version of DistilBERT.
+1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
+1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (from NAVER), released together with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.
+1. **[DPR](https://huggingface.co/docs/transformers/model_doc/dpr)** (from Facebook) released with the paper [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
+1. **[DPT](https://huggingface.co/docs/transformers/master/model_doc/dpt)** (from Intel Labs) released with the paper [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) by René Ranftl, Alexey Bochkovskiy, Vladlen Koltun.
+1. **[EfficientFormer](https://huggingface.co/docs/transformers/model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
+1. **[EfficientNet](https://huggingface.co/docs/transformers/model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
+1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
+1. **[EnCodec](https://huggingface.co/docs/transformers/model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
+1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
+1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
+1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.
+1. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (from Meta AI) are transformer protein language models.  **ESM-1b** was released with the paper [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118) by Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus. **ESM-1v** was released with the paper [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648) by Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives. **ESM-2 and ESMFold** were released with the paper [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) by Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives.
+1. **[Falcon](https://huggingface.co/docs/transformers/model_doc/falcon)** (from Technology Innovation Institute) by Almazrouei, Ebtesam and Alobeidli, Hamza and Alshamsi, Abdulaziz and Cappelli, Alessandro and Cojocaru, Ruxandra and Debbah, Merouane and Goffinet, Etienne and Heslow, Daniel and Launay, Julien and Malartic, Quentin and Noune, Badreddine and Pannier, Baptiste and Penedo, Guilherme.
+1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei
+1. **[FLAN-UL2](https://huggingface.co/docs/transformers/model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei
+1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
+1. **[FLAVA](https://huggingface.co/docs/transformers/model_doc/flava)** (from Facebook AI) released with the paper [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela.
+1. **[FNet](https://huggingface.co/docs/transformers/model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon.
+1. **[FocalNet](https://huggingface.co/docs/transformers/model_doc/focalnet)** (from Microsoft Research) released with the paper [Focal Modulation Networks](https://arxiv.org/abs/2203.11926) by Jianwei Yang, Chunyuan Li, Xiyang Dai, Lu Yuan, Jianfeng Gao.
+1. **[Funnel Transformer](https://huggingface.co/docs/transformers/model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
+1. **[GIT](https://huggingface.co/docs/transformers/model_doc/git)** (from Microsoft Research) released with the paper [GIT: A Generative Image-to-text Transformer for Vision and Language](https://arxiv.org/abs/2205.14100) by Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, Lijuan Wang.
+1. **[GLPN](https://huggingface.co/docs/transformers/model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.
+1. **[GPT](https://huggingface.co/docs/transformers/model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
+1. **[GPT Neo](https://huggingface.co/docs/transformers/model_doc/gpt_neo)** (from EleutherAI) released in the repository [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) by Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy.
+1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach
+1. **[GPT NeoX Japanese](https://huggingface.co/docs/transformers/model_doc/gpt_neox_japanese)** (from ABEJA) released by Shinya Otani, Takayoshi Makabe, Anuj Arora, and Kyo Hattori.
+1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
+1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
+1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren.
+1. **[GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode)** (from BigCode) released with the paper [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988) by Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra.
+1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
+1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
+1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (from Allegro.pl, AGH University of Science and Technology) released with the paper [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) by Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.
+1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
+1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
+1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh.
+1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
+1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
+1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
+1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
+1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
+1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.
+1. **[LayoutLMv3](https://huggingface.co/docs/transformers/model_doc/layoutlmv3)** (from Microsoft Research Asia) released with the paper [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) by Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei.
+1. **[LayoutXLM](https://huggingface.co/docs/transformers/model_doc/layoutxlm)** (from Microsoft Research Asia) released with the paper [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) by Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei.
+1. **[LED](https://huggingface.co/docs/transformers/model_doc/led)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
+1. **[LeViT](https://huggingface.co/docs/transformers/model_doc/levit)** (from Meta AI) released with the paper [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) by Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze.
+1. **[LiLT](https://huggingface.co/docs/transformers/model_doc/lilt)** (from South China University of Technology) released with the paper [LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding](https://arxiv.org/abs/2202.13669) by Jiapeng Wang, Lianwen Jin, Kai Ding.
+1. **[LLaMA](https://huggingface.co/docs/transformers/model_doc/llama)** (from The FAIR team of Meta AI) released with the paper [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971) by Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, Guillaume Lample.
+1. **[Llama2](https://huggingface.co/docs/transformers/model_doc/llama2)** (from The FAIR team of Meta AI) released with the paper [Llama2: Open Foundation and Fine-Tuned Chat Models](https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/XXX) by Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushka rMishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing EllenTan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, Thomas Scialom.
+1. **[Longformer](https://huggingface.co/docs/transformers/model_doc/longformer)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
+1. **[LongT5](https://huggingface.co/docs/transformers/model_doc/longt5)** (from Google AI) released with the paper [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) by Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang.
+1. **[LUKE](https://huggingface.co/docs/transformers/model_doc/luke)** (from Studio Ousia) released with the paper [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto.
+1. **[LXMERT](https://huggingface.co/docs/transformers/model_doc/lxmert)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) by Hao Tan and Mohit Bansal.
+1. **[M-CTC-T](https://huggingface.co/docs/transformers/model_doc/mctct)** (from Facebook) released with the paper [Pseudo-Labeling For Massively Multilingual Speech Recognition](https://arxiv.org/abs/2111.00161) by Loren Lugosch, Tatiana Likhomanenko, Gabriel Synnaeve, and Ronan Collobert.
+1. **[M2M100](https://huggingface.co/docs/transformers/model_doc/m2m_100)** (from Facebook) released with the paper [Beyond English-Centric Multilingual Machine Translation](https://arxiv.org/abs/2010.11125) by Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, Armand Joulin.
+1. **[MarianMT](https://huggingface.co/docs/transformers/model_doc/marian)** Machine translation models trained using [OPUS](http://opus.nlpl.eu/) data by Jörg Tiedemann. The [Marian Framework](https://marian-nmt.github.io/) is being developed by the Microsoft Translator Team.
+1. **[MarkupLM](https://huggingface.co/docs/transformers/model_doc/markuplm)** (from Microsoft Research Asia) released with the paper [MarkupLM: Pre-training of Text and Markup Language for Visually-rich Document Understanding](https://arxiv.org/abs/2110.08518) by Junlong Li, Yiheng Xu, Lei Cui, Furu Wei.
+1. **[Mask2Former](https://huggingface.co/docs/transformers/model_doc/mask2former)** (from FAIR and UIUC) released with the paper [Masked-attention Mask Transformer for Universal Image Segmentation](https://arxiv.org/abs/2112.01527) by Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, Rohit Girdhar.
+1. **[MaskFormer](https://huggingface.co/docs/transformers/model_doc/maskformer)** (from Meta and UIUC) released with the paper [Per-Pixel Classification is Not All You Need for Semantic Segmentation](https://arxiv.org/abs/2107.06278) by Bowen Cheng, Alexander G. Schwing, Alexander Kirillov.
+1. **[MatCha](https://huggingface.co/docs/transformers/model_doc/matcha)** (from Google AI) released with the paper [MatCha: Enhancing Visual Language Pretraining with Math Reasoning and Chart Derendering](https://arxiv.org/abs/2212.09662) by Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, Julian Martin Eisenschlos.
+1. **[mBART](https://huggingface.co/docs/transformers/model_doc/mbart)** (from Facebook) released with the paper [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
+1. **[mBART-50](https://huggingface.co/docs/transformers/model_doc/mbart)** (from Facebook) released with the paper [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) by Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, Angela Fan.
+1. **[MEGA](https://huggingface.co/docs/transformers/model_doc/mega)** (from Meta/USC/CMU/SJTU) released with the paper [Mega: Moving Average Equipped Gated Attention](https://arxiv.org/abs/2209.10655) by Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer.
+1. **[Megatron-BERT](https://huggingface.co/docs/transformers/model_doc/megatron-bert)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
+1. **[Megatron-GPT2](https://huggingface.co/docs/transformers/model_doc/megatron_gpt2)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
+1. **[MGP-STR](https://huggingface.co/docs/transformers/model_doc/mgp-str)** (from Alibaba Research) released with the paper [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592) by Peng Wang, Cheng Da, and Cong Yao.
+1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
+1. **[MMS](https://huggingface.co/docs/transformers/model_doc/mms)** (from Facebook) released with the paper [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516) by Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli.
+1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
+1. **[MobileNetV1](https://huggingface.co/docs/transformers/model_doc/mobilenet_v1)** (from Google Inc.) released with the paper [MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications](https://arxiv.org/abs/1704.04861) by Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam.
+1. **[MobileNetV2](https://huggingface.co/docs/transformers/model_doc/mobilenet_v2)** (from Google Inc.) released with the paper [MobileNetV2: Inverted Residuals and Linear Bottlenecks](https://arxiv.org/abs/1801.04381) by Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen.
+1. **[MobileViT](https://huggingface.co/docs/transformers/model_doc/mobilevit)** (from Apple) released with the paper [MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer](https://arxiv.org/abs/2110.02178) by Sachin Mehta and Mohammad Rastegari.
+1. **[MobileViTV2](https://huggingface.co/docs/transformers/model_doc/mobilevitv2)** (from Apple) released with the paper [Separable Self-attention for Mobile Vision Transformers](https://arxiv.org/abs/2206.02680) by Sachin Mehta and Mohammad Rastegari.
+1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
+1. **[MPT](https://huggingface.co/docs/transformers/model_doc/mpt)** (from MosaiML) released with the repository [llm-foundry](https://github.com/mosaicml/llm-foundry/) by the MosaicML NLP Team.
+1. **[MRA](https://huggingface.co/docs/transformers/model_doc/mra)** (from the University of Wisconsin - Madison) released with the paper [Multi Resolution Analysis (MRA) for Approximate Self-Attention](https://arxiv.org/abs/2207.10284) by Zhanpeng Zeng, Sourav Pal, Jeffery Kline, Glenn M Fung, Vikas Singh.
+1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
+1. **[MusicGen](https://huggingface.co/docs/transformers/model_doc/musicgen)** (from Meta) released with the paper [Simple and Controllable Music Generation](https://arxiv.org/abs/2306.05284) by Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi and Alexandre Défossez.
+1. **[MVP](https://huggingface.co/docs/transformers/model_doc/mvp)** (from RUC AI Box) released with the paper [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen.
+1. **[NAT](https://huggingface.co/docs/transformers/model_doc/nat)** (from SHI Labs) released with the paper [Neighborhood Attention Transformer](https://arxiv.org/abs/2204.07143) by Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi.
+1. **[Nezha](https://huggingface.co/docs/transformers/model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu.
+1. **[NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
+1. **[NLLB-MOE](https://huggingface.co/docs/transformers/model_doc/nllb-moe)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
+1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
+1. **[OneFormer](https://huggingface.co/docs/transformers/model_doc/oneformer)** (from SHI Labs) released with the paper [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) by Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi.
+1. **[OpenLlama](https://huggingface.co/docs/transformers/model_doc/open-llama)** (from [s-JoL](https://huggingface.co/s-JoL)) released in [Open-Llama](https://github.com/s-JoL/Open-Llama).
+1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al.
+1. **[OWL-ViT](https://huggingface.co/docs/transformers/model_doc/owlvit)** (from Google AI) released with the paper [Simple Open-Vocabulary Object Detection with Vision Transformers](https://arxiv.org/abs/2205.06230) by Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby.
+1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
+1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, and Peter J. Liu.
+1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
+1. **[Persimmon](https://huggingface.co/docs/transformers/main/model_doc/persimmon)** (from ADEPT) released in a [blog post](https://www.adept.ai/blog/persimmon-8b) by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
+1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
+1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
+1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
+1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi and Kyogu Lee.
+1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
+1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
+1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
+1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
+1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
+1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
+1. **[RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)** (from META Platforms) released with the paper [Designing Network Design Space](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár.
+1. **[RemBERT](https://huggingface.co/docs/transformers/model_doc/rembert)** (from Google Research) released with the paper [Rethinking embedding coupling in pre-trained language models](https://arxiv.org/abs/2010.12821) by Hyung Won Chung, Thibault Févry, Henry Tsai, M. Johnson, Sebastian Ruder.
+1. **[ResNet](https://huggingface.co/docs/transformers/model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
+1. **[RoBERTa](https://huggingface.co/docs/transformers/model_doc/roberta)** (from Facebook), released together with the paper [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
+1. **[RoBERTa-PreLayerNorm](https://huggingface.co/docs/transformers/model_doc/roberta-prelayernorm)** (from Facebook) released with the paper [fairseq: A Fast, Extensible Toolkit for Sequence Modeling](https://arxiv.org/abs/1904.01038) by Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, Michael Auli.
+1. **[RoCBert](https://huggingface.co/docs/transformers/model_doc/roc_bert)** (from WeChatAI) released with the paper [RoCBert: Robust Chinese Bert with Multimodal Contrastive Pretraining](https://aclanthology.org/2022.acl-long.65.pdf) by HuiSu, WeiweiShi, XiaoyuShen, XiaoZhou, TuoJi, JiaruiFang, JieZhou.
+1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
+1. **[RWKV](https://huggingface.co/docs/transformers/model_doc/rwkv)** (from Bo Peng), released on [this repo](https://github.com/BlinkDL/RWKV-LM) by Bo Peng.
+1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
+1. **[Segment Anything](https://huggingface.co/docs/transformers/model_doc/sam)** (from Meta AI) released with the paper [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf) by Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.
+1. **[SEW](https://huggingface.co/docs/transformers/model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
+1. **[SEW-D](https://huggingface.co/docs/transformers/model_doc/sew_d)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
+1. **[SpeechT5](https://huggingface.co/docs/transformers/model_doc/speecht5)** (from Microsoft Research) released with the paper [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) by Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei.
+1. **[SpeechToTextTransformer](https://huggingface.co/docs/transformers/model_doc/speech_to_text)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino.
+1. **[SpeechToTextTransformer2](https://huggingface.co/docs/transformers/model_doc/speech_to_text_2)** (from Facebook), released together with the paper [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau.
+1. **[Splinter](https://huggingface.co/docs/transformers/model_doc/splinter)** (from Tel Aviv University), released together with the paper [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) by Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy.
+1. **[SqueezeBERT](https://huggingface.co/docs/transformers/model_doc/squeezebert)** (from Berkeley) released with the paper [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://arxiv.org/abs/2006.11316) by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
+1. **[SwiftFormer](https://huggingface.co/docs/transformers/model_doc/swiftformer)** (from MBZUAI) released with the paper [SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications](https://arxiv.org/abs/2303.15446) by Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, Fahad Shahbaz Khan.
+1. **[Swin Transformer](https://huggingface.co/docs/transformers/model_doc/swin)** (from Microsoft) released with the paper [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://arxiv.org/abs/2103.14030) by Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo.
+1. **[Swin Transformer V2](https://huggingface.co/docs/transformers/model_doc/swinv2)** (from Microsoft) released with the paper [Swin Transformer V2: Scaling Up Capacity and Resolution](https://arxiv.org/abs/2111.09883) by Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, Baining Guo.
+1. **[Swin2SR](https://huggingface.co/docs/transformers/model_doc/swin2sr)** (from University of Würzburg) released with the paper [Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration](https://arxiv.org/abs/2209.11345) by Marcos V. Conde, Ui-Jin Choi, Maxime Burchi, Radu Timofte.
+1. **[SwitchTransformers](https://huggingface.co/docs/transformers/model_doc/switch_transformers)** (from Google) released with the paper [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://arxiv.org/abs/2101.03961) by William Fedus, Barret Zoph, Noam Shazeer.
+1. **[T5](https://huggingface.co/docs/transformers/model_doc/t5)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
+1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (from Google AI) released in the repository [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
+1. **[Table Transformer](https://huggingface.co/docs/transformers/model_doc/table-transformer)** (from Microsoft Research) released with the paper [PubTables-1M: Towards Comprehensive Table Extraction From Unstructured Documents](https://arxiv.org/abs/2110.00061) by Brandon Smock, Rohith Pesala, Robin Abraham.
+1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
+1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
+1. **[Time Series Transformer](https://huggingface.co/docs/transformers/model_doc/time_series_transformer)** (from HuggingFace).
+1. **[TimeSformer](https://huggingface.co/docs/transformers/model_doc/timesformer)** (from Facebook) released with the paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Gedas Bertasius, Heng Wang, Lorenzo Torresani.
+1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
+1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
+1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
+1. **[TVLT](https://huggingface.co/docs/transformers/model_doc/tvlt)** (from UNC Chapel Hill) released with the paper [TVLT: Textless Vision-Language Transformer](https://arxiv.org/abs/2209.14156) by Zineng Tang, Jaemin Cho, Yixin Nie, Mohit Bansal.
+1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
+1. **[UMT5](https://huggingface.co/docs/transformers/model_doc/umt5)** (from Google Research) released with the paper [UniMax: Fairer and More Effective Language Sampling for Large-Scale Multilingual Pretraining](https://openreview.net/forum?id=kXwdL1cWOAi) by Hyung Won Chung, Xavier Garcia, Adam Roberts, Yi Tay, Orhan Firat, Sharan Narang, Noah Constant.
+1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
+1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
+1. **[UPerNet](https://huggingface.co/docs/transformers/model_doc/upernet)** (from Peking University) released with the paper [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221) by Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun.
+1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://arxiv.org/abs/2202.09741) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
+1. **[VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae)** (from Multimedia Computing Group, Nanjing University) released with the paper [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) by Zhan Tong, Yibing Song, Jue Wang, Limin Wang.
+1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
+1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
+1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
+1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
+1. **[ViTMatte](https://huggingface.co/docs/transformers/main/model_doc/vitmatte)** (from HUST-VL) rreleased with the paper [ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers](https://arxiv.org/abs/2305.15272) by Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang.
+1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
+1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
+1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
+1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
+1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/transformers/model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
+1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
+1. **[Whisper](https://huggingface.co/docs/transformers/model_doc/whisper)** (from OpenAI) released with the paper [Robust Speech Recognition via Large-Scale Weak Supervision](https://cdn.openai.com/papers/whisper.pdf) by Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, Ilya Sutskever.
+1. **[X-CLIP](https://huggingface.co/docs/transformers/model_doc/xclip)** (from Microsoft Research) released with the paper [Expanding Language-Image Pretrained Models for General Video Recognition](https://arxiv.org/abs/2208.02816) by Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, Haibin Ling.
+1. **[X-MOD](https://huggingface.co/docs/transformers/model_doc/xmod)** (from Meta AI) released with the paper [Lifting the Curse of Multilinguality by Pre-training Modular Transformers](http://dx.doi.org/10.18653/v1/2022.naacl-main.255) by Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, Mikel Artetxe.
+1. **[XGLM](https://huggingface.co/docs/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
+1. **[XLM](https://huggingface.co/docs/transformers/model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
+1. **[XLM-ProphetNet](https://huggingface.co/docs/transformers/model_doc/xlm-prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
+1. **[XLM-RoBERTa](https://huggingface.co/docs/transformers/model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
+1. **[XLM-RoBERTa-XL](https://huggingface.co/docs/transformers/model_doc/xlm-roberta-xl)** (from Facebook AI), released together with the paper [Larger-Scale Transformers for Multilingual Masked Language Modeling](https://arxiv.org/abs/2105.00572) by Naman Goyal, Jingfei Du, Myle Ott, Giri Anantharaman, Alexis Conneau.
+1. **[XLM-V](https://huggingface.co/docs/transformers/model_doc/xlm-v)** (from Meta AI) released with the paper [XLM-V: Overcoming the Vocabulary Bottleneck in Multilingual Masked Language Models](https://arxiv.org/abs/2301.10472) by Davis Liang, Hila Gonen, Yuning Mao, Rui Hou, Naman Goyal, Marjan Ghazvininejad, Luke Zettlemoyer, Madian Khabsa.
+1. **[XLNet](https://huggingface.co/docs/transformers/model_doc/xlnet)** (from Google/CMU) released with the paper [​XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
+1. **[XLS-R](https://huggingface.co/docs/transformers/model_doc/xls_r)** (from Facebook AI) released with the paper [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) by Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli.
+1. **[XLSR-Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/xlsr_wav2vec2)** (from Facebook AI) released with the paper [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://arxiv.org/abs/2006.13979) by Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
+1. **[YOLOS](https://huggingface.co/docs/transformers/model_doc/yolos)** (from Huazhong University of Science & Technology) released with the paper [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) by Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu.
+1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (from the University of Wisconsin - Madison) released with the paper [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://arxiv.org/abs/2111.09714) by Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh.
+1. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
+
+Чтобы проверить, есть ли у каждой модели реализация на Flax, PyTorch или TensorFlow, или связанный с ней токенизатор, поддерживаемый библиотекой 🤗 Tokenizers, обратитесь к [этой таблице](https://huggingface.co/docs/transformers/index#supported-frameworks).
+
+Эти реализации были протестированы на нескольких наборах данных (см. примеры скриптов) и должны соответствовать производительности оригинальных реализаций. Более подробную информацию о производительности можно найти в разделе "Примеры" [документации](https://github.com/huggingface/transformers/tree/main/examples).
+
+
+## Изучи больше
+
+| Секция | Описание |
+|-|-|
+| [Документация](https://huggingface.co/docs/transformers/) | Полная документация по API и гайды |
+| [Краткие описания задач](https://huggingface.co/docs/transformers/task_summary) | Задачи поддерживаются 🤗 Transformers |
+| [Пособие по предварительной обработке](https://huggingface.co/docs/transformers/preprocessing) | Использование класса `Tokenizer` для подготовки данных для моделей |
+| [Обучение и доработка](https://huggingface.co/docs/transformers/training) | Использование моделей, предоставляемых 🤗 Transformers, в цикле обучения PyTorch/TensorFlow и API `Trainer`. |
+| [Быстрый тур: Тонкая настройка/скрипты использования](https://github.com/huggingface/transformers/tree/main/examples) | Примеры скриптов для тонкой настройки моделей на широком спектре задач |
+| [Совместное использование и загрузка моделей](https://huggingface.co/docs/transformers/model_sharing) | Загружайте и делитесь с сообществом своими доработанными моделями |
+
+## Цитирование
+
+Теперь у нас есть [статья](https://www.aclweb.org/anthology/2020.emnlp-demos.6/), которую можно цитировать для библиотеки 🤗 Transformers:
+```bibtex
+@inproceedings{wolf-etal-2020-transformers,
+    title = "Transformers: State-of-the-Art Natural Language Processing",
+    author = "Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and Rémi Louf and Morgan Funtowicz and Joe Davison and Sam Shleifer and Patrick von Platen and Clara Ma and Yacine Jernite and Julien Plu and Canwen Xu and Teven Le Scao and Sylvain Gugger and Mariama Drame and Quentin Lhoest and Alexander M. Rush",
+    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
+    month = oct,
+    year = "2020",
+    address = "Online",
+    publisher = "Association for Computational Linguistics",
+    url = "https://www.aclweb.org/anthology/2020.emnlp-demos.6",
+    pages = "38--45"
+}
+```

README_zh-hans.md

@@ -43,7 +43,7 @@ checkpoint: 检查点
     <br>
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers_logo_name.png" width="400"/>
     <br>
-<p>
+</p>
 <p align="center">
     <a href="https://circleci.com/gh/huggingface/transformers">
         <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/main">
@@ -72,7 +72,7 @@ checkpoint: 检查点
         <a href="https://github.com/huggingface/transformers/blob/main/README_es.md">Español</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_ja.md">日本語</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_hd.md">हिन्दी</a>
-    <p>
+    </p>
 </h4>
 
 <h3 align="center">
@@ -260,6 +260,7 @@ conda install -c huggingface transformers
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (来自 Alexa) 伴随论文 [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) 由 Adrian de Wynter and Daniel J. Perry 发布。
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (来自 NAVER CLOVA) 伴随论文 [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) 由 Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park 发布。
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (来自 Google Research) 伴随论文 [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) 由 Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel 发布。
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (来自 Inria/Facebook/Sorbonne) 伴随论文 [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) 由 Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot 发布。
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (来自 Google Research) 伴随论文 [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) 由 Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting 发布。
@@ -323,6 +324,7 @@ conda install -c huggingface transformers
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by 坂本俊之(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (来自 UCSD, NVIDIA) 伴随论文 [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) 由 Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang 发布。
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (来自 Allegro.pl, AGH University of Science and Technology) 伴随论文 [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) 由 Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik 发布。
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (来自 Facebook) 伴随论文 [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) 由 Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed 发布。
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (来自 Berkeley) 伴随论文 [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) 由 Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer 发布。
 1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
@@ -356,6 +358,7 @@ conda install -c huggingface transformers
 1. **[Megatron-BERT](https://huggingface.co/docs/transformers/model_doc/megatron-bert)** (来自 NVIDIA) 伴随论文 [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) 由 Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro 发布。
 1. **[Megatron-GPT2](https://huggingface.co/docs/transformers/model_doc/megatron_gpt2)** (来自 NVIDIA) 伴随论文 [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) 由 Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro 发布。
 1. **[MGP-STR](https://huggingface.co/docs/transformers/model_doc/mgp-str)** (来自 Alibaba Research) 伴随论文 [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592) 由 Peng Wang, Cheng Da, and Cong Yao 发布。
+1. **[Mistral](https://huggingface.co/docs/transformers/model_doc/mistral)** (from Mistral AI) by The Mistral AI team: Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.. 
 1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (来自 Studio Ousia) 伴随论文 [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) 由 Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka 发布。
 1. **[MMS](https://huggingface.co/docs/transformers/model_doc/mms)** (来自 Facebook) 伴随论文 [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516) 由 Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli 发布。
 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (来自 CMU/Google Brain) 伴随论文 [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) 由 Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou 发布。
@@ -373,6 +376,7 @@ conda install -c huggingface transformers
 1. **[Nezha](https://huggingface.co/docs/transformers/model_doc/nezha)** (来自华为诺亚方舟实验室) 伴随论文 [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) 由 Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu 发布。
 1. **[NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)** (来自 Meta) 伴随论文 [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) 由 the NLLB team 发布。
 1. **[NLLB-MOE](https://huggingface.co/docs/transformers/model_doc/nllb-moe)** (来自 Meta) 伴随论文 [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) 由 the NLLB team 发布。
+1. **[Nougat](https://huggingface.co/docs/transformers/model_doc/nougat)** (来自 Meta AI) 伴随论文 [Nougat: Neural Optical Understanding for Academic Documents](https://arxiv.org/abs/2308.13418) 由 Lukas Blecher, Guillem Cucurull, Thomas Scialom, Robert Stojnic 发布。
 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (来自 the University of Wisconsin - Madison) 伴随论文 [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) 由 Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh 发布。
 1. **[OneFormer](https://huggingface.co/docs/transformers/model_doc/oneformer)** (来自 SHI Labs)  伴随论文 [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) 由 Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi 发布。
 1. **[OpenLlama](https://huggingface.co/docs/transformers/model_doc/open-llama)** (来自 [s-JoL](https://huggingface.co/s-JoL)) 由 [Open-Llama](https://github.com/s-JoL/Open-Llama) 发布.
@@ -381,11 +385,12 @@ conda install -c huggingface transformers
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (来自 Google) 伴随论文 [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) 由 Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu 发布。
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (来自 Google) 伴随论文 [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) 由 Jason Phang, Yao Zhao, Peter J. Liu 发布。
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (来自 Deepmind) 伴随论文 [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) 由 Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira 发布。
+1. **[Persimmon](https://huggingface.co/docs/transformers/model_doc/persimmon)** (来自 ADEPT) 伴随论文 [blog post](https://www.adept.ai/blog/persimmon-8b) 由 Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani 发布。
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (来自 VinAI Research) 伴随论文 [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) 由 Dat Quoc Nguyen and Anh Tuan Nguyen 发布。
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (来自 Google) 伴随论文 [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) 由 Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova 发布。
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (来自 UCLA NLP) 伴随论文 [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) 由 Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang 发布。
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (来自 Sea AI Labs) 伴随论文 [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) 由 Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng 发布。
-1. **[Pop2Piano](https://huggingface.co/docs/transformers/main/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (来自 Microsoft Research) 伴随论文 [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) 由 Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou 发布。
 1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (来自 Nanjing University, The University of Hong Kong etc.) 伴随论文 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) 由 Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao 发布。
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (来自 NVIDIA) 伴随论文 [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) 由 Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius 发布。
@@ -436,10 +441,11 @@ conda install -c huggingface transformers
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (来自 Google AI) 伴随论文 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 由 Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 发布。
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (来自 UCLA NLP) 伴随论文 [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) 由 Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang 发布。
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (来自 Google AI) 伴随论文 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 由 Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 发布。
-1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (来自 Meta AI) 伴随论文 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) 由 Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He 发布。
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (来自 Meta AI) 伴随论文 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) 由 Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He 发布。
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (来自 Meta AI) 伴随论文 [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) 由 Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick 发布。
+1. **[ViTMatte](https://huggingface.co/docs/transformers/model_doc/vitmatte)** (来自 HUST-VL) 伴随论文 [ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers](https://arxiv.org/abs/2305.15272) 由 Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang 发布。
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (来自 Meta AI) 伴随论文 [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas 发布.
-1. **[VITS](https://huggingface.co/docs/transformers/main/model_doc/vits)** (来自 Kakao Enterprise) 伴随论文 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) 由 Jaehyeon Kim, Jungil Kong, Juhee Son 发布。
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (来自 Kakao Enterprise) 伴随论文 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) 由 Jaehyeon Kim, Jungil Kong, Juhee Son 发布。
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (来自 Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) 由 Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (来自 Facebook AI) 伴随论文 [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) 由 Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli 发布。
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (来自 Facebook AI) 伴随论文 [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) 由 Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino 发布。
@@ -458,7 +464,7 @@ conda install -c huggingface transformers
 1. **[XLS-R](https://huggingface.co/docs/transformers/model_doc/xls_r)** (来自 Facebook AI) 伴随论文 [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) 由 Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli 发布。
 1. **[XLSR-Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/xlsr_wav2vec2)** (来自 Facebook AI) 伴随论文 [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://arxiv.org/abs/2006.13979) 由 Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli 发布。
 1. **[YOLOS](https://huggingface.co/docs/transformers/model_doc/yolos)** (来自 Huazhong University of Science & Technology) 伴随论文 [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) 由 Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu 发布。
-1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (来自 the University of Wisconsin - Madison) 伴随论文 [You Only Sample (Almost) 由 Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh 发布。
+1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (来自 the University of Wisconsin - Madison) 伴随论文 [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://arxiv.org/abs/2111.09714) 由 Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh 发布。
 1. 想要贡献新的模型？我们这里有一份**详细指引和模板**来引导你添加新的模型。你可以在 [`templates`](./templates) 目录中找到他们。记得查看 [贡献指南](./CONTRIBUTING.md) 并在开始写 PR 前联系维护人员或开一个新的 issue 来获得反馈。
 
 要检查某个模型是否已有 Flax、PyTorch 或 TensorFlow 的实现，或其是否在 🤗 Tokenizers 库中有对应词符化器（tokenizer），敬请参阅[此表](https://huggingface.co/docs/transformers/index#supported-frameworks)。

README_zh-hant.md

@@ -55,7 +55,7 @@ user: 使用者
     <br>
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers_logo_name.png" width="400"/>
     <br>
-<p>
+</p>
 <p align="center">
     <a href="https://circleci.com/gh/huggingface/transformers">
         <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/main">
@@ -84,7 +84,7 @@ user: 使用者
         <a href="https://github.com/huggingface/transformers/blob/main/README_es.md">Español</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_ja.md">日本語</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/README_hd.md">हिन्दी</a>
-    <p>
+    </p>
 </h4>
 
 <h3 align="center">
@@ -272,6 +272,7 @@ conda install -c huggingface transformers
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (from NAVER CLOVA) released with the paper [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
@@ -335,6 +336,7 @@ conda install -c huggingface transformers
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by 坂本俊之(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (from Allegro.pl, AGH University of Science and Technology) released with the paper [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) by Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
 1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
@@ -368,6 +370,7 @@ conda install -c huggingface transformers
 1. **[Megatron-BERT](https://huggingface.co/docs/transformers/model_doc/megatron-bert)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
 1. **[Megatron-GPT2](https://huggingface.co/docs/transformers/model_doc/megatron_gpt2)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
 1. **[MGP-STR](https://huggingface.co/docs/transformers/model_doc/mgp-str)** (from Alibaba Research) released with the paper [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592) by Peng Wang, Cheng Da, and Cong Yao.
+1. **[Mistral](https://huggingface.co/docs/transformers/model_doc/mistral)** (from Mistral AI) by The Mistral AI team: Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.. 
 1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
 1. **[MMS](https://huggingface.co/docs/transformers/model_doc/mms)** (from Facebook) released with the paper [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516) by Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli.
 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
@@ -385,6 +388,7 @@ conda install -c huggingface transformers
 1. **[Nezha](https://huggingface.co/docs/transformers/model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu.
 1. **[NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
 1. **[NLLB-MOE](https://huggingface.co/docs/transformers/model_doc/nllb-moe)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
+1. **[Nougat](https://huggingface.co/docs/transformers/model_doc/nougat)** (from Meta AI) released with the paper [Nougat: Neural Optical Understanding for Academic Documents](https://arxiv.org/abs/2308.13418) by Lukas Blecher, Guillem Cucurull, Thomas Scialom, Robert Stojnic.
 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
 1. **[OneFormer](https://huggingface.co/docs/transformers/model_doc/oneformer)** (from SHI Labs) released with the paper [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) by Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi.
 1. **[OpenLlama](https://huggingface.co/docs/transformers/model_doc/open-llama)** (from [s-JoL](https://huggingface.co/s-JoL)) released in [Open-Llama](https://github.com/s-JoL/Open-Llama).
@@ -393,11 +397,12 @@ conda install -c huggingface transformers
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, Peter J. Liu.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
+1. **[Persimmon](https://huggingface.co/docs/transformers/model_doc/persimmon)** (from ADEPT) released with the paper [blog post](https://www.adept.ai/blog/persimmon-8b) by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
-1. **[Pop2Piano](https://huggingface.co/docs/transformers/main/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
 1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
@@ -448,10 +453,11 @@ conda install -c huggingface transformers
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
-1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
+1. **[ViTMatte](https://huggingface.co/docs/transformers/model_doc/vitmatte)** (from HUST-VL) released with the paper [ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers](https://arxiv.org/abs/2305.15272) by Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
-1. **[VITS](https://huggingface.co/docs/transformers/main/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
@@ -470,7 +476,7 @@ conda install -c huggingface transformers
 1. **[XLS-R](https://huggingface.co/docs/transformers/model_doc/xls_r)** (from Facebook AI) released with the paper [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) by Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli.
 1. **[XLSR-Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/xlsr_wav2vec2)** (from Facebook AI) released with the paper [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://arxiv.org/abs/2006.13979) by Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
 1. **[YOLOS](https://huggingface.co/docs/transformers/model_doc/yolos)** (from Huazhong University of Science & Technology) released with the paper [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) by Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu.
-1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (from the University of Wisconsin - Madison) released with the paper [You Only Sample (Almost) by Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh.
+1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (from the University of Wisconsin - Madison) released with the paper [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://arxiv.org/abs/2111.09714) by Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh.
 1. 想要貢獻新的模型？我們這裡有一份**詳細指引和模板**來引導你加入新的模型。你可以在 [`templates`](./templates) 目錄中找到它們。記得查看[貢獻指引](./CONTRIBUTING.md)並在開始寫 PR 前聯繫維護人員或開一個新的 issue 來獲得 feedbacks。
 
 要檢查某個模型是否已有 Flax、PyTorch 或 TensorFlow 的實作，或其是否在🤗 Tokenizers 函式庫中有對應的 tokenizer，敬請參閱[此表](https://huggingface.co/docs/transformers/index#supported-frameworks)。

docker/transformers-pytorch-amd-gpu/Dockerfile

@@ -0,0 +1,31 @@
+FROM rocm/pytorch:rocm5.6_ubuntu20.04_py3.8_pytorch_2.0.1
+LABEL maintainer="Hugging Face"
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+RUN apt update && \
+    apt install -y --no-install-recommends git libsndfile1-dev tesseract-ocr espeak-ng python3 python3-pip ffmpeg && \
+    apt clean && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN python3 -m pip install --no-cache-dir --upgrade pip setuptools ninja git+https://github.com/facebookresearch/detectron2.git pytesseract "itsdangerous<2.1.0"
+
+# If set to nothing, will install the latest version
+ARG PYTORCH='2.0.1'
+ARG TORCH_VISION='0.15.2'
+ARG TORCH_AUDIO='2.0.2'
+ARG ROCM='5.6'
+
+RUN git clone --depth 1 --branch v$TORCH_AUDIO https://github.com/pytorch/audio.git
+RUN cd audio && USE_ROCM=1 USE_CUDA=0 python setup.py install
+
+ARG REF=main
+WORKDIR /
+RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
+RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-torch,testing,video]
+
+RUN python3 -m pip uninstall -y tensorflow flax
+
+# 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/README.md

@@ -81,7 +81,7 @@ The `preview` command only works with existing doc files. When you add a complet
 
 ## Adding a new element to the navigation bar
 
-Accepted files are Markdown (.md or .md).
+Accepted files are Markdown (.md).
 
 Create a file with its extension and put it in the source directory. You can then link it to the toc-tree by putting
 the filename without the extension in the [`_toctree.yml`](https://github.com/huggingface/transformers/blob/main/docs/source/en/_toctree.yml) file.
@@ -364,9 +364,6 @@ We use pytests' [doctest integration](https://docs.pytest.org/doctest.html) to v
 For Transformers, the doctests are run on a daily basis via GitHub Actions as can be 
 seen [here](https://github.com/huggingface/transformers/actions/workflows/doctests.yml).
 
-To include your example in the daily doctests, you need to add the filename that
-contains the example docstring to the [documentation_tests.txt](../utils/documentation_tests.txt).
-
 ### For Python files
 
 Run all the tests in the docstrings of a given file with the following command, here is how we test the modeling file of Wav2Vec2 for instance:

docs/source/de/_toctree.yml

@@ -15,8 +15,28 @@
     title: Vorverarbeiten
   - local: training
     title: Optimierung eines vortrainierten Modells
+  - local: run_scripts
+    title: Trainieren mit einem Skript
   - local: accelerate
     title: Verteiltes Training mit 🤗 Accelerate
+  - local: peft
+    title: Laden und Trainieren von Adaptern mit 🤗 PEFT
   - local: model_sharing
     title: Ein Modell teilen
+  - local: transformers_agents
+    title: Agents
+  - local: llm_tutorial
+    title: Generation with LLMs
   title: Tutorials
+- sections:
+  - local: add_new_model
+    title: Wie fügt man ein Modell zu 🤗 Transformers hinzu?
+  - local: add_tensorflow_model
+    title: Wie konvertiert man ein 🤗 Transformers-Modell in TensorFlow?
+  - local: add_new_pipeline
+    title: Wie fügt man eine Pipeline zu 🤗 Transformers hinzu?
+  - local: testing
+    title: Testen
+  - local: pr_checks
+    title: Überprüfung einer Pull Request
+  title: Contribute

docs/source/de/add_new_model.md

@@ -0,0 +1,895 @@
+<!--Copyright 2020 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
+
+⚠️ 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.
+
+-->
+
+# Wie kann ich ein Modell zu 🤗 Transformers hinzufügen?
+
+Die 🤗 Transformers-Bibliothek ist dank der Beiträge der Community oft in der Lage, neue Modelle anzubieten. Aber das kann ein anspruchsvolles Projekt sein und erfordert eine eingehende Kenntnis der 🤗 Transformers-Bibliothek und des zu implementierenden Modells. Bei Hugging Face versuchen wir, mehr Mitgliedern der Community die Möglichkeit zu geben, aktiv Modelle hinzuzufügen, und wir haben diese Anleitung zusammengestellt, die Sie durch den Prozess des Hinzufügens eines PyTorch-Modells führt (stellen Sie sicher, dass Sie [PyTorch installiert haben](https://pytorch.org/get-started/locally/)).
+
+<Tip>
+
+Wenn Sie daran interessiert sind, ein TensorFlow-Modell zu implementieren, werfen Sie einen Blick in die Anleitung [How to convert a 🤗 Transformers model to TensorFlow](add_tensorflow_model)!
+
+</Tip>
+
+Auf dem Weg dorthin, werden Sie:
+
+- Einblicke in bewährte Open-Source-Verfahren erhalten
+- die Konstruktionsprinzipien hinter einer der beliebtesten Deep-Learning-Bibliotheken verstehen
+- lernen Sie, wie Sie große Modelle effizient testen können
+- lernen Sie, wie Sie Python-Hilfsprogramme wie `black`, `ruff` und `make fix-copies` integrieren, um sauberen und lesbaren Code zu gewährleisten
+
+Ein Mitglied des Hugging Face-Teams wird Ihnen dabei zur Seite stehen, damit Sie nicht alleine sind. 🤗 ❤️
+
+Um loszulegen, öffnen Sie eine [New model addition](https://github.com/huggingface/transformers/issues/new?assignees=&labels=New+model&template=new-model-addition.yml) Ausgabe für das Modell, das Sie in 🤗 Transformers sehen möchten. Wenn Sie nicht besonders wählerisch sind, wenn es darum geht, ein bestimmtes Modell beizusteuern, können Sie nach dem [New model label](https://github.com/huggingface/transformers/labels/New%20model) filtern, um zu sehen, ob es noch unbeanspruchte Modellanfragen gibt, und daran arbeiten.
+
+Sobald Sie eine neue Modellanfrage eröffnet haben, sollten Sie sich zunächst mit 🤗 Transformers vertraut machen, falls Sie das noch nicht sind!
+
+## Allgemeiner Überblick über 🤗 Transformers
+
+Zunächst sollten Sie sich einen allgemeinen Überblick über 🤗 Transformers verschaffen. 🤗 Transformers ist eine sehr meinungsfreudige Bibliothek, es ist also möglich, dass
+Es besteht also die Möglichkeit, dass Sie mit einigen der Philosophien oder Designentscheidungen der Bibliothek nicht einverstanden sind. Aus unserer Erfahrung heraus haben wir jedoch
+dass die grundlegenden Designentscheidungen und Philosophien der Bibliothek entscheidend sind, um 🤗 Transformers effizient zu skalieren.
+Transformatoren zu skalieren und gleichzeitig die Wartungskosten auf einem vernünftigen Niveau zu halten.
+
+Ein guter erster Ansatzpunkt, um die Bibliothek besser zu verstehen, ist die Lektüre der [Dokumentation unserer Philosophie](Philosophie). Als Ergebnis unserer Arbeitsweise gibt es einige Entscheidungen, die wir versuchen, auf alle Modelle anzuwenden:
+
+- Komposition wird im Allgemeinen gegenüber Abstraktion bevorzugt
+- Die Duplizierung von Code ist nicht immer schlecht, wenn sie die Lesbarkeit oder Zugänglichkeit eines Modells stark verbessert
+- Modelldateien sind so in sich geschlossen wie möglich, so dass Sie, wenn Sie den Code eines bestimmten Modells lesen, idealerweise nur
+  in die entsprechende Datei `modeling_....py` schauen müssen.
+
+Unserer Meinung nach ist der Code der Bibliothek nicht nur ein Mittel, um ein Produkt bereitzustellen, *z.B.* die Möglichkeit, BERT für
+Inferenz zu verwenden, sondern auch als das Produkt selbst, das wir verbessern wollen. Wenn Sie also ein Modell hinzufügen, ist der Benutzer nicht nur die
+Person, die Ihr Modell verwenden wird, sondern auch jeder, der Ihren Code liest, zu verstehen versucht und ihn möglicherweise verbessert.
+
+Lassen Sie uns daher ein wenig tiefer in das allgemeine Design der Bibliothek einsteigen.
+
+### Überblick über die Modelle
+
+Um ein Modell erfolgreich hinzuzufügen, ist es wichtig, die Interaktion zwischen Ihrem Modell und seiner Konfiguration zu verstehen,
+[`PreTrainedModel`] und [`PretrainedConfig`]. Als Beispiel werden wir
+das Modell, das zu 🤗 Transformers hinzugefügt werden soll, `BrandNewBert` nennen.
+
+Schauen wir uns das mal an:
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers_overview.png"/>
+
+Wie Sie sehen, machen wir in 🤗 Transformers von der Vererbung Gebrauch, aber wir beschränken die Abstraktionsebene auf ein absolutes Minimum.
+Minimum. Es gibt nie mehr als zwei Abstraktionsebenen für ein Modell in der Bibliothek. `BrandNewBertModel`
+erbt von `BrandNewBertPreTrainedModel`, das wiederum von [`PreTrainedModel`] erbt und
+das war's. In der Regel wollen wir sicherstellen, dass ein neues Modell nur von
+[`PreTrainedModel`] abhängt. Die wichtigen Funktionalitäten, die jedem neuen Modell automatisch zur Verfügung gestellt werden, sind
+Modell automatisch bereitgestellt werden, sind [`~PreTrainedModel.from_pretrained`] und
+[`~PreTrainedModel.save_pretrained`], die für die Serialisierung und Deserialisierung verwendet werden. Alle
+anderen wichtigen Funktionalitäten, wie `BrandNewBertModel.forward` sollten vollständig in der neuen
+Skript `modeling_brand_new_bert.py` definiert werden. Als nächstes wollen wir sicherstellen, dass ein Modell mit einer bestimmten Kopfebene, wie z.B.
+`BrandNewBertForMaskedLM` nicht von `BrandNewBertModel` erbt, sondern `BrandNewBertModel` verwendet
+als Komponente, die im Forward Pass aufgerufen werden kann, um die Abstraktionsebene niedrig zu halten. Jedes neue Modell erfordert eine
+Konfigurationsklasse, genannt `BrandNewBertConfig`. Diese Konfiguration wird immer als ein Attribut in
+[PreTrainedModel] gespeichert und kann daher über das Attribut `config` für alle Klassen aufgerufen werden
+die von `BrandNewBertPreTrainedModel` erben:
+
+```python
+model = BrandNewBertModel.from_pretrained("brandy/brand_new_bert")
+model.config  # model has access to its config
+```
+
+Ähnlich wie das Modell erbt die Konfiguration grundlegende Serialisierungs- und Deserialisierungsfunktionalitäten von
+[`PretrainedConfig`]. Beachten Sie, dass die Konfiguration und das Modell immer in zwei verschiedene Formate serialisiert werden
+unterschiedliche Formate serialisiert werden - das Modell in eine *pytorch_model.bin* Datei und die Konfiguration in eine *config.json* Datei. Aufruf von
+[~PreTrainedModel.save_pretrained`] wird automatisch
+[~PretrainedConfig.save_pretrained`] auf, so dass sowohl das Modell als auch die Konfiguration gespeichert werden.
+
+
+### Code-Stil
+
+Wenn Sie Ihr neues Modell kodieren, sollten Sie daran denken, dass Transformers eine Bibliothek mit vielen Meinungen ist und dass wir selbst ein paar Macken haben
+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). 
+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.
+3. Generell ziehen wir längeren expliziten Code einem kurzen magischen Code vor.
+4. Vermeiden Sie die Unterklassifizierung von `nn.Sequential` in PyTorch, sondern unterklassifizieren Sie `nn.Module` und schreiben Sie den Vorwärtspass, so dass jeder
+   so dass jeder, der Ihren Code verwendet, ihn schnell debuggen kann, indem er Druckanweisungen oder Haltepunkte hinzufügt.
+5. Ihre Funktionssignatur sollte mit einer Typ-Annotation versehen sein. Im Übrigen sind gute Variablennamen viel lesbarer und verständlicher
+   verständlicher als Typ-Anmerkungen.
+
+### Übersicht der Tokenizer
+
+Noch nicht ganz fertig :-( Dieser Abschnitt wird bald hinzugefügt!
+
+## Schritt-für-Schritt-Rezept zum Hinzufügen eines Modells zu 🤗 Transformers
+
+Jeder hat andere Vorlieben, was die Portierung eines Modells angeht. Daher kann es sehr hilfreich sein, wenn Sie sich Zusammenfassungen ansehen
+wie andere Mitwirkende Modelle auf Hugging Face portiert haben. Hier ist eine Liste von Blogbeiträgen aus der Community, wie man ein Modell portiert:
+
+1. [Portierung eines GPT2-Modells](https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28) von [Thomas](https://huggingface.co/thomwolf)
+2. [Portierung des WMT19 MT-Modells](https://huggingface.co/blog/porting-fsmt) von [Stas](https://huggingface.co/stas)
+
+Aus Erfahrung können wir Ihnen sagen, dass die wichtigsten Dinge, die Sie beim Hinzufügen eines Modells beachten müssen, sind:
+
+- Erfinden Sie das Rad nicht neu! Die meisten Teile des Codes, den Sie für das neue 🤗 Transformers-Modell hinzufügen werden, existieren bereits
+  irgendwo in 🤗 Transformers. Nehmen Sie sich etwas Zeit, um ähnliche, bereits vorhandene Modelle und Tokenizer zu finden, die Sie kopieren können
+  von. [grep](https://www.gnu.org/software/grep/) und [rg](https://github.com/BurntSushi/ripgrep) sind Ihre
+  Freunde. Beachten Sie, dass es sehr gut möglich ist, dass der Tokenizer Ihres Modells auf einer Modellimplementierung basiert und
+  und der Modellierungscode Ihres Modells auf einer anderen. *Z.B.* Der Modellierungscode von FSMT basiert auf BART, während der Tokenizer-Code von FSMT
+  auf XLM basiert.
+- Es handelt sich eher um eine technische als um eine wissenschaftliche Herausforderung. Sie sollten mehr Zeit auf die Schaffung einer
+  eine effiziente Debugging-Umgebung zu schaffen, als zu versuchen, alle theoretischen Aspekte des Modells in dem Papier zu verstehen.
+- Bitten Sie um Hilfe, wenn Sie nicht weiterkommen! Modelle sind der Kernbestandteil von 🤗 Transformers, so dass wir bei Hugging Face mehr als
+  mehr als glücklich, Ihnen bei jedem Schritt zu helfen, um Ihr Modell hinzuzufügen. Zögern Sie nicht zu fragen, wenn Sie merken, dass Sie nicht weiterkommen.
+  Fortschritte machen.
+
+Im Folgenden versuchen wir, Ihnen ein allgemeines Rezept an die Hand zu geben, das uns bei der Portierung eines Modells auf 🤗 Transformers am nützlichsten erschien.
+
+Die folgende Liste ist eine Zusammenfassung all dessen, was getan werden muss, um ein Modell hinzuzufügen und kann von Ihnen als To-Do verwendet werden
+Liste verwenden:
+
+☐ (Optional) Verstehen der theoretischen Aspekte des Modells<br>
+☐ Vorbereiten der 🤗 Transformers-Entwicklungsumgebung<br>
+☐ Debugging-Umgebung des ursprünglichen Repositorys eingerichtet<br>
+☐ Skript erstellt, das den Durchlauf `forward()` unter Verwendung des ursprünglichen Repositorys und des Checkpoints erfolgreich durchführt<br>
+☐ Erfolgreich das Modellskelett zu 🤗 Transformers hinzugefügt<br>
+☐ Erfolgreiche Umwandlung des ursprünglichen Prüfpunkts in den 🤗 Transformers-Prüfpunkt<br>
+☐ Erfolgreich den Durchlauf `forward()` in 🤗 Transformers ausgeführt, der eine identische Ausgabe wie der ursprüngliche Prüfpunkt liefert<br>
+☐ Modell-Tests in 🤗 Transformers abgeschlossen<br>
+☐ Erfolgreich Tokenizer in 🤗 Transformers hinzugefügt<br>
+☐ End-to-End-Integrationstests ausgeführt<br>
+☐ Docs fertiggestellt<br>
+☐ Modellgewichte in den Hub hochgeladen<br>
+☐ Die Pull-Anfrage eingereicht<br>
+☐ (Optional) Hinzufügen eines Demo-Notizbuchs
+
+Für den Anfang empfehlen wir in der Regel, mit einem guten theoretischen Verständnis von `BrandNewBert` zu beginnen. Wie auch immer,
+wenn Sie es vorziehen, die theoretischen Aspekte des Modells *on-the-job* zu verstehen, dann ist es völlig in Ordnung, direkt in die
+in die Code-Basis von `BrandNewBert` einzutauchen. Diese Option könnte für Sie besser geeignet sein, wenn Ihre technischen Fähigkeiten besser sind als
+als Ihre theoretischen Fähigkeiten, wenn Sie Schwierigkeiten haben, die Arbeit von `BrandNewBert` zu verstehen, oder wenn Sie einfach Spaß am Programmieren
+mehr Spaß am Programmieren haben als am Lesen wissenschaftlicher Abhandlungen.
+
+### 1. (Optional) Theoretische Aspekte von BrandNewBert
+
+Sie sollten sich etwas Zeit nehmen, um die Abhandlung von *BrandNewBert* zu lesen, falls eine solche Beschreibung existiert. Möglicherweise gibt es große
+Abschnitte des Papiers, die schwer zu verstehen sind. Wenn das der Fall ist, ist das in Ordnung - machen Sie sich keine Sorgen! Das Ziel ist
+ist es nicht, ein tiefes theoretisches Verständnis des Papiers zu erlangen, sondern die notwendigen Informationen zu extrahieren, um
+das Modell effektiv in 🤗 Transformers zu implementieren. Das heißt, Sie müssen nicht zu viel Zeit auf die
+theoretischen Aspekten verbringen, sondern sich lieber auf die praktischen Aspekte konzentrieren, nämlich:
+
+- Welche Art von Modell ist *brand_new_bert*? BERT-ähnliches Modell nur für den Encoder? GPT2-ähnliches reines Decoder-Modell? BART-ähnliches
+  Encoder-Decoder-Modell? Sehen Sie sich die [model_summary](model_summary) an, wenn Sie mit den Unterschieden zwischen diesen Modellen nicht vertraut sind.
+- Was sind die Anwendungen von *brand_new_bert*? Textklassifizierung? Texterzeugung? Seq2Seq-Aufgaben, *z.B.,*
+  Zusammenfassungen?
+- Was ist die neue Eigenschaft des Modells, die es von BERT/GPT-2/BART unterscheidet?
+- Welches der bereits existierenden [🤗 Transformers-Modelle](https://huggingface.co/transformers/#contents) ist am ähnlichsten
+  ähnlich wie *brand_new_bert*?
+- Welche Art von Tokenizer wird verwendet? Ein Satzteil-Tokenisierer? Ein Wortstück-Tokenisierer? Ist es derselbe Tokenisierer, der für
+  für BERT oder BART?
+
+Nachdem Sie das Gefühl haben, einen guten Überblick über die Architektur des Modells erhalten zu haben, können Sie dem
+Hugging Face Team schreiben und Ihre Fragen stellen. Dazu können Fragen zur Architektur des Modells gehören,
+seiner Aufmerksamkeitsebene usw. Wir werden Ihnen gerne weiterhelfen.
+
+### 2. Bereiten Sie als nächstes Ihre Umgebung vor
+
+1. Forken Sie das [Repository](https://github.com/huggingface/transformers), indem Sie auf der Seite des Repositorys auf die Schaltfläche 'Fork' klicken.
+   Seite des Repositorys klicken. Dadurch wird eine Kopie des Codes unter Ihrem GitHub-Benutzerkonto erstellt.
+
+2. Klonen Sie Ihren `transformers` Fork auf Ihre lokale Festplatte und fügen Sie das Basis-Repository als Remote hinzu:
+
+```bash
+git clone https://github.com/[your Github handle]/transformers.git
+cd transformers
+git remote add upstream https://github.com/huggingface/transformers.git
+```
+
+3. Richten Sie eine Entwicklungsumgebung ein, indem Sie z.B. den folgenden Befehl ausführen:
+
+```bash
+python -m venv .env
+source .env/bin/activate
+pip install -e ".[dev]"
+```
+
+Abhängig von Ihrem Betriebssystem und da die Anzahl der optionalen Abhängigkeiten von Transformers wächst, kann es sein, dass Sie bei diesem Befehl einen
+Fehler mit diesem Befehl. Stellen Sie in diesem Fall sicher, dass Sie das Deep Learning Framework, mit dem Sie arbeiten, installieren
+(PyTorch, TensorFlow und/oder Flax) und führen Sie es aus:
+
+```bash
+pip install -e ".[quality]"
+```
+
+was für die meisten Anwendungsfälle ausreichend sein sollte. Sie können dann zum übergeordneten Verzeichnis zurückkehren
+
+```bash
+cd ..
+```
+
+4. Wir empfehlen, die PyTorch-Version von *brand_new_bert* zu Transformers hinzuzufügen. Um PyTorch zu installieren, folgen Sie bitte den
+   Anweisungen auf https://pytorch.org/get-started/locally/.
+
+**Anmerkung:** Sie müssen CUDA nicht installiert haben. Es reicht aus, das neue Modell auf der CPU zum Laufen zu bringen.
+
+5. Um *brand_new_bert* zu portieren, benötigen Sie außerdem Zugriff auf das Original-Repository:
+
+```bash
+git clone https://github.com/org_that_created_brand_new_bert_org/brand_new_bert.git
+cd brand_new_bert
+pip install -e .
+```
+
+Jetzt haben Sie eine Entwicklungsumgebung eingerichtet, um *brand_new_bert* auf 🤗 Transformers zu portieren.
+
+### 3.-4. Führen Sie einen Pre-Training-Checkpoint mit dem Original-Repository durch
+
+Zunächst werden Sie mit dem ursprünglichen *brand_new_bert* Repository arbeiten. Oft ist die ursprüngliche Implementierung sehr
+"forschungslastig". Das bedeutet, dass es an Dokumentation mangeln kann und der Code schwer zu verstehen sein kann. Aber das sollte
+genau Ihre Motivation sein, *brand_new_bert* neu zu implementieren. Eines unserer Hauptziele bei Hugging Face ist es, *die Menschen dazu zu bringen
+auf den Schultern von Giganten zu stehen*, was sich hier sehr gut darin ausdrückt, dass wir ein funktionierendes Modell nehmen und es umschreiben, um es so
+es so **zugänglich, benutzerfreundlich und schön** wie möglich zu machen. Dies ist die wichtigste Motivation für die Neuimplementierung von
+Modelle in 🤗 Transformers umzuwandeln - der Versuch, komplexe neue NLP-Technologie für **jeden** zugänglich zu machen.
+
+Sie sollten damit beginnen, indem Sie in das Original-Repository eintauchen.
+
+Die erfolgreiche Ausführung des offiziellen Pre-Trainingsmodells im Original-Repository ist oft **der schwierigste** Schritt.
+Unserer Erfahrung nach ist es sehr wichtig, dass Sie einige Zeit damit verbringen, sich mit der ursprünglichen Code-Basis vertraut zu machen. Sie müssen
+das Folgende herausfinden:
+
+- Wo finden Sie die vortrainierten Gewichte?
+- Wie lädt man die vorab trainierten Gewichte in das entsprechende Modell?
+- Wie kann der Tokenizer unabhängig vom Modell ausgeführt werden?
+- Verfolgen Sie einen Forward Pass, damit Sie wissen, welche Klassen und Funktionen für einen einfachen Forward Pass erforderlich sind. Normalerweise,
+  müssen Sie nur diese Funktionen reimplementieren.
+- Sie müssen in der Lage sein, die wichtigen Komponenten des Modells zu finden: Wo befindet sich die Klasse des Modells? Gibt es Unterklassen des Modells,
+  *z.B.* EncoderModel, DecoderModel? Wo befindet sich die Selbstaufmerksamkeitsschicht? Gibt es mehrere verschiedene Aufmerksamkeitsebenen,
+  *z.B.* *Selbstaufmerksamkeit*, *Kreuzaufmerksamkeit*...?
+- Wie können Sie das Modell in der ursprünglichen Umgebung des Repo debuggen? Müssen Sie *print* Anweisungen hinzufügen, können Sie
+  mit einem interaktiven Debugger wie *ipdb* arbeiten oder sollten Sie eine effiziente IDE zum Debuggen des Modells verwenden, wie z.B. PyCharm?
+
+Es ist sehr wichtig, dass Sie, bevor Sie mit der Portierung beginnen, den Code im Original-Repository **effizient** debuggen können
+Repository können! Denken Sie auch daran, dass Sie mit einer Open-Source-Bibliothek arbeiten, also zögern Sie nicht, ein Problem oder
+oder sogar eine Pull-Anfrage im Original-Repository zu stellen. Die Betreuer dieses Repositorys sind wahrscheinlich sehr froh darüber
+dass jemand in ihren Code schaut!
+
+An diesem Punkt liegt es wirklich an Ihnen, welche Debugging-Umgebung und Strategie Sie zum Debuggen des ursprünglichen
+Modell zu debuggen. Wir raten dringend davon ab, eine kostspielige GPU-Umgebung einzurichten, sondern arbeiten Sie einfach auf einer CPU, sowohl wenn Sie mit dem
+in das ursprüngliche Repository einzutauchen und auch, wenn Sie beginnen, die 🤗 Transformers-Implementierung des Modells zu schreiben. Nur
+ganz am Ende, wenn das Modell bereits erfolgreich auf 🤗 Transformers portiert wurde, sollte man überprüfen, ob das
+Modell auch auf der GPU wie erwartet funktioniert.
+
+Im Allgemeinen gibt es zwei mögliche Debugging-Umgebungen für die Ausführung des Originalmodells
+
+- [Jupyter notebooks](https://jupyter.org/) / [google colab](https://colab.research.google.com/notebooks/intro.ipynb)
+- Lokale Python-Skripte.
+
+Jupyter-Notebooks haben den Vorteil, dass sie eine zellenweise Ausführung ermöglichen, was hilfreich sein kann, um logische Komponenten besser voneinander zu trennen und
+logische Komponenten voneinander zu trennen und schnellere Debugging-Zyklen zu haben, da Zwischenergebnisse gespeichert werden können. Außerdem,
+Außerdem lassen sich Notebooks oft leichter mit anderen Mitwirkenden teilen, was sehr hilfreich sein kann, wenn Sie das Hugging Face Team um Hilfe bitten möchten.
+Face Team um Hilfe bitten. Wenn Sie mit Jupyter-Notizbüchern vertraut sind, empfehlen wir Ihnen dringend, mit ihnen zu arbeiten.
+
+Der offensichtliche Nachteil von Jupyter-Notizbüchern ist, dass Sie, wenn Sie nicht daran gewöhnt sind, mit ihnen zu arbeiten, einige Zeit damit verbringen müssen
+einige Zeit damit verbringen müssen, sich an die neue Programmierumgebung zu gewöhnen, und dass Sie möglicherweise Ihre bekannten Debugging-Tools nicht mehr verwenden können
+wie z.B. `ipdb` nicht mehr verwenden können.
+
+Für jede Codebasis ist es immer ein guter erster Schritt, einen **kleinen** vortrainierten Checkpoint zu laden und in der Lage zu sein, einen
+einzelnen Vorwärtsdurchlauf mit einem Dummy-Integer-Vektor von Eingabe-IDs als Eingabe zu reproduzieren. Ein solches Skript könnte wie folgt aussehen (in
+Pseudocode):
+
+```python
+model = BrandNewBertModel.load_pretrained_checkpoint("/path/to/checkpoint/")
+input_ids = [0, 4, 5, 2, 3, 7, 9]  # vector of input ids
+original_output = model.predict(input_ids)
+```
+
+Was die Debugging-Strategie anbelangt, so können Sie im Allgemeinen aus mehreren Strategien wählen:
+
+- Zerlegen Sie das ursprüngliche Modell in viele kleine testbare Komponenten und führen Sie für jede dieser Komponenten einen Vorwärtsdurchlauf zur
+  Überprüfung
+- Zerlegen Sie das ursprüngliche Modell nur in den ursprünglichen *Tokenizer* und das ursprüngliche *Modell*, führen Sie einen Vorwärtsdurchlauf für diese Komponenten durch
+  und verwenden Sie dazwischenliegende Druckanweisungen oder Haltepunkte zur Überprüfung.
+
+Auch hier bleibt es Ihnen überlassen, welche Strategie Sie wählen. Oft ist die eine oder die andere Strategie vorteilhaft, je nach der ursprünglichen Codebasis
+Basis.
+
+Wenn die ursprüngliche Codebasis es Ihnen erlaubt, das Modell in kleinere Teilkomponenten zu zerlegen, *z.B.* wenn die ursprüngliche
+Code-Basis problemlos im Eager-Modus ausgeführt werden kann, lohnt es sich in der Regel, dies zu tun. Es gibt einige wichtige Vorteile
+am Anfang den schwierigeren Weg zu gehen:
+
+- Wenn Sie später das ursprüngliche Modell mit der Hugging Face-Implementierung vergleichen, können Sie automatisch überprüfen, ob
+  für jede Komponente einzeln überprüfen, ob die entsprechende Komponente der 🤗 Transformers-Implementierung übereinstimmt, anstatt sich auf
+  anstatt sich auf den visuellen Vergleich über Druckanweisungen zu verlassen
+- können Sie das große Problem der Portierung eines Modells in kleinere Probleme der Portierung einzelner Komponenten zerlegen
+  einzelnen Komponenten zu zerlegen und so Ihre Arbeit besser zu strukturieren
+- Die Aufteilung des Modells in logisch sinnvolle Komponenten hilft Ihnen, einen besseren Überblick über das Design des Modells zu bekommen
+  und somit das Modell besser zu verstehen
+- In einem späteren Stadium helfen Ihnen diese komponentenweisen Tests dabei, sicherzustellen, dass keine Regressionen auftreten, während Sie fortfahren
+  Ihren Code ändern
+
+[Lysandre's](https://gist.github.com/LysandreJik/db4c948f6b4483960de5cbac598ad4ed) Integrationstests für ELECTRA
+gibt ein schönes Beispiel dafür, wie dies geschehen kann.
+
+Wenn die ursprüngliche Codebasis jedoch sehr komplex ist oder nur die Ausführung von Zwischenkomponenten in einem kompilierten Modus erlaubt,
+könnte es zu zeitaufwändig oder sogar unmöglich sein, das Modell in kleinere testbare Teilkomponenten zu zerlegen. Ein gutes
+Beispiel ist die [T5's MeshTensorFlow](https://github.com/tensorflow/mesh/tree/master/mesh_tensorflow) Bibliothek, die sehr komplex ist
+sehr komplex ist und keine einfache Möglichkeit bietet, das Modell in seine Unterkomponenten zu zerlegen. Bei solchen Bibliotheken ist man
+oft auf die Überprüfung von Druckanweisungen angewiesen.
+
+Unabhängig davon, welche Strategie Sie wählen, ist die empfohlene Vorgehensweise oft die gleiche, nämlich dass Sie mit der Fehlersuche in den
+die Anfangsebenen zuerst und die Endebenen zuletzt debuggen.
+
+Es wird empfohlen, dass Sie die Ausgaben der folgenden Ebenen abrufen, entweder durch Druckanweisungen oder Unterkomponentenfunktionen
+Schichten in der folgenden Reihenfolge abrufen:
+
+1. Rufen Sie die Eingabe-IDs ab, die an das Modell übergeben wurden
+2. Rufen Sie die Worteinbettungen ab
+3. Rufen Sie die Eingabe der ersten Transformer-Schicht ab
+4. Rufen Sie die Ausgabe der ersten Transformer-Schicht ab
+5. Rufen Sie die Ausgabe der folgenden n - 1 Transformer-Schichten ab
+6. Rufen Sie die Ausgabe des gesamten BrandNewBert Modells ab
+
+Die Eingabe-IDs sollten dabei aus einem Array von Ganzzahlen bestehen, *z.B.* `input_ids = [0, 4, 4, 3, 2, 4, 1, 7, 19]`
+
+Die Ausgaben der folgenden Schichten bestehen oft aus mehrdimensionalen Float-Arrays und können wie folgt aussehen:
+
+```
+[[
+ [-0.1465, -0.6501,  0.1993,  ...,  0.1451,  0.3430,  0.6024],
+ [-0.4417, -0.5920,  0.3450,  ..., -0.3062,  0.6182,  0.7132],
+ [-0.5009, -0.7122,  0.4548,  ..., -0.3662,  0.6091,  0.7648],
+ ...,
+ [-0.5613, -0.6332,  0.4324,  ..., -0.3792,  0.7372,  0.9288],
+ [-0.5416, -0.6345,  0.4180,  ..., -0.3564,  0.6992,  0.9191],
+ [-0.5334, -0.6403,  0.4271,  ..., -0.3339,  0.6533,  0.8694]]],
+```
+
+Wir erwarten, dass jedes zu 🤗 Transformers hinzugefügte Modell eine Reihe von Integrationstests besteht, was bedeutet, dass das ursprüngliche
+Modell und die neu implementierte Version in 🤗 Transformers exakt dieselbe Ausgabe liefern müssen, und zwar mit einer Genauigkeit von 0,001!
+Da es normal ist, dass das exakt gleiche Modell, das in verschiedenen Bibliotheken geschrieben wurde, je nach Bibliotheksrahmen eine leicht unterschiedliche Ausgabe liefern kann
+eine leicht unterschiedliche Ausgabe liefern kann, akzeptieren wir eine Fehlertoleranz von 1e-3 (0,001). Es reicht nicht aus, wenn das Modell
+fast das gleiche Ergebnis liefert, sie müssen fast identisch sein. Daher werden Sie sicherlich die Zwischenergebnisse
+Zwischenergebnisse der 🤗 Transformers-Version mehrfach mit den Zwischenergebnissen der ursprünglichen Implementierung von
+*brand_new_bert* vergleichen. In diesem Fall ist eine **effiziente** Debugging-Umgebung des ursprünglichen Repositorys absolut
+wichtig ist. Hier sind einige Ratschläge, um Ihre Debugging-Umgebung so effizient wie möglich zu gestalten.
+
+- Finden Sie den besten Weg, um Zwischenergebnisse zu debuggen. Ist das ursprüngliche Repository in PyTorch geschrieben? Dann sollten Sie
+  dann sollten Sie sich wahrscheinlich die Zeit nehmen, ein längeres Skript zu schreiben, das das ursprüngliche Modell in kleinere Unterkomponenten zerlegt, um
+  Zwischenwerte abzurufen. Ist das ursprüngliche Repository in Tensorflow 1 geschrieben? Dann müssen Sie sich möglicherweise auf die
+  TensorFlow Druckoperationen wie [tf.print](https://www.tensorflow.org/api_docs/python/tf/print) verlassen, um die
+  Zwischenwerte auszugeben. Ist das ursprüngliche Repository in Jax geschrieben? Dann stellen Sie sicher, dass das Modell **nicht jitted** ist, wenn
+  wenn Sie den Vorwärtsdurchlauf ausführen, *z.B.* schauen Sie sich [dieser Link](https://github.com/google/jax/issues/196) an.
+- Verwenden Sie den kleinsten vortrainierten Prüfpunkt, den Sie finden können. Je kleiner der Prüfpunkt ist, desto schneller wird Ihr Debugging-Zyklus
+  wird. Es ist nicht effizient, wenn Ihr vorab trainiertes Modell so groß ist, dass Ihr Vorwärtsdurchlauf mehr als 10 Sekunden dauert.
+  Falls nur sehr große Checkpoints verfügbar sind, kann es sinnvoller sein, ein Dummy-Modell in der neuen
+  Umgebung mit zufällig initialisierten Gewichten zu erstellen und diese Gewichte zum Vergleich mit der 🤗 Transformers-Version
+  Ihres Modells
+- Vergewissern Sie sich, dass Sie den einfachsten Weg wählen, um einen Forward Pass im ursprünglichen Repository aufzurufen. Idealerweise sollten Sie
+  die Funktion im originalen Repository finden, die **nur** einen einzigen Vorwärtspass aufruft, *d.h.* die oft aufgerufen wird
+  Vorhersagen", "Auswerten", "Vorwärts" oder "Aufruf" genannt wird. Sie wollen keine Funktion debuggen, die `forward` aufruft
+  mehrfach aufruft, *z.B.* um Text zu erzeugen, wie `autoregressive_sample`, `generate`.
+- Versuchen Sie, die Tokenisierung vom *Forward*-Pass des Modells zu trennen. Wenn das Original-Repository Beispiele zeigt, bei denen
+  Sie eine Zeichenkette eingeben müssen, dann versuchen Sie herauszufinden, an welcher Stelle im Vorwärtsaufruf die Zeichenketteneingabe in Eingabe-IDs geändert wird
+  geändert wird und beginnen Sie an dieser Stelle. Das könnte bedeuten, dass Sie möglicherweise selbst ein kleines Skript schreiben oder den
+  Originalcode so ändern müssen, dass Sie die ids direkt eingeben können, anstatt eine Zeichenkette einzugeben.
+- Vergewissern Sie sich, dass sich das Modell in Ihrem Debugging-Setup **nicht** im Trainingsmodus befindet, der oft dazu führt, dass das Modell
+  Dies führt häufig zu zufälligen Ergebnissen, da das Modell mehrere Dropout-Schichten enthält. Stellen Sie sicher, dass der Vorwärtsdurchlauf in Ihrer Debugging
+  Umgebung **deterministisch** ist, damit die Dropout-Schichten nicht verwendet werden. Oder verwenden Sie *transformers.utils.set_seed*.
+  wenn sich die alte und die neue Implementierung im selben Framework befinden.
+
+Im folgenden Abschnitt finden Sie genauere Details/Tipps, wie Sie dies für *brand_new_bert* tun können.
+
+### 5.-14. Portierung von BrandNewBert auf 🤗 Transformatoren
+
+Als nächstes können Sie endlich damit beginnen, neuen Code zu 🤗 Transformers hinzuzufügen. Gehen Sie in den Klon Ihres 🤗 Transformers Forks:
+
+```bash
+cd transformers
+```
+
+In dem speziellen Fall, dass Sie ein Modell hinzufügen, dessen Architektur genau mit der Modellarchitektur eines
+Modells übereinstimmt, müssen Sie nur ein Konvertierungsskript hinzufügen, wie in [diesem Abschnitt](#write-a-conversion-script) beschrieben.
+In diesem Fall können Sie einfach die gesamte Modellarchitektur des bereits vorhandenen Modells wiederverwenden.
+
+Andernfalls beginnen wir mit der Erstellung eines neuen Modells. Sie haben hier zwei Möglichkeiten:
+
+- `transformers-cli add-new-model-like`, um ein neues Modell wie ein bestehendes hinzuzufügen
+- `transformers-cli add-new-model`, um ein neues Modell aus unserer Vorlage hinzuzufügen (sieht dann aus wie BERT oder Bart, je nachdem, welche Art von Modell Sie wählen)
+
+In beiden Fällen werden Sie mit einem Fragebogen aufgefordert, die grundlegenden Informationen zu Ihrem Modell auszufüllen. Für den zweiten Befehl müssen Sie `cookiecutter` installieren, weitere Informationen dazu finden Sie [hier](https://github.com/huggingface/transformers/tree/main/templates/adding_a_new_model).
+
+**Eröffnen Sie einen Pull Request auf dem Haupt-Repositorium huggingface/transformers**
+
+Bevor Sie mit der Anpassung des automatisch generierten Codes beginnen, ist es nun an der Zeit, einen "Work in progress (WIP)" Pull
+Anfrage, *z.B.* "[WIP] Add *brand_new_bert*", in 🤗 Transformers zu öffnen, damit Sie und das Hugging Face Team
+Seite an Seite an der Integration des Modells in 🤗 Transformers arbeiten können.
+
+Sie sollten Folgendes tun:
+
+1. Erstellen Sie eine Verzweigung mit einem beschreibenden Namen von Ihrer Hauptverzweigung
+
+```bash
+git checkout -b add_brand_new_bert
+```
+
+2. Bestätigen Sie den automatisch generierten Code:
+
+```bash
+git add .
+git commit
+```
+
+3. Abrufen und zurücksetzen auf die aktuelle Haupt
+
+```bash
+git fetch upstream
+git rebase upstream/main
+```
+
+4. Übertragen Sie die Änderungen auf Ihr Konto mit:
+
+```bash
+git push -u origin a-descriptive-name-for-my-changes
+```
+
+5. Wenn Sie zufrieden sind, gehen Sie auf die Webseite Ihrer Abspaltung auf GitHub. Klicken Sie auf "Pull request". Stellen Sie sicher, dass Sie das
+   GitHub-Handle einiger Mitglieder des Hugging Face-Teams als Reviewer hinzuzufügen, damit das Hugging Face-Team über zukünftige Änderungen informiert wird.
+   zukünftige Änderungen benachrichtigt wird.
+
+6. Ändern Sie den PR in einen Entwurf, indem Sie auf der rechten Seite der GitHub-Pull-Request-Webseite auf "In Entwurf umwandeln" klicken.
+
+Vergessen Sie im Folgenden nicht, wenn Sie Fortschritte gemacht haben, Ihre Arbeit zu committen und in Ihr Konto zu pushen, damit sie in der Pull-Anfrage erscheint.
+damit sie in der Pull-Anfrage angezeigt wird. Außerdem sollten Sie darauf achten, dass Sie Ihre Arbeit von Zeit zu Zeit mit dem aktuellen main
+von Zeit zu Zeit zu aktualisieren, indem Sie dies tun:
+
+```bash
+git fetch upstream
+git merge upstream/main
+```
+
+Generell sollten Sie alle Fragen, die Sie in Bezug auf das Modell oder Ihre Implementierung haben, in Ihrem PR stellen und
+in der PR diskutiert/gelöst werden. Auf diese Weise wird das Hugging Face Team immer benachrichtigt, wenn Sie neuen Code einreichen oder
+wenn Sie eine Frage haben. Es ist oft sehr hilfreich, das Hugging Face-Team auf Ihren hinzugefügten Code hinzuweisen, damit das Hugging Face-Team Ihr Problem oder Ihre Frage besser verstehen kann.
+Face-Team Ihr Problem oder Ihre Frage besser verstehen kann.
+
+Gehen Sie dazu auf die Registerkarte "Geänderte Dateien", auf der Sie alle Ihre Änderungen sehen, gehen Sie zu einer Zeile, zu der Sie eine Frage stellen möchten
+eine Frage stellen möchten, und klicken Sie auf das "+"-Symbol, um einen Kommentar hinzuzufügen. Wenn eine Frage oder ein Problem gelöst wurde,
+können Sie auf die Schaltfläche "Lösen" des erstellten Kommentars klicken.
+
+Auf dieselbe Weise wird das Hugging Face-Team Kommentare öffnen, wenn es Ihren Code überprüft. Wir empfehlen, die meisten Fragen
+auf GitHub in Ihrem PR zu stellen. Für einige sehr allgemeine Fragen, die für die Öffentlichkeit nicht sehr nützlich sind, können Sie das
+Hugging Face Team per Slack oder E-Mail zu stellen.
+
+**5. Passen Sie den Code der generierten Modelle für brand_new_bert** an.
+
+Zunächst werden wir uns nur auf das Modell selbst konzentrieren und uns nicht um den Tokenizer kümmern. Den gesamten relevanten Code sollten Sie
+finden Sie in den generierten Dateien `src/transformers/models/brand_new_bert/modeling_brand_new_bert.py` und
+`src/transformers/models/brand_new_bert/configuration_brand_new_bert.py`.
+
+Jetzt können Sie endlich mit dem Programmieren beginnen :). Der generierte Code in
+`src/transformers/models/brand_new_bert/modeling_brand_new_bert.py` wird entweder die gleiche Architektur wie BERT haben, wenn
+wenn es sich um ein reines Encoder-Modell handelt oder BART, wenn es sich um ein Encoder-Decoder-Modell handelt. An diesem Punkt sollten Sie sich daran erinnern, was
+was Sie am Anfang über die theoretischen Aspekte des Modells gelernt haben: *Wie unterscheidet sich das Modell von BERT oder
+BART?*". Implementieren Sie diese Änderungen, was oft bedeutet, dass Sie die *Selbstaufmerksamkeitsschicht*, die Reihenfolge der Normalisierungsschicht usw. ändern müssen.
+Schicht usw... Auch hier ist es oft nützlich, sich die ähnliche Architektur bereits bestehender Modelle in Transformers anzusehen, um ein besseres Gefühl dafür zu bekommen
+ein besseres Gefühl dafür zu bekommen, wie Ihr Modell implementiert werden sollte.
+
+**Beachten Sie**, dass Sie an diesem Punkt nicht sehr sicher sein müssen, dass Ihr Code völlig korrekt oder sauber ist. Vielmehr ist es
+Sie sollten vielmehr eine erste *unbereinigte*, kopierte Version des ursprünglichen Codes in
+src/transformers/models/brand_new_bert/modeling_brand_new_bert.py" hinzuzufügen, bis Sie das Gefühl haben, dass der gesamte notwendige Code
+hinzugefügt wurde. Unserer Erfahrung nach ist es viel effizienter, schnell eine erste Version des erforderlichen Codes hinzuzufügen und
+den Code iterativ mit dem Konvertierungsskript zu verbessern/korrigieren, wie im nächsten Abschnitt beschrieben. Das einzige, was
+zu diesem Zeitpunkt funktionieren muss, ist, dass Sie die 🤗 Transformers-Implementierung von *brand_new_bert* instanziieren können, *d.h.* der
+folgende Befehl sollte funktionieren:
+
+```python
+from transformers import BrandNewBertModel, BrandNewBertConfig
+
+model = BrandNewBertModel(BrandNewBertConfig())
+```
+
+Der obige Befehl erstellt ein Modell gemäß den Standardparametern, die in `BrandNewBertConfig()` definiert sind, mit
+zufälligen Gewichten und stellt damit sicher, dass die `init()` Methoden aller Komponenten funktionieren.
+
+Beachten Sie, dass alle zufälligen Initialisierungen in der Methode `_init_weights` Ihres `BrandnewBertPreTrainedModel` stattfinden sollten.
+Klasse erfolgen sollte. Sie sollte alle Blattmodule in Abhängigkeit von den Variablen der Konfiguration initialisieren. Hier ist ein Beispiel mit der
+BERT `_init_weights` Methode:
+
+```py
+def _init_weights(self, module):
+    """Initialize the weights"""
+    if isinstance(module, nn.Linear):
+        module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        if module.bias is not None:
+            module.bias.data.zero_()
+    elif isinstance(module, nn.Embedding):
+        module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        if module.padding_idx is not None:
+            module.weight.data[module.padding_idx].zero_()
+    elif isinstance(module, nn.LayerNorm):
+        module.bias.data.zero_()
+        module.weight.data.fill_(1.0)
+```
+
+Sie können weitere benutzerdefinierte Schemata verwenden, wenn Sie eine spezielle Initialisierung für einige Module benötigen. Zum Beispiel in
+`Wav2Vec2ForPreTraining` müssen die letzten beiden linearen Schichten die Initialisierung des regulären PyTorch `nn.Linear` haben.
+aber alle anderen sollten eine Initialisierung wie oben verwenden. Dies ist wie folgt kodiert:
+
+```py
+def _init_weights(self, module):
+    """Initialize the weights"""
+    if isinstnace(module, Wav2Vec2ForPreTraining):
+        module.project_hid.reset_parameters()
+        module.project_q.reset_parameters()
+        module.project_hid._is_hf_initialized = True
+        module.project_q._is_hf_initialized = True
+    elif isinstance(module, nn.Linear):
+        module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        if module.bias is not None:
+            module.bias.data.zero_()
+```
+
+Das Flag `_is_hf_initialized` wird intern verwendet, um sicherzustellen, dass wir ein Submodul nur einmal initialisieren. Wenn Sie es auf
+True` für `module.project_q` und `module.project_hid` setzen, stellen wir sicher, dass die benutzerdefinierte Initialisierung, die wir vorgenommen haben, später nicht überschrieben wird,
+die Funktion `_init_weights` nicht auf sie angewendet wird.
+
+**6. Schreiben Sie ein Konvertierungsskript**
+
+Als nächstes sollten Sie ein Konvertierungsskript schreiben, mit dem Sie den Checkpoint, den Sie zum Debuggen von *brand_new_bert* im
+im ursprünglichen Repository in einen Prüfpunkt konvertieren, der mit Ihrer gerade erstellten 🤗 Transformers-Implementierung von
+*brand_new_bert*. Es ist nicht ratsam, das Konvertierungsskript von Grund auf neu zu schreiben, sondern die bereits
+bestehenden Konvertierungsskripten in 🤗 Transformers nach einem Skript zu suchen, das für die Konvertierung eines ähnlichen Modells verwendet wurde, das im
+demselben Framework wie *brand_new_bert* geschrieben wurde. Normalerweise reicht es aus, ein bereits vorhandenes Konvertierungsskript zu kopieren und
+es für Ihren Anwendungsfall leicht anzupassen. Zögern Sie nicht, das Hugging Face Team zu bitten, Sie auf ein ähnliches, bereits vorhandenes
+Konvertierungsskript für Ihr Modell zu finden.
+
+- Wenn Sie ein Modell von TensorFlow nach PyTorch portieren, ist ein guter Ausgangspunkt das Konvertierungsskript von BERT [hier] (https://github.com/huggingface/transformers/blob/7acfa95afb8194f8f9c1f4d2c6028224dbed35a2/src/transformers/models/bert/modeling_bert.py#L91)
+- Wenn Sie ein Modell von PyTorch nach PyTorch portieren, ist ein guter Ausgangspunkt das Konvertierungsskript von BART [hier](https://github.com/huggingface/transformers/blob/main/src/transformers/models/bart/convert_bart_original_pytorch_checkpoint_to_pytorch.py)
+
+Im Folgenden werden wir kurz erklären, wie PyTorch-Modelle Ebenengewichte speichern und Ebenennamen definieren. In PyTorch wird der
+Name einer Ebene durch den Namen des Klassenattributs definiert, das Sie der Ebene geben. Lassen Sie uns ein Dummy-Modell in
+PyTorch, das wir `SimpleModel` nennen, wie folgt:
+
+```python
+from torch import nn
+
+
+class SimpleModel(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.dense = nn.Linear(10, 10)
+        self.intermediate = nn.Linear(10, 10)
+        self.layer_norm = nn.LayerNorm(10)
+```
+
+Jetzt können wir eine Instanz dieser Modelldefinition erstellen, die alle Gewichte ausfüllt: `dense`, `intermediate`,
+`layer_norm` mit zufälligen Gewichten. Wir können das Modell ausdrucken, um seine Architektur zu sehen
+
+```python
+model = SimpleModel()
+
+print(model)
+```
+
+Dies gibt folgendes aus:
+
+```
+SimpleModel(
+  (dense): Linear(in_features=10, out_features=10, bias=True)
+  (intermediate): Linear(in_features=10, out_features=10, bias=True)
+  (layer_norm): LayerNorm((10,), eps=1e-05, elementwise_affine=True)
+)
+```
+
+Wir können sehen, dass die Ebenennamen durch den Namen des Klassenattributs in PyTorch definiert sind. Sie können die Gewichtswerte
+Werte einer bestimmten Ebene anzeigen lassen:
+
+```python
+print(model.dense.weight.data)
+```
+
+um zu sehen, dass die Gewichte zufällig initialisiert wurden
+
+```
+tensor([[-0.0818,  0.2207, -0.0749, -0.0030,  0.0045, -0.1569, -0.1598,  0.0212,
+         -0.2077,  0.2157],
+        [ 0.1044,  0.0201,  0.0990,  0.2482,  0.3116,  0.2509,  0.2866, -0.2190,
+          0.2166, -0.0212],
+        [-0.2000,  0.1107, -0.1999, -0.3119,  0.1559,  0.0993,  0.1776, -0.1950,
+         -0.1023, -0.0447],
+        [-0.0888, -0.1092,  0.2281,  0.0336,  0.1817, -0.0115,  0.2096,  0.1415,
+         -0.1876, -0.2467],
+        [ 0.2208, -0.2352, -0.1426, -0.2636, -0.2889, -0.2061, -0.2849, -0.0465,
+          0.2577,  0.0402],
+        [ 0.1502,  0.2465,  0.2566,  0.0693,  0.2352, -0.0530,  0.1859, -0.0604,
+          0.2132,  0.1680],
+        [ 0.1733, -0.2407, -0.1721,  0.1484,  0.0358, -0.0633, -0.0721, -0.0090,
+          0.2707, -0.2509],
+        [-0.1173,  0.1561,  0.2945,  0.0595, -0.1996,  0.2988, -0.0802,  0.0407,
+          0.1829, -0.1568],
+        [-0.1164, -0.2228, -0.0403,  0.0428,  0.1339,  0.0047,  0.1967,  0.2923,
+          0.0333, -0.0536],
+        [-0.1492, -0.1616,  0.1057,  0.1950, -0.2807, -0.2710, -0.1586,  0.0739,
+          0.2220,  0.2358]]).
+```
+
+Im Konvertierungsskript sollten Sie diese zufällig initialisierten Gewichte mit den genauen Gewichten der
+entsprechenden Ebene im Kontrollpunkt. *Z.B.*
+
+```python
+# retrieve matching layer weights, e.g. by
+# recursive algorithm
+layer_name = "dense"
+pretrained_weight = array_of_dense_layer
+
+model_pointer = getattr(model, "dense")
+
+model_pointer.weight.data = torch.from_numpy(pretrained_weight)
+```
+
+Dabei müssen Sie sicherstellen, dass jedes zufällig initialisierte Gewicht Ihres PyTorch-Modells und sein entsprechendes
+Checkpoint-Gewicht in **Form und Name** genau übereinstimmen. Zu diesem Zweck ist es **notwendig**, assert
+Anweisungen für die Form hinzuzufügen und die Namen der Checkpoint-Gewichte auszugeben. Sie sollten z.B. Anweisungen hinzufügen wie:
+
+```python
+assert (
+    model_pointer.weight.shape == pretrained_weight.shape
+), f"Pointer shape of random weight {model_pointer.shape} and array shape of checkpoint weight {pretrained_weight.shape} mismatched"
+```
+
+Außerdem sollten Sie die Namen der beiden Gewichte ausdrucken, um sicherzustellen, dass sie übereinstimmen, *z.B.*.
+
+```python
+logger.info(f"Initialize PyTorch weight {layer_name} from {pretrained_weight.name}")
+```
+
+Wenn entweder die Form oder der Name nicht übereinstimmt, haben Sie wahrscheinlich das falsche Kontrollpunktgewicht einer zufällig
+Ebene der 🤗 Transformers-Implementierung zugewiesen.
+
+Eine falsche Form ist höchstwahrscheinlich auf eine falsche Einstellung der Konfigurationsparameter in `BrandNewBertConfig()` zurückzuführen, die
+nicht genau mit denen übereinstimmen, die für den zu konvertierenden Prüfpunkt verwendet wurden. Es könnte aber auch sein, dass
+die PyTorch-Implementierung eines Layers erfordert, dass das Gewicht vorher transponiert wird.
+
+Schließlich sollten Sie auch überprüfen, ob **alle** erforderlichen Gewichte initialisiert sind und alle Checkpoint-Gewichte ausgeben, die
+die nicht zur Initialisierung verwendet wurden, um sicherzustellen, dass das Modell korrekt konvertiert wurde. Es ist völlig normal, dass die
+Konvertierungsversuche entweder mit einer falschen Shape-Anweisung oder einer falschen Namenszuweisung fehlschlagen. Das liegt höchstwahrscheinlich daran, dass entweder
+Sie haben falsche Parameter in `BrandNewBertConfig()` verwendet, haben eine falsche Architektur in der 🤗 Transformers
+Implementierung, Sie haben einen Fehler in den `init()` Funktionen einer der Komponenten der 🤗 Transformers
+Implementierung oder Sie müssen eine der Kontrollpunktgewichte transponieren.
+
+Dieser Schritt sollte mit dem vorherigen Schritt wiederholt werden, bis alle Gewichte des Kontrollpunkts korrekt in das
+Transformers-Modell geladen sind. Nachdem Sie den Prüfpunkt korrekt in die 🤗 Transformers-Implementierung geladen haben, können Sie das Modell
+das Modell unter einem Ordner Ihrer Wahl `/path/to/converted/checkpoint/folder` speichern, der dann sowohl ein
+Datei `pytorch_model.bin` und eine Datei `config.json` enthalten sollte:
+
+```python
+model.save_pretrained("/path/to/converted/checkpoint/folder")
+```
+
+**7. Implementieren Sie den Vorwärtspass**
+
+Nachdem es Ihnen gelungen ist, die trainierten Gewichte korrekt in die 🤗 Transformers-Implementierung zu laden, sollten Sie nun dafür sorgen
+sicherstellen, dass der Forward Pass korrekt implementiert ist. In [Machen Sie sich mit dem ursprünglichen Repository vertraut](#34-run-a-pretrained-checkpoint-using-the-original-repository) haben Sie bereits ein Skript erstellt, das einen Forward Pass
+Durchlauf des Modells unter Verwendung des Original-Repositorys durchführt. Jetzt sollten Sie ein analoges Skript schreiben, das die 🤗 Transformers
+Implementierung anstelle der Originalimplementierung verwenden. Es sollte wie folgt aussehen:
+
+```python
+model = BrandNewBertModel.from_pretrained("/path/to/converted/checkpoint/folder")
+input_ids = [0, 4, 4, 3, 2, 4, 1, 7, 19]
+output = model(input_ids).last_hidden_states
+```
+
+Es ist sehr wahrscheinlich, dass die 🤗 Transformers-Implementierung und die ursprüngliche Modell-Implementierung nicht genau die gleiche Ausgabe liefern.
+beim ersten Mal nicht die gleiche Ausgabe liefern oder dass der Vorwärtsdurchlauf einen Fehler auslöst. Seien Sie nicht enttäuscht - das ist zu erwarten! Erstens,
+sollten Sie sicherstellen, dass der Vorwärtsdurchlauf keine Fehler auslöst. Es passiert oft, dass die falschen Dimensionen verwendet werden
+verwendet werden, was zu einem *Dimensionality mismatch* Fehler führt oder dass der falsche Datentyp verwendet wird, *z.B.* `torch.long`
+anstelle von `torch.float32`. Zögern Sie nicht, das Hugging Face Team um Hilfe zu bitten, wenn Sie bestimmte Fehler nicht lösen können.
+bestimmte Fehler nicht lösen können.
+
+Um sicherzustellen, dass die Implementierung von 🤗 Transformers korrekt funktioniert, müssen Sie sicherstellen, dass die Ausgaben
+einer Genauigkeit von `1e-3` entsprechen. Zunächst sollten Sie sicherstellen, dass die Ausgabeformen identisch sind, *d.h.*.
+Die Ausgabeform *outputs.shape* sollte für das Skript der 🤗 Transformers-Implementierung und die ursprüngliche
+Implementierung ergeben. Als nächstes sollten Sie sicherstellen, dass auch die Ausgabewerte identisch sind. Dies ist einer der schwierigsten
+Teile des Hinzufügens eines neuen Modells. Häufige Fehler, warum die Ausgaben nicht identisch sind, sind:
+
+- Einige Ebenen wurden nicht hinzugefügt, *d.h.* eine *Aktivierungsebene* wurde nicht hinzugefügt, oder die Restverbindung wurde vergessen
+- Die Worteinbettungsmatrix wurde nicht gebunden
+- Es werden die falschen Positionseinbettungen verwendet, da die ursprüngliche Implementierung einen Offset verwendet
+- Dropout wird während des Vorwärtsdurchlaufs angewendet. Um dies zu beheben, stellen Sie sicher, dass *model.training auf False* steht und dass keine Dropout
+  Schicht während des Vorwärtsdurchlaufs fälschlicherweise aktiviert wird, *d.h.* übergeben Sie *self.training* an [PyTorch's functional dropout](https://pytorch.org/docs/stable/nn.functional.html?highlight=dropout#torch.nn.functional.dropout)
+
+Der beste Weg, das Problem zu beheben, besteht normalerweise darin, sich den Vorwärtsdurchlauf der ursprünglichen Implementierung und die 🤗
+Transformers-Implementierung nebeneinander zu sehen und zu prüfen, ob es Unterschiede gibt. Idealerweise sollten Sie die
+Zwischenergebnisse beider Implementierungen des Vorwärtsdurchlaufs debuggen/ausdrucken, um die genaue Position im Netzwerk zu finden, an der die 🤗
+Transformers-Implementierung eine andere Ausgabe zeigt als die ursprüngliche Implementierung. Stellen Sie zunächst sicher, dass die
+hartcodierten `input_ids` in beiden Skripten identisch sind. Überprüfen Sie dann, ob die Ausgaben der ersten Transformation von
+der `input_ids` (normalerweise die Worteinbettungen) identisch sind. Und dann arbeiten Sie sich bis zur allerletzten Schicht des
+Netzwerks. Irgendwann werden Sie einen Unterschied zwischen den beiden Implementierungen feststellen, der Sie auf den Fehler
+in der Implementierung von 🤗 Transformers hinweist. Unserer Erfahrung nach ist ein einfacher und effizienter Weg, viele Druckanweisungen hinzuzufügen
+sowohl in der Original-Implementierung als auch in der 🤗 Transformers-Implementierung an den gleichen Stellen im Netzwerk
+hinzuzufügen und nacheinander Druckanweisungen zu entfernen, die dieselben Werte für Zwischenpräsentationen anzeigen.
+
+Wenn Sie sicher sind, dass beide Implementierungen die gleiche Ausgabe liefern, überprüfen Sie die Ausgaben mit
+`torch.allclose(original_output, output, atol=1e-3)` überprüfen, haben Sie den schwierigsten Teil hinter sich! Herzlichen Glückwunsch - die
+Arbeit, die noch zu erledigen ist, sollte ein Kinderspiel sein 😊.
+
+**8. Hinzufügen aller notwendigen Modelltests**
+
+An diesem Punkt haben Sie erfolgreich ein neues Modell hinzugefügt. Es ist jedoch sehr gut möglich, dass das Modell noch nicht
+noch nicht vollständig mit dem erforderlichen Design übereinstimmt. Um sicherzustellen, dass die Implementierung vollständig kompatibel mit 🤗 Transformers ist, sollten alle
+gemeinsamen Tests bestehen. Der Cookiecutter sollte automatisch eine Testdatei für Ihr Modell hinzugefügt haben, wahrscheinlich unter
+demselben `tests/models/brand_new_bert/test_modeling_brand_new_bert.py`. Führen Sie diese Testdatei aus, um zu überprüfen, ob alle gängigen
+Tests bestehen:
+
+```bash
+pytest tests/models/brand_new_bert/test_modeling_brand_new_bert.py
+```
+
+Nachdem Sie alle allgemeinen Tests festgelegt haben, müssen Sie nun sicherstellen, dass all die schöne Arbeit, die Sie geleistet haben, gut getestet ist, damit
+
+- a) die Community Ihre Arbeit leicht nachvollziehen kann, indem sie sich spezifische Tests von *brand_new_bert* ansieht
+- b) zukünftige Änderungen an Ihrem Modell keine wichtigen Funktionen des Modells zerstören.
+
+Als erstes sollten Sie Integrationstests hinzufügen. Diese Integrationstests tun im Wesentlichen dasselbe wie die Debugging-Skripte
+die Sie zuvor zur Implementierung des Modells in 🤗 Transformers verwendet haben. Eine Vorlage für diese Modelltests wurde bereits von dem
+Cookiecutter hinzugefügt, die `BrandNewBertModelIntegrationTests` heißt und nur noch von Ihnen ausgefüllt werden muss. Um sicherzustellen, dass diese
+Tests erfolgreich sind, führen Sie
+
+```bash
+RUN_SLOW=1 pytest -sv tests/models/brand_new_bert/test_modeling_brand_new_bert.py::BrandNewBertModelIntegrationTests
+```
+
+<Tip>
+
+Falls Sie Windows verwenden, sollten Sie `RUN_SLOW=1` durch `SET RUN_SLOW=1` ersetzen.
+
+</Tip>
+
+Zweitens sollten alle Funktionen, die speziell für *brand_new_bert* sind, zusätzlich in einem separaten Test getestet werden unter
+`BrandNewBertModelTester`/``BrandNewBertModelTest`. Dieser Teil wird oft vergessen, ist aber in zweierlei Hinsicht äußerst nützlich
+Weise:
+
+- Er hilft dabei, das Wissen, das Sie während der Modellerweiterung erworben haben, an die Community weiterzugeben, indem er zeigt, wie die
+  speziellen Funktionen von *brand_new_bert* funktionieren sollten.
+- Künftige Mitwirkende können Änderungen am Modell schnell testen, indem sie diese speziellen Tests ausführen.
+
+
+**9. Implementieren Sie den Tokenizer**
+
+Als nächstes sollten wir den Tokenizer von *brand_new_bert* hinzufügen. Normalerweise ist der Tokenizer äquivalent oder sehr ähnlich zu einem
+bereits vorhandenen Tokenizer von 🤗 Transformers.
+
+Es ist sehr wichtig, die ursprüngliche Tokenizer-Datei zu finden/extrahieren und es zu schaffen, diese Datei in die 🤗
+Transformers Implementierung des Tokenizers zu laden.
+
+Um sicherzustellen, dass der Tokenizer korrekt funktioniert, empfiehlt es sich, zunächst ein Skript im ursprünglichen Repository zu erstellen
+zu erstellen, das eine Zeichenkette eingibt und die `input_ids` zurückgibt. Es könnte etwa so aussehen (in Pseudocode):
+
+```python
+input_str = "This is a long example input string containing special characters .$?-, numbers 2872 234 12 and words."
+model = BrandNewBertModel.load_pretrained_checkpoint("/path/to/checkpoint/")
+input_ids = model.tokenize(input_str)
+```
+
+Möglicherweise müssen Sie noch einmal einen Blick in das ursprüngliche Repository werfen, um die richtige Tokenizer-Funktion zu finden, oder Sie müssen
+Sie müssen vielleicht sogar Änderungen an Ihrem Klon des Original-Repositorys vornehmen, um nur die `input_ids` auszugeben. Nach dem Schreiben
+ein funktionierendes Tokenisierungsskript geschrieben, das das ursprüngliche Repository verwendet, sollten Sie ein analoges Skript für 🤗 Transformers
+erstellt werden. Es sollte ähnlich wie dieses aussehen:
+
+```python
+from transformers import BrandNewBertTokenizer
+
+input_str = "This is a long example input string containing special characters .$?-, numbers 2872 234 12 and words."
+
+tokenizer = BrandNewBertTokenizer.from_pretrained("/path/to/tokenizer/folder/")
+
+input_ids = tokenizer(input_str).input_ids
+```
+
+Wenn beide `input_ids` die gleichen Werte ergeben, sollte als letzter Schritt auch eine Tokenizer-Testdatei hinzugefügt werden.
+
+Analog zu den Modellierungstestdateien von *brand_new_bert* sollten auch die Tokenisierungs-Testdateien von *brand_new_bert*
+eine Reihe von fest kodierten Integrationstests enthalten.
+
+**10. Führen Sie End-to-End-Integrationstests aus**
+
+Nachdem Sie den Tokenizer hinzugefügt haben, sollten Sie auch ein paar End-to-End-Integrationstests, die sowohl das Modell als auch den
+Tokenizer zu `tests/models/brand_new_bert/test_modeling_brand_new_bert.py` in 🤗 Transformers.
+Ein solcher Test sollte bei einem aussagekräftigen
+Text-zu-Text-Beispiel zeigen, dass die Implementierung von 🤗 Transformers wie erwartet funktioniert. Ein aussagekräftiges Text-zu-Text-Beispiel kann
+z.B. *ein Quell-zu-Ziel-Übersetzungspaar, ein Artikel-zu-Zusammenfassung-Paar, ein Frage-zu-Antwort-Paar, usw... Wenn keiner der
+der portierten Prüfpunkte in einer nachgelagerten Aufgabe feinabgestimmt wurde, genügt es, sich einfach auf die Modelltests zu verlassen. In einem
+letzten Schritt, um sicherzustellen, dass das Modell voll funktionsfähig ist, sollten Sie alle Tests auch auf der GPU durchführen. Es kann
+Es kann vorkommen, dass Sie vergessen haben, einige `.to(self.device)` Anweisungen zu internen Tensoren des Modells hinzuzufügen, was in einem solchen
+Test zu einem Fehler führen würde. Falls Sie keinen Zugang zu einem Grafikprozessor haben, kann das Hugging Face Team diese Tests für Sie durchführen.
+Tests für Sie übernehmen.
+
+**11. Docstring hinzufügen**
+
+Nun sind alle notwendigen Funktionen für *brand_new_bert* hinzugefügt - Sie sind fast fertig! Das Einzige, was Sie noch hinzufügen müssen, ist
+ein schöner Docstring und eine Doku-Seite. Der Cookiecutter sollte eine Vorlagendatei namens
+`docs/source/model_doc/brand_new_bert.md` hinzugefügt haben, die Sie ausfüllen sollten. Die Benutzer Ihres Modells werden in der Regel zuerst einen Blick auf
+diese Seite ansehen, bevor sie Ihr Modell verwenden. Daher muss die Dokumentation verständlich und prägnant sein. Es ist sehr nützlich für
+die Gemeinschaft, einige *Tipps* hinzuzufügen, um zu zeigen, wie das Modell verwendet werden sollte. Zögern Sie nicht, das Hugging Face-Team anzupingen
+bezüglich der Docstrings.
+
+Stellen Sie als nächstes sicher, dass der zu `src/transformers/models/brand_new_bert/modeling_brand_new_bert.py` hinzugefügte docstring
+korrekt ist und alle erforderlichen Eingaben und Ausgaben enthält. Wir haben eine ausführliche Anleitung zum Schreiben von Dokumentationen und unserem Docstring-Format [hier](writing-documentation). Es ist immer gut, sich daran zu erinnern, dass die Dokumentation
+mindestens so sorgfältig behandelt werden sollte wie der Code in 🤗 Transformers, denn die Dokumentation ist in der Regel der erste Kontaktpunkt der
+Berührungspunkt der Community mit dem Modell ist.
+
+**Code refactor**
+
+Großartig, jetzt haben Sie den gesamten erforderlichen Code für *brand_new_bert* hinzugefügt. An diesem Punkt sollten Sie einige mögliche
+falschen Codestil korrigieren, indem Sie ausführen:
+
+```bash
+make style
+```
+
+und überprüfen Sie, ob Ihr Kodierungsstil die Qualitätsprüfung besteht:
+
+```bash
+make quality
+```
+
+Es gibt noch ein paar andere sehr strenge Designtests in 🤗 Transformers, die möglicherweise noch fehlschlagen, was sich in den
+den Tests Ihres Pull Requests. Dies liegt oft an fehlenden Informationen im Docstring oder an einer falschen
+Benennung. Das Hugging Face Team wird Ihnen sicherlich helfen, wenn Sie hier nicht weiterkommen.
+
+Und schließlich ist es immer eine gute Idee, den eigenen Code zu refaktorisieren, nachdem man sichergestellt hat, dass er korrekt funktioniert. Wenn alle
+Tests bestanden haben, ist es nun an der Zeit, den hinzugefügten Code noch einmal durchzugehen und einige Überarbeitungen vorzunehmen.
+
+Sie haben nun den Codierungsteil abgeschlossen, herzlichen Glückwunsch! 🎉 Sie sind großartig! 😎
+
+**12. Laden Sie die Modelle in den Model Hub hoch**
+
+In diesem letzten Teil sollten Sie alle Checkpoints konvertieren und in den Modell-Hub hochladen und eine Modellkarte für jeden
+hochgeladenen Modell-Kontrollpunkt. Sie können sich mit den Hub-Funktionen vertraut machen, indem Sie unsere [Model sharing and uploading Page](model_sharing) lesen. Hier sollten Sie mit dem Hugging Face-Team zusammenarbeiten, um einen passenden Namen für jeden
+Checkpoint festzulegen und die erforderlichen Zugriffsrechte zu erhalten, um das Modell unter der Organisation des Autors *brand_new_bert* hochladen zu können.
+*brand_new_bert*. Die Methode `push_to_hub`, die in allen Modellen in `transformers` vorhanden ist, ist ein schneller und effizienter Weg, Ihren Checkpoint in den Hub zu pushen. Ein kleines Snippet ist unten eingefügt:
+
+```python
+brand_new_bert.push_to_hub("brand_new_bert")
+# Uncomment the following line to push to an organization.
+# brand_new_bert.push_to_hub("<organization>/brand_new_bert")
+```
+
+Es lohnt sich, etwas Zeit darauf zu verwenden, für jeden Kontrollpunkt passende Musterkarten zu erstellen. Die Modellkarten sollten die
+spezifischen Merkmale dieses bestimmten Prüfpunkts hervorheben, * z.B.* auf welchem Datensatz wurde der Prüfpunkt
+vortrainiert/abgestimmt? Für welche nachgelagerte Aufgabe sollte das Modell verwendet werden? Und fügen Sie auch etwas Code bei, wie Sie
+wie das Modell korrekt verwendet wird.
+
+**13. (Optional) Notizbuch hinzufügen**
+
+Es ist sehr hilfreich, ein Notizbuch hinzuzufügen, in dem im Detail gezeigt wird, wie *brand_new_bert* für Schlussfolgerungen verwendet werden kann und/oder
+bei einer nachgelagerten Aufgabe feinabgestimmt wird. Dies ist nicht zwingend erforderlich, um Ihren PR zusammenzuführen, aber sehr nützlich für die Gemeinschaft.
+
+**14. Reichen Sie Ihren fertigen PR ein**
+
+Sie sind jetzt mit der Programmierung fertig und können zum letzten Schritt übergehen, nämlich der Zusammenführung Ihres PR mit main. Normalerweise hat das
+Hugging Face Team Ihnen an diesem Punkt bereits geholfen haben, aber es lohnt sich, sich etwas Zeit zu nehmen, um Ihrem fertigen
+PR eine schöne Beschreibung zu geben und eventuell Kommentare zu Ihrem Code hinzuzufügen, wenn Sie Ihren Gutachter auf bestimmte Designentscheidungen hinweisen wollen.
+Gutachter hinweisen wollen.
+
+### Teilen Sie Ihre Arbeit!!
+
+Jetzt ist es an der Zeit, von der Community Anerkennung für Ihre Arbeit zu bekommen! Die Fertigstellung einer Modellergänzung ist ein wichtiger
+Beitrag zu Transformers und der gesamten NLP-Gemeinschaft. Ihr Code und die portierten vortrainierten Modelle werden sicherlich
+von Hunderten und vielleicht sogar Tausenden von Entwicklern und Forschern genutzt werden. Sie sollten stolz auf Ihre Arbeit sein und Ihre
+Ihre Leistung mit der Gemeinschaft teilen.
+
+**Sie haben ein weiteres Modell erstellt, das für jeden in der Community super einfach zugänglich ist! 🤯**

docs/source/de/add_new_pipeline.md

@@ -0,0 +1,258 @@
+<!--Copyright 2020 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
+
+⚠️ 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.
+
+-->
+
+# Wie erstellt man eine benutzerdefinierte Pipeline?
+
+In dieser Anleitung sehen wir uns an, wie Sie eine benutzerdefinierte Pipeline erstellen und sie auf dem [Hub](hf.co/models) freigeben oder sie der
+🤗 Transformers-Bibliothek hinzufügen.
+
+Zuallererst müssen Sie entscheiden, welche Roheingaben die Pipeline verarbeiten kann. Es kann sich um Strings, rohe Bytes,
+Dictionaries oder was auch immer die wahrscheinlichste gewünschte Eingabe ist. Versuchen Sie, diese Eingaben so rein wie möglich in Python zu halten
+denn das macht die Kompatibilität einfacher (auch mit anderen Sprachen über JSON). Dies werden die Eingaben der
+Pipeline (`Vorverarbeitung`).
+
+Definieren Sie dann die `Outputs`. Dieselbe Richtlinie wie für die Eingänge. Je einfacher, desto besser. Dies werden die Ausgaben der
+Methode `Postprocess`.
+
+Beginnen Sie damit, die Basisklasse `Pipeline` mit den 4 Methoden zu erben, die für die Implementierung von `preprocess` benötigt werden,
+Weiterleiten", "Nachbearbeitung" und "Parameter säubern".
+
+
+```python
+from transformers import Pipeline
+
+
+class MyPipeline(Pipeline):
+    def _sanitize_parameters(self, **kwargs):
+        preprocess_kwargs = {}
+        if "maybe_arg" in kwargs:
+            preprocess_kwargs["maybe_arg"] = kwargs["maybe_arg"]
+        return preprocess_kwargs, {}, {}
+
+    def preprocess(self, inputs, maybe_arg=2):
+        model_input = Tensor(inputs["input_ids"])
+        return {"model_input": model_input}
+
+    def _forward(self, model_inputs):
+        # model_inputs == {"model_input": model_input}
+        outputs = self.model(**model_inputs)
+        # Maybe {"logits": Tensor(...)}
+        return outputs
+
+    def postprocess(self, model_outputs):
+        best_class = model_outputs["logits"].softmax(-1)
+        return best_class
+```
+
+Die Struktur dieser Aufteilung soll eine relativ nahtlose Unterstützung für CPU/GPU ermöglichen und gleichzeitig die Durchführung von
+Vor-/Nachbearbeitung auf der CPU in verschiedenen Threads
+
+Preprocess" nimmt die ursprünglich definierten Eingaben und wandelt sie in etwas um, das in das Modell eingespeist werden kann. Es kann
+mehr Informationen enthalten und ist normalerweise ein `Dict`.
+
+`_forward` ist das Implementierungsdetail und ist nicht dafür gedacht, direkt aufgerufen zu werden. Weiterleiten" ist die bevorzugte
+aufgerufene Methode, da sie Sicherheitsvorkehrungen enthält, die sicherstellen, dass alles auf dem erwarteten Gerät funktioniert. Wenn etwas
+mit einem realen Modell verknüpft ist, gehört es in die Methode `_forward`, alles andere gehört in die Methoden preprocess/postprocess.
+
+Die Methode `Postprocess` nimmt die Ausgabe von `_forward` und verwandelt sie in die endgültige Ausgabe, die zuvor festgelegt wurde.
+zuvor entschieden wurde.
+
+Die Methode `_sanitize_parameters` ermöglicht es dem Benutzer, beliebige Parameter zu übergeben, wann immer er möchte, sei es bei der Initialisierung
+Zeit `pipeline(...., maybe_arg=4)` oder zur Aufrufzeit `pipe = pipeline(...); output = pipe(...., maybe_arg=4)`.
+
+Die Rückgabe von `_sanitize_parameters` sind die 3 Dicts von kwargs, die direkt an `preprocess` übergeben werden,
+`_forward` und `postprocess` übergeben werden. Füllen Sie nichts aus, wenn der Aufrufer keinen zusätzlichen Parameter angegeben hat. Das
+erlaubt es, die Standardargumente in der Funktionsdefinition beizubehalten, was immer "natürlicher" ist.
+
+Ein klassisches Beispiel wäre das Argument `top_k` in der Nachbearbeitung bei Klassifizierungsaufgaben.
+
+```python
+>>> pipe = pipeline("my-new-task")
+>>> pipe("This is a test")
+[{"label": "1-star", "score": 0.8}, {"label": "2-star", "score": 0.1}, {"label": "3-star", "score": 0.05}
+{"label": "4-star", "score": 0.025}, {"label": "5-star", "score": 0.025}]
+
+>>> pipe("This is a test", top_k=2)
+[{"label": "1-star", "score": 0.8}, {"label": "2-star", "score": 0.1}]
+```
+
+In order to achieve that, we'll update our `postprocess` method with a default parameter to `5`. and edit
+`_sanitize_parameters` to allow this new parameter.
+
+
+```python
+def postprocess(self, model_outputs, top_k=5):
+    best_class = model_outputs["logits"].softmax(-1)
+    # Add logic to handle top_k
+    return best_class
+
+
+def _sanitize_parameters(self, **kwargs):
+    preprocess_kwargs = {}
+    if "maybe_arg" in kwargs:
+        preprocess_kwargs["maybe_arg"] = kwargs["maybe_arg"]
+
+    postprocess_kwargs = {}
+    if "top_k" in kwargs:
+        postprocess_kwargs["top_k"] = kwargs["top_k"]
+    return preprocess_kwargs, {}, postprocess_kwargs
+```
+
+Versuchen Sie, die Eingaben/Ausgaben sehr einfach und idealerweise JSON-serialisierbar zu halten, da dies die Verwendung der Pipeline sehr einfach macht
+ohne dass die Benutzer neue Arten von Objekten verstehen müssen. Es ist auch relativ üblich, viele verschiedene Arten von Argumenten zu unterstützen
+von Argumenten zu unterstützen (Audiodateien, die Dateinamen, URLs oder reine Bytes sein können).
+
+
+
+## Hinzufügen zur Liste der unterstützten Aufgaben
+
+Um Ihre `neue Aufgabe` in die Liste der unterstützten Aufgaben aufzunehmen, müssen Sie sie zur `PIPELINE_REGISTRY` hinzufügen:
+
+```python
+from transformers.pipelines import PIPELINE_REGISTRY
+
+PIPELINE_REGISTRY.register_pipeline(
+    "new-task",
+    pipeline_class=MyPipeline,
+    pt_model=AutoModelForSequenceClassification,
+)
+```
+
+Wenn Sie möchten, können Sie ein Standardmodell angeben. In diesem Fall sollte es mit einer bestimmten Revision (die der Name einer Verzweigung oder ein Commit-Hash sein kann, hier haben wir `"abcdef"` genommen) sowie mit dem Typ versehen sein:
+
+```python
+PIPELINE_REGISTRY.register_pipeline(
+    "new-task",
+    pipeline_class=MyPipeline,
+    pt_model=AutoModelForSequenceClassification,
+    default={"pt": ("user/awesome_model", "abcdef")},
+    type="text",  # current support type: text, audio, image, multimodal
+)
+```
+
+## Teilen Sie Ihre Pipeline auf dem Hub
+
+Um Ihre benutzerdefinierte Pipeline auf dem Hub freizugeben, müssen Sie lediglich den benutzerdefinierten Code Ihrer `Pipeline`-Unterklasse in einer
+Python-Datei speichern. Nehmen wir zum Beispiel an, Sie möchten eine benutzerdefinierte Pipeline für die Klassifizierung von Satzpaaren wie folgt verwenden:
+
+```py
+import numpy as np
+
+from transformers import Pipeline
+
+
+def softmax(outputs):
+    maxes = np.max(outputs, axis=-1, keepdims=True)
+    shifted_exp = np.exp(outputs - maxes)
+    return shifted_exp / shifted_exp.sum(axis=-1, keepdims=True)
+
+
+class PairClassificationPipeline(Pipeline):
+    def _sanitize_parameters(self, **kwargs):
+        preprocess_kwargs = {}
+        if "second_text" in kwargs:
+            preprocess_kwargs["second_text"] = kwargs["second_text"]
+        return preprocess_kwargs, {}, {}
+
+    def preprocess(self, text, second_text=None):
+        return self.tokenizer(text, text_pair=second_text, return_tensors=self.framework)
+
+    def _forward(self, model_inputs):
+        return self.model(**model_inputs)
+
+    def postprocess(self, model_outputs):
+        logits = model_outputs.logits[0].numpy()
+        probabilities = softmax(logits)
+
+        best_class = np.argmax(probabilities)
+        label = self.model.config.id2label[best_class]
+        score = probabilities[best_class].item()
+        logits = logits.tolist()
+        return {"label": label, "score": score, "logits": logits}
+```
+
+Die Implementierung ist Framework-unabhängig und funktioniert für PyTorch- und TensorFlow-Modelle. Wenn wir dies in einer Datei
+einer Datei namens `pair_classification.py` gespeichert haben, können wir sie importieren und wie folgt registrieren:
+
+```py
+from pair_classification import PairClassificationPipeline
+from transformers.pipelines import PIPELINE_REGISTRY
+from transformers import AutoModelForSequenceClassification, TFAutoModelForSequenceClassification
+
+PIPELINE_REGISTRY.register_pipeline(
+    "pair-classification",
+    pipeline_class=PairClassificationPipeline,
+    pt_model=AutoModelForSequenceClassification,
+    tf_model=TFAutoModelForSequenceClassification,
+)
+```
+
+Sobald dies geschehen ist, können wir es mit einem vortrainierten Modell verwenden. Zum Beispiel wurde `sgugger/finetuned-bert-mrpc` auf den
+auf den MRPC-Datensatz abgestimmt, der Satzpaare als Paraphrasen oder nicht klassifiziert.
+
+```py
+from transformers import pipeline
+
+classifier = pipeline("pair-classification", model="sgugger/finetuned-bert-mrpc")
+```
+
+Dann können wir sie auf dem Hub mit der Methode `save_pretrained` in einem `Repository` freigeben:
+
+```py
+from huggingface_hub import Repository
+
+repo = Repository("test-dynamic-pipeline", clone_from="{your_username}/test-dynamic-pipeline")
+classifier.save_pretrained("test-dynamic-pipeline")
+repo.push_to_hub()
+```
+
+Dadurch wird die Datei, in der Sie `PairClassificationPipeline` definiert haben, in den Ordner `"test-dynamic-pipeline"` kopiert,
+und speichert das Modell und den Tokenizer der Pipeline, bevor Sie alles in das Repository verschieben
+`{Ihr_Benutzername}/test-dynamic-pipeline`. Danach kann jeder die Pipeline verwenden, solange er die Option
+`trust_remote_code=True` angeben:
+
+```py
+from transformers import pipeline
+
+classifier = pipeline(model="{your_username}/test-dynamic-pipeline", trust_remote_code=True)
+```
+
+## Hinzufügen der Pipeline zu 🤗 Transformers
+
+Wenn Sie Ihre Pipeline zu 🤗 Transformers beitragen möchten, müssen Sie ein neues Modul im Untermodul `pipelines` hinzufügen
+mit dem Code Ihrer Pipeline hinzufügen. Fügen Sie es dann der Liste der in `pipelines/__init__.py` definierten Aufgaben hinzu.
+
+Dann müssen Sie noch Tests hinzufügen. Erstellen Sie eine neue Datei `tests/test_pipelines_MY_PIPELINE.py` mit Beispielen für die anderen Tests.
+
+Die Funktion `run_pipeline_test` ist sehr allgemein gehalten und läuft auf kleinen Zufallsmodellen auf jeder möglichen
+Architektur, wie durch `model_mapping` und `tf_model_mapping` definiert.
+
+Dies ist sehr wichtig, um die zukünftige Kompatibilität zu testen, d.h. wenn jemand ein neues Modell für
+`XXXForQuestionAnswering` hinzufügt, wird der Pipeline-Test versuchen, mit diesem Modell zu arbeiten. Da die Modelle zufällig sind, ist es
+ist es unmöglich, die tatsächlichen Werte zu überprüfen. Deshalb gibt es eine Hilfsfunktion `ANY`, die einfach versucht, die
+Ausgabe der Pipeline TYPE.
+
+Außerdem *müssen* Sie 2 (idealerweise 4) Tests implementieren.
+
+- test_small_model_pt` : Definieren Sie 1 kleines Modell für diese Pipeline (es spielt keine Rolle, ob die Ergebnisse keinen Sinn ergeben)
+  und testen Sie die Ausgaben der Pipeline. Die Ergebnisse sollten die gleichen sein wie bei `test_small_model_tf`.
+- test_small_model_tf : Definieren Sie 1 kleines Modell für diese Pipeline (es spielt keine Rolle, ob die Ergebnisse keinen Sinn ergeben)
+  und testen Sie die Ausgaben der Pipeline. Die Ergebnisse sollten die gleichen sein wie bei `test_small_model_pt`.
+- test_large_model_pt` (`optional`): Testet die Pipeline an einer echten Pipeline, bei der die Ergebnisse
+  Sinn machen. Diese Tests sind langsam und sollten als solche gekennzeichnet werden. Hier geht es darum, die Pipeline zu präsentieren und sicherzustellen
+  sicherzustellen, dass es in zukünftigen Versionen keine Abweichungen gibt.
+- test_large_model_tf` (`optional`): Testet die Pipeline an einer echten Pipeline, bei der die Ergebnisse
+  Sinn machen. Diese Tests sind langsam und sollten als solche gekennzeichnet werden. Hier geht es darum, die Pipeline zu präsentieren und sicherzustellen
+  sicherzustellen, dass es in zukünftigen Versionen keine Abweichungen gibt.

docs/source/de/add_tensorflow_model.md

@@ -0,0 +1,356 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+
+⚠️ 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.
+
+-->
+
+# Wie konvertiert man ein 🤗 Transformers-Modell in TensorFlow?
+
+Die Tatsache, dass mehrere Frameworks für die Verwendung mit 🤗 Transformers zur Verfügung stehen, gibt Ihnen die Flexibilität, deren Stärken beim Entwurf Ihrer Anwendung auszuspielen.
+Ihre Anwendung zu entwerfen, aber das bedeutet auch, dass die Kompatibilität für jedes Modell einzeln hinzugefügt werden muss. Die gute Nachricht ist, dass
+das Hinzufügen von TensorFlow-Kompatibilität zu einem bestehenden Modell einfacher ist als [das Hinzufügen eines neuen Modells von Grund auf](add_new_model)!
+Ob Sie ein tieferes Verständnis für große TensorFlow-Modelle haben möchten, einen wichtigen Open-Source-Beitrag leisten oder
+TensorFlow für das Modell Ihrer Wahl aktivieren wollen, dieser Leitfaden ist für Sie.
+
+Dieser Leitfaden befähigt Sie, ein Mitglied unserer Gemeinschaft, TensorFlow-Modellgewichte und/oder
+Architekturen beizusteuern, die in 🤗 Transformers verwendet werden sollen, und zwar mit minimaler Betreuung durch das Hugging Face Team. Das Schreiben eines neuen Modells
+ist keine Kleinigkeit, aber ich hoffe, dass dieser Leitfaden dazu beiträgt, dass es weniger eine Achterbahnfahrt 🎢 und mehr ein Spaziergang im Park 🚶 ist.
+Die Nutzung unserer kollektiven Erfahrungen ist absolut entscheidend, um diesen Prozess immer einfacher zu machen, und deshalb möchten wir
+ermutigen Sie daher, Verbesserungsvorschläge für diesen Leitfaden zu machen!
+
+Bevor Sie tiefer eintauchen, empfehlen wir Ihnen, die folgenden Ressourcen zu lesen, wenn Sie neu in 🤗 Transformers sind:
+- [Allgemeiner Überblick über 🤗 Transformers](add_new_model#general-overview-of-transformers)
+- [Die TensorFlow-Philosophie von Hugging Face](https://huggingface.co/blog/tensorflow-philosophy)
+
+Im Rest dieses Leitfadens werden Sie lernen, was nötig ist, um eine neue TensorFlow Modellarchitektur hinzuzufügen, die
+Verfahren zur Konvertierung von PyTorch in TensorFlow-Modellgewichte und wie Sie Unstimmigkeiten zwischen ML
+Frameworks. Legen Sie los!
+
+<Tip>
+
+Sind Sie unsicher, ob das Modell, das Sie verwenden möchten, bereits eine entsprechende TensorFlow-Architektur hat?
+
+&nbsp;
+
+Überprüfen Sie das Feld `model_type` in der `config.json` des Modells Ihrer Wahl
+([Beispiel](https://huggingface.co/bert-base-uncased/blob/main/config.json#L14)). Wenn der entsprechende Modellordner in
+🤗 Transformers eine Datei hat, deren Name mit "modeling_tf" beginnt, bedeutet dies, dass es eine entsprechende TensorFlow
+Architektur hat ([Beispiel](https://github.com/huggingface/transformers/tree/main/src/transformers/models/bert)).
+
+</Tip>
+
+
+## Schritt-für-Schritt-Anleitung zum Hinzufügen von TensorFlow-Modellarchitektur-Code
+
+Es gibt viele Möglichkeiten, eine große Modellarchitektur zu entwerfen, und viele Möglichkeiten, diesen Entwurf zu implementieren. Wie auch immer,
+Sie erinnern sich vielleicht an unseren [allgemeinen Überblick über 🤗 Transformers](add_new_model#general-overview-of-transformers)
+wissen, dass wir ein meinungsfreudiger Haufen sind - die Benutzerfreundlichkeit von 🤗 Transformers hängt von konsistenten Designentscheidungen ab. Aus
+Erfahrung können wir Ihnen ein paar wichtige Dinge über das Hinzufügen von TensorFlow-Modellen sagen:
+
+- Erfinden Sie das Rad nicht neu! In den meisten Fällen gibt es mindestens zwei Referenzimplementierungen, die Sie überprüfen sollten: das
+PyTorch-Äquivalent des Modells, das Sie implementieren, und andere TensorFlow-Modelle für dieselbe Klasse von Problemen.
+- Gute Modellimplementierungen überleben den Test der Zeit. Dies geschieht nicht, weil der Code hübsch ist, sondern eher
+sondern weil der Code klar, einfach zu debuggen und darauf aufzubauen ist. Wenn Sie den Maintainern das Leben mit Ihrer
+TensorFlow-Implementierung leicht machen, indem Sie die gleichen Muster wie in anderen TensorFlow-Modellen nachbilden und die Abweichung
+zur PyTorch-Implementierung minimieren, stellen Sie sicher, dass Ihr Beitrag lange Bestand haben wird.
+- Bitten Sie um Hilfe, wenn Sie nicht weiterkommen! Das 🤗 Transformers-Team ist da, um zu helfen, und wir haben wahrscheinlich Lösungen für die gleichen
+Probleme gefunden, vor denen Sie stehen.
+
+Hier finden Sie einen Überblick über die Schritte, die zum Hinzufügen einer TensorFlow-Modellarchitektur erforderlich sind:
+1. Wählen Sie das Modell, das Sie konvertieren möchten
+2. Bereiten Sie die Transformers-Entwicklungsumgebung vor.
+3. (Optional) Verstehen Sie die theoretischen Aspekte und die bestehende Implementierung
+4. Implementieren Sie die Modellarchitektur
+5. Implementieren Sie Modelltests
+6. Reichen Sie den Pull-Antrag ein
+7. (Optional) Erstellen Sie Demos und teilen Sie diese mit der Welt
+
+### 1.-3. Bereiten Sie Ihren Modellbeitrag vor
+
+**1. Wählen Sie das Modell, das Sie konvertieren möchten**
+
+Beginnen wir mit den Grundlagen: Als erstes müssen Sie die Architektur kennen, die Sie konvertieren möchten. Wenn Sie
+Sie sich nicht auf eine bestimmte Architektur festgelegt haben, ist es eine gute Möglichkeit, das 🤗 Transformers-Team um Vorschläge zu bitten.
+Wir werden Sie zu den wichtigsten Architekturen führen, die auf der TensorFlow-Seite noch fehlen.
+Seite fehlen. Wenn das spezifische Modell, das Sie mit TensorFlow verwenden möchten, bereits eine Implementierung der TensorFlow-Architektur in
+🤗 Transformers, aber es fehlen Gewichte, können Sie direkt in den
+Abschnitt [Gewichtskonvertierung](#adding-tensorflow-weights-to-hub)
+auf dieser Seite.
+
+Der Einfachheit halber wird im Rest dieser Anleitung davon ausgegangen, dass Sie sich entschieden haben, mit der TensorFlow-Version von
+*BrandNewBert* (dasselbe Beispiel wie in der [Anleitung](add_new_model), um ein neues Modell von Grund auf hinzuzufügen).
+
+<Tip>
+
+Bevor Sie mit der Arbeit an einer TensorFlow-Modellarchitektur beginnen, sollten Sie sich vergewissern, dass es keine laufenden Bemühungen in dieser Richtung gibt.
+Sie können nach `BrandNewBert` auf der
+[pull request GitHub page](https://github.com/huggingface/transformers/pulls?q=is%3Apr), um zu bestätigen, dass es keine
+TensorFlow-bezogene Pull-Anfrage gibt.
+
+</Tip>
+
+
+**2. Transformers-Entwicklungsumgebung vorbereiten**
+
+Nachdem Sie die Modellarchitektur ausgewählt haben, öffnen Sie einen PR-Entwurf, um Ihre Absicht zu signalisieren, daran zu arbeiten. Folgen Sie den
+Anweisungen, um Ihre Umgebung einzurichten und einen PR-Entwurf zu öffnen.
+
+1. Forken Sie das [repository](https://github.com/huggingface/transformers), indem Sie auf der Seite des Repositorys auf die Schaltfläche 'Fork' klicken.
+   Seite des Repositorys klicken. Dadurch wird eine Kopie des Codes unter Ihrem GitHub-Benutzerkonto erstellt.
+
+2. Klonen Sie Ihren `transformers` Fork auf Ihre lokale Festplatte und fügen Sie das Basis-Repository als Remote hinzu:
+
+```bash
+git clone https://github.com/[your Github handle]/transformers.git
+cd transformers
+git remote add upstream https://github.com/huggingface/transformers.git
+```
+
+3. Richten Sie eine Entwicklungsumgebung ein, indem Sie z.B. den folgenden Befehl ausführen:
+
+```bash
+python -m venv .env
+source .env/bin/activate
+pip install -e ".[dev]"
+```
+
+Abhängig von Ihrem Betriebssystem und da die Anzahl der optionalen Abhängigkeiten von Transformers wächst, kann es sein, dass Sie bei diesem Befehl einen
+Fehler mit diesem Befehl erhalten. Wenn das der Fall ist, stellen Sie sicher, dass Sie TensorFlow installieren und dann ausführen:
+
+```bash
+pip install -e ".[quality]"
+```
+
+**Hinweis:** Sie müssen CUDA nicht installiert haben. Es reicht aus, das neue Modell auf der CPU laufen zu lassen.
+
+4. Erstellen Sie eine Verzweigung mit einem beschreibenden Namen von Ihrer Hauptverzweigung
+
+```bash
+git checkout -b add_tf_brand_new_bert
+```
+
+5. Abrufen und zurücksetzen auf die aktuelle Hauptversion
+
+```bash
+git fetch upstream
+git rebase upstream/main
+```
+
+6. Fügen Sie eine leere `.py` Datei in `transformers/src/models/brandnewbert/` mit dem Namen `modeling_tf_brandnewbert.py` hinzu. Dies wird
+Ihre TensorFlow-Modelldatei sein.
+
+7. Übertragen Sie die Änderungen auf Ihr Konto mit:
+
+```bash
+git add .
+git commit -m "initial commit"
+git push -u origin add_tf_brand_new_bert
+```
+
+8. Wenn Sie zufrieden sind, gehen Sie auf die Webseite Ihrer Abspaltung auf GitHub. Klicken Sie auf "Pull request". Stellen Sie sicher, dass Sie das
+   GitHub-Handle einiger Mitglieder des Hugging Face-Teams als Reviewer hinzuzufügen, damit das Hugging Face-Team über zukünftige Änderungen informiert wird.
+   zukünftige Änderungen benachrichtigt wird.
+
+9. Ändern Sie den PR in einen Entwurf, indem Sie auf der rechten Seite der GitHub-Pull-Request-Webseite auf "In Entwurf umwandeln" klicken.
+
+
+Jetzt haben Sie eine Entwicklungsumgebung eingerichtet, um *BrandNewBert* nach TensorFlow in 🤗 Transformers zu portieren.
+
+
+**3. (Optional) Verstehen Sie die theoretischen Aspekte und die bestehende Implementierung**
+
+Sie sollten sich etwas Zeit nehmen, um die Arbeit von *BrandNewBert* zu lesen, falls eine solche Beschreibung existiert. Möglicherweise gibt es große
+Abschnitte des Papiers, die schwer zu verstehen sind. Wenn das der Fall ist, ist das in Ordnung - machen Sie sich keine Sorgen! Das Ziel ist
+ist es nicht, ein tiefes theoretisches Verständnis des Papiers zu erlangen, sondern die notwendigen Informationen zu extrahieren, um
+das Modell mit Hilfe von TensorFlow effektiv in 🤗 Transformers neu zu implementieren. Das heißt, Sie müssen nicht zu viel Zeit auf die
+viel Zeit auf die theoretischen Aspekte verwenden, sondern sich lieber auf die praktischen Aspekte konzentrieren, nämlich auf die bestehende Modelldokumentation
+Seite (z.B. [model docs for BERT](model_doc/bert)).
+
+Nachdem Sie die Grundlagen der Modelle, die Sie implementieren wollen, verstanden haben, ist es wichtig, die bestehende
+Implementierung zu verstehen. Dies ist eine gute Gelegenheit, sich zu vergewissern, dass eine funktionierende Implementierung mit Ihren Erwartungen an das
+Modell entspricht, und um technische Herausforderungen auf der TensorFlow-Seite vorauszusehen.
+
+Es ist ganz natürlich, dass Sie sich von der Menge an Informationen, die Sie gerade aufgesogen haben, überwältigt fühlen. Es ist
+Es ist definitiv nicht erforderlich, dass Sie in dieser Phase alle Facetten des Modells verstehen. Dennoch empfehlen wir Ihnen dringend
+ermutigen wir Sie, alle dringenden Fragen in unserem [Forum](https://discuss.huggingface.co/) zu klären.
+
+
+### 4. Implementierung des Modells
+
+Jetzt ist es an der Zeit, endlich mit dem Programmieren zu beginnen. Als Ausgangspunkt empfehlen wir die PyTorch-Datei selbst: Kopieren Sie den Inhalt von
+modeling_brand_new_bert.py` in `src/transformers/models/brand_new_bert/` nach
+modeling_tf_brand_new_bert.py`. Das Ziel dieses Abschnitts ist es, die Datei zu ändern und die Importstruktur von
+🤗 Transformers zu aktualisieren, so dass Sie `TFBrandNewBert` und
+`TFBrandNewBert.from_pretrained(model_repo, from_pt=True)` erfolgreich ein funktionierendes TensorFlow *BrandNewBert* Modell lädt.
+
+Leider gibt es kein Rezept, um ein PyTorch-Modell in TensorFlow zu konvertieren. Sie können jedoch unsere Auswahl an
+Tipps befolgen, um den Prozess so reibungslos wie möglich zu gestalten:
+- Stellen Sie `TF` dem Namen aller Klassen voran (z.B. wird `BrandNewBert` zu `TFBrandNewBert`).
+- Die meisten PyTorch-Operationen haben einen direkten TensorFlow-Ersatz. Zum Beispiel entspricht `torch.nn.Linear` der Klasse
+  `tf.keras.layers.Dense`, `torch.nn.Dropout` entspricht `tf.keras.layers.Dropout`, usw. Wenn Sie sich nicht sicher sind
+  über eine bestimmte Operation nicht sicher sind, können Sie die [TensorFlow-Dokumentation](https://www.tensorflow.org/api_docs/python/tf)
+  oder die [PyTorch-Dokumentation](https://pytorch.org/docs/stable/).
+- Suchen Sie nach Mustern in der Codebasis von 🤗 Transformers. Wenn Sie auf eine bestimmte Operation stoßen, für die es keinen direkten Ersatz gibt
+   Ersatz hat, stehen die Chancen gut, dass jemand anderes bereits das gleiche Problem hatte.
+- Behalten Sie standardmäßig die gleichen Variablennamen und die gleiche Struktur wie in PyTorch bei. Dies erleichtert die Fehlersuche, die Verfolgung von
+   Probleme zu verfolgen und spätere Korrekturen vorzunehmen.
+- Einige Ebenen haben in jedem Framework unterschiedliche Standardwerte. Ein bemerkenswertes Beispiel ist die Schicht für die Batch-Normalisierung
+   epsilon (`1e-5` in [PyTorch](https://pytorch.org/docs/stable/generated/torch.nn.BatchNorm2d.html#torch.nn.BatchNorm2d)
+   und `1e-3` in [TensorFlow](https://www.tensorflow.org/api_docs/python/tf/keras/layers/BatchNormalization)).
+   Prüfen Sie die Dokumentation genau!
+- Die Variablen `nn.Parameter` von PyTorch müssen in der Regel innerhalb von TF Layer's `build()` initialisiert werden. Siehe das folgende
+   Beispiel: [PyTorch](https://github.com/huggingface/transformers/blob/655f72a6896c0533b1bdee519ed65a059c2425ac/src/transformers/models/vit_mae/modeling_vit_mae.py#L212) /
+   [TensorFlow](https://github.com/huggingface/transformers/blob/655f72a6896c0533b1bdee519ed65a059c2425ac/src/transformers/models/vit_mae/modeling_tf_vit_mae.py#L220)
+- Wenn das PyTorch-Modell ein `#copied from ...` am Anfang einer Funktion hat, stehen die Chancen gut, dass Ihr TensorFlow-Modell diese Funktion auch
+   diese Funktion von der Architektur ausleihen kann, von der sie kopiert wurde, vorausgesetzt, es hat eine TensorFlow-Architektur.
+- Die korrekte Zuweisung des Attributs `name` in TensorFlow-Funktionen ist entscheidend, um das `from_pt=True` Gewicht zu erreichen
+   Cross-Loading. Name" ist fast immer der Name der entsprechenden Variablen im PyTorch-Code. Wenn `name` nicht
+   nicht richtig gesetzt ist, sehen Sie dies in der Fehlermeldung beim Laden der Modellgewichte.
+- Die Logik der Basismodellklasse, `BrandNewBertModel`, befindet sich in `TFBrandNewBertMainLayer`, einer Keras
+   Schicht-Unterklasse ([Beispiel](https://github.com/huggingface/transformers/blob/4fd32a1f499e45f009c2c0dea4d81c321cba7e02/src/transformers/models/bert/modeling_tf_bert.py#L719)).
+   TFBrandNewBertModel" ist lediglich ein Wrapper für diese Schicht.
+- Keras-Modelle müssen erstellt werden, um die vorher trainierten Gewichte zu laden. Aus diesem Grund muss `TFBrandNewBertPreTrainedModel`
+   ein Beispiel für die Eingaben in das Modell enthalten, die `dummy_inputs`
+   ([Beispiel](https://github.com/huggingface/transformers/blob/4fd32a1f499e45f009c2c0dea4d81c321cba7e02/src/transformers/models/bert/modeling_tf_bert.py#L916)).
+- Wenn Sie nicht weiterkommen, fragen Sie nach Hilfe - wir sind für Sie da! 🤗
+
+Neben der Modelldatei selbst müssen Sie auch die Verweise auf die Modellklassen und die zugehörigen
+Dokumentationsseiten hinzufügen. Sie können diesen Teil ganz nach den Mustern in anderen PRs erledigen
+([Beispiel](https://github.com/huggingface/transformers/pull/18020/files)). Hier ist eine Liste der erforderlichen manuellen
+Änderungen:
+- Fügen Sie alle öffentlichen Klassen von *BrandNewBert* in `src/transformers/__init__.py` ein.
+- Fügen Sie *BrandNewBert* Klassen zu den entsprechenden Auto Klassen in `src/transformers/models/auto/modeling_tf_auto.py` hinzu.
+- Fügen Sie die *BrandNewBert* zugehörigen Klassen für träges Laden in `src/transformers/utils/dummy_tf_objects.py` hinzu.
+- Aktualisieren Sie die Importstrukturen für die öffentlichen Klassen in `src/transformers/models/brand_new_bert/__init__.py`.
+- Fügen Sie die Dokumentationszeiger auf die öffentlichen Methoden von *BrandNewBert* in `docs/source/de/model_doc/brand_new_bert.md` hinzu.
+- Fügen Sie sich selbst zur Liste der Mitwirkenden an *BrandNewBert* in `docs/source/de/model_doc/brand_new_bert.md` hinzu.
+- Fügen Sie schließlich ein grünes Häkchen ✅ in der TensorFlow-Spalte von *BrandNewBert* in `docs/source/de/index.md` hinzu.
+
+Wenn Sie mit Ihrer Implementierung zufrieden sind, führen Sie die folgende Checkliste aus, um zu bestätigen, dass Ihre Modellarchitektur
+fertig ist:
+1. Alle Schichten, die sich zur Trainingszeit anders verhalten (z.B. Dropout), werden mit einem `Training` Argument aufgerufen, das
+von den Top-Level-Klassen weitergegeben wird
+2. Sie haben `#copied from ...` verwendet, wann immer es möglich war.
+3. Die Funktion `TFBrandNewBertMainLayer` und alle Klassen, die sie verwenden, haben ihre Funktion `call` mit `@unpack_inputs` dekoriert
+4. TFBrandNewBertMainLayer` ist mit `@keras_serializable` dekoriert
+5. Ein TensorFlow-Modell kann aus PyTorch-Gewichten mit `TFBrandNewBert.from_pretrained(model_repo, from_pt=True)` geladen werden.
+6. Sie können das TensorFlow Modell mit dem erwarteten Eingabeformat aufrufen
+
+
+### 5. Modell-Tests hinzufügen
+
+Hurra, Sie haben ein TensorFlow-Modell implementiert! Jetzt ist es an der Zeit, Tests hinzuzufügen, um sicherzustellen, dass sich Ihr Modell wie erwartet verhält.
+erwartet. Wie im vorigen Abschnitt schlagen wir vor, dass Sie zunächst die Datei `test_modeling_brand_new_bert.py` in
+`tests/models/brand_new_bert/` in die Datei `test_modeling_tf_brand_new_bert.py` zu kopieren und dann die notwendigen
+TensorFlow-Ersetzungen vornehmen. Für den Moment sollten Sie in allen Aufrufen von `.from_pretrained()` das Flag `from_pt=True` verwenden, um die
+die vorhandenen PyTorch-Gewichte zu laden.
+
+Wenn Sie damit fertig sind, kommt der Moment der Wahrheit: Führen Sie die Tests durch! 😬
+
+```bash
+NVIDIA_TF32_OVERRIDE=0 RUN_SLOW=1 RUN_PT_TF_CROSS_TESTS=1 \
+py.test -vv tests/models/brand_new_bert/test_modeling_tf_brand_new_bert.py
+```
+
+Das wahrscheinlichste Ergebnis ist, dass Sie eine Reihe von Fehlern sehen werden. Machen Sie sich keine Sorgen, das ist zu erwarten! Das Debuggen von ML-Modellen ist
+notorisch schwierig, und der Schlüssel zum Erfolg ist Geduld (und `breakpoint()`). Nach unserer Erfahrung sind die schwierigsten
+Probleme aus subtilen Unstimmigkeiten zwischen ML-Frameworks, zu denen wir am Ende dieses Leitfadens ein paar Hinweise geben.
+In anderen Fällen kann es sein, dass ein allgemeiner Test nicht direkt auf Ihr Modell anwendbar ist; in diesem Fall empfehlen wir eine Überschreibung
+auf der Ebene der Modelltestklasse. Zögern Sie nicht, in Ihrem Entwurf einer Pull-Anfrage um Hilfe zu bitten, wenn
+Sie nicht weiterkommen.
+
+Wenn alle Tests erfolgreich waren, können Sie Ihr Modell in die 🤗 Transformers-Bibliothek aufnehmen! 🎉
+
+### 6.-7. Stellen Sie sicher, dass jeder Ihr Modell verwenden kann
+
+**6. Reichen Sie den Pull Request ein**
+
+Sobald Sie mit der Implementierung und den Tests fertig sind, ist es an der Zeit, eine Pull-Anfrage einzureichen. Bevor Sie Ihren Code einreichen,
+führen Sie unser Dienstprogramm zur Codeformatierung, `make fixup` 🪄, aus. Damit werden automatisch alle Formatierungsfehler behoben, die dazu führen würden, dass
+unsere automatischen Prüfungen fehlschlagen würden.
+
+Nun ist es an der Zeit, Ihren Entwurf einer Pull-Anfrage in eine echte Pull-Anfrage umzuwandeln. Klicken Sie dazu auf die Schaltfläche "Bereit für
+Review" und fügen Sie Joao (`@gante`) und Matt (`@Rocketknight1`) als Reviewer hinzu. Eine Modell-Pull-Anfrage benötigt
+mindestens 3 Reviewer, aber sie werden sich darum kümmern, geeignete zusätzliche Reviewer für Ihr Modell zu finden.
+
+Nachdem alle Gutachter mit dem Stand Ihres PR zufrieden sind, entfernen Sie als letzten Aktionspunkt das Flag `from_pt=True` in
+.from_pretrained()-Aufrufen zu entfernen. Da es keine TensorFlow-Gewichte gibt, müssen Sie sie hinzufügen! Lesen Sie den Abschnitt
+unten, um zu erfahren, wie Sie dies tun können.
+
+Wenn schließlich die TensorFlow-Gewichte zusammengeführt werden, Sie mindestens 3 Genehmigungen von Prüfern haben und alle CI-Checks grün sind
+grün sind, überprüfen Sie die Tests ein letztes Mal lokal
+
+```bash
+NVIDIA_TF32_OVERRIDE=0 RUN_SLOW=1 RUN_PT_TF_CROSS_TESTS=1 \
+py.test -vv tests/models/brand_new_bert/test_modeling_tf_brand_new_bert.py
+```
+
+und wir werden Ihren PR zusammenführen! Herzlichen Glückwunsch zu dem Meilenstein 🎉.
+
+**7. (Optional) Erstellen Sie Demos und teilen Sie sie mit der Welt**
+
+Eine der schwierigsten Aufgaben bei Open-Source ist die Entdeckung. Wie können die anderen Benutzer von der Existenz Ihres
+fabelhaften TensorFlow-Beitrags erfahren? Mit der richtigen Kommunikation, natürlich! 📣
+
+Es gibt vor allem zwei Möglichkeiten, Ihr Modell mit der Community zu teilen:
+- Erstellen Sie Demos. Dazu gehören Gradio-Demos, Notebooks und andere unterhaltsame Möglichkeiten, Ihr Modell vorzuführen. Wir raten Ihnen
+   ermutigen Sie, ein Notizbuch zu unseren [community-driven demos](https://huggingface.co/docs/transformers/community) hinzuzufügen.
+- Teilen Sie Geschichten in sozialen Medien wie Twitter und LinkedIn. Sie sollten stolz auf Ihre Arbeit sein und sie mit der
+   Ihre Leistung mit der Community teilen - Ihr Modell kann nun von Tausenden von Ingenieuren und Forschern auf der ganzen Welt genutzt werden
+   der Welt genutzt werden 🌍! Wir werden Ihre Beiträge gerne retweeten und Ihnen helfen, Ihre Arbeit mit der Community zu teilen.
+
+
+## Hinzufügen von TensorFlow-Gewichten zum 🤗 Hub
+
+Unter der Annahme, dass die TensorFlow-Modellarchitektur in 🤗 Transformers verfügbar ist, ist die Umwandlung von PyTorch-Gewichten in
+TensorFlow-Gewichte ist ein Kinderspiel!
+
+Hier sehen Sie, wie es geht:
+1. Stellen Sie sicher, dass Sie in Ihrem Terminal bei Ihrem Hugging Face Konto angemeldet sind. Sie können sich mit dem folgenden Befehl anmelden
+   `huggingface-cli login` (Ihre Zugangstoken finden Sie [hier](https://huggingface.co/settings/tokens))
+2. Führen Sie `transformers-cli pt-to-tf --model-name foo/bar` aus, wobei `foo/bar` der Name des Modell-Repositorys ist
+   ist, das die PyTorch-Gewichte enthält, die Sie konvertieren möchten.
+3. Markieren Sie `@joaogante` und `@Rocketknight1` in dem 🤗 Hub PR, den der obige Befehl gerade erstellt hat
+
+Das war's! 🎉
+
+
+## Fehlersuche in verschiedenen ML-Frameworks 🐛
+
+Irgendwann, wenn Sie eine neue Architektur hinzufügen oder TensorFlow-Gewichte für eine bestehende Architektur erstellen, werden Sie
+stoßen Sie vielleicht auf Fehler, die sich über Unstimmigkeiten zwischen PyTorch und TensorFlow beschweren. Sie könnten sich sogar dazu entschließen, den
+Modellarchitektur-Code für die beiden Frameworks zu öffnen, und stellen fest, dass sie identisch aussehen. Was ist denn da los? 🤔
+
+Lassen Sie uns zunächst darüber sprechen, warum es wichtig ist, diese Diskrepanzen zu verstehen. Viele Community-Mitglieder werden 🤗
+Transformers-Modelle und vertrauen darauf, dass sich unsere Modelle wie erwartet verhalten. Wenn es eine große Diskrepanz gibt
+zwischen den beiden Frameworks auftritt, bedeutet dies, dass das Modell nicht der Referenzimplementierung für mindestens eines der Frameworks folgt.
+der Frameworks folgt. Dies kann zu stillen Fehlern führen, bei denen das Modell zwar läuft, aber eine schlechte Leistung aufweist. Dies ist
+wohl schlimmer als ein Modell, das überhaupt nicht läuft! Aus diesem Grund streben wir an, dass die Abweichung zwischen den Frameworks kleiner als
+1e-5" in allen Phasen des Modells.
+
+Wie bei anderen numerischen Problemen auch, steckt der Teufel im Detail. Und wie bei jedem detailorientierten Handwerk ist die geheime
+Zutat hier Geduld. Hier ist unser Vorschlag für den Arbeitsablauf, wenn Sie auf diese Art von Problemen stoßen:
+1. Lokalisieren Sie die Quelle der Abweichungen. Das Modell, das Sie konvertieren, hat wahrscheinlich bis zu einem gewissen Punkt nahezu identische innere Variablen.
+   bestimmten Punkt. Platzieren Sie `Breakpoint()`-Anweisungen in den Architekturen der beiden Frameworks und vergleichen Sie die Werte der
+   numerischen Variablen von oben nach unten, bis Sie die Quelle der Probleme gefunden haben.
+2. Nachdem Sie nun die Ursache des Problems gefunden haben, setzen Sie sich mit dem 🤗 Transformers-Team in Verbindung. Es ist möglich
+   dass wir ein ähnliches Problem schon einmal gesehen haben und umgehend eine Lösung anbieten können. Als Ausweichmöglichkeit können Sie beliebte Seiten
+   wie StackOverflow und GitHub-Probleme.
+3. Wenn keine Lösung in Sicht ist, bedeutet das, dass Sie tiefer gehen müssen. Die gute Nachricht ist, dass Sie das Problem gefunden haben.
+   Problem ausfindig gemacht haben, so dass Sie sich auf die problematische Anweisung konzentrieren und den Rest des Modells ausblenden können! Die schlechte Nachricht ist
+   dass Sie sich in die Quellimplementierung der besagten Anweisung einarbeiten müssen. In manchen Fällen finden Sie vielleicht ein
+   Problem mit einer Referenzimplementierung - verzichten Sie nicht darauf, ein Problem im Upstream-Repository zu öffnen.
+
+In einigen Fällen können wir nach Rücksprache mit dem 🤗 Transformers-Team zu dem Schluss kommen, dass die Behebung der Abweichung nicht machbar ist.
+Wenn die Abweichung in den Ausgabeschichten des Modells sehr klein ist (aber möglicherweise groß in den versteckten Zuständen), können wir
+könnten wir beschließen, sie zu ignorieren und das Modell zu verteilen. Die oben erwähnte CLI `pt-to-tf` hat ein `--max-error`
+Flag, um die Fehlermeldung bei der Gewichtskonvertierung zu überschreiben.

docs/source/de/llm_tutorial.md

@@ -0,0 +1,221 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+
+# Generation with LLMs
+
+[[open-in-colab]]
+
+LLMs (Large Language Models) sind die Schlüsselkomponente bei der Texterstellung. Kurz gesagt, bestehen sie aus großen, vortrainierten Transformationsmodellen, die darauf trainiert sind, das nächste Wort (oder genauer gesagt Token) aus einem Eingabetext vorherzusagen. Da sie jeweils ein Token vorhersagen, müssen Sie etwas Aufwändigeres tun, um neue Sätze zu generieren, als nur das Modell aufzurufen - Sie müssen eine autoregressive Generierung durchführen.
+
+Die autoregressive Generierung ist ein Verfahren zur Inferenzzeit, bei dem ein Modell mit seinen eigenen generierten Ausgaben iterativ aufgerufen wird, wenn einige anfängliche Eingaben vorliegen. In 🤗 Transformers wird dies von der Methode [`~generation.GenerationMixin.generate`] übernommen, die allen Modellen mit generativen Fähigkeiten zur Verfügung steht.
+
+Dieses Tutorial zeigt Ihnen, wie Sie:
+
+* Text mit einem LLM generieren
+* Vermeiden Sie häufige Fallstricke
+* Nächste Schritte, damit Sie das Beste aus Ihrem LLM herausholen können
+
+Bevor Sie beginnen, stellen Sie sicher, dass Sie alle erforderlichen Bibliotheken installiert haben:
+
+```bash
+pip install transformers bitsandbytes>=0.39.0 -q
+```
+
+
+## Text generieren
+
+Ein Sprachmodell, das für [causal language modeling](tasks/language_modeling) trainiert wurde, nimmt eine Folge von Text-Token als Eingabe und gibt die Wahrscheinlichkeitsverteilung für das nächste Token zurück.
+
+<!-- [GIF 1 -- FWD PASS] -->
+<figure class="image table text-center m-0 w-full">
+    <video
+        style="max-width: 90%; margin: auto;"
+        autoplay loop muted playsinline
+        src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/assisted-generation/gif_1_1080p.mov"
+    ></video>
+    <figcaption>"Forward pass of an LLM"</figcaption>
+</figure>
+
+Ein wichtiger Aspekt der autoregressiven Generierung mit LLMs ist die Auswahl des nächsten Tokens aus dieser Wahrscheinlichkeitsverteilung. In diesem Schritt ist alles möglich, solange Sie am Ende ein Token für die nächste Iteration haben. Das heißt, es kann so einfach sein wie die Auswahl des wahrscheinlichsten Tokens aus der Wahrscheinlichkeitsverteilung oder so komplex wie die Anwendung von einem Dutzend Transformationen vor der Stichprobenziehung aus der resultierenden Verteilung.
+
+<!-- [GIF 2 -- TEXT GENERATION] -->
+<figure class="image table text-center m-0 w-full">
+    <video
+        style="max-width: 90%; margin: auto;"
+        autoplay loop muted playsinline
+        src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/assisted-generation/gif_2_1080p.mov"
+    ></video>
+    <figcaption>"Die autoregressive Generierung wählt iterativ das nächste Token aus einer Wahrscheinlichkeitsverteilung aus, um Text zu erzeugen"</figcaption>
+</figure>
+
+Der oben dargestellte Prozess wird iterativ wiederholt, bis eine bestimmte Abbruchbedingung erreicht ist. Im Idealfall wird die Abbruchbedingung vom Modell vorgegeben, das lernen sollte, wann es ein Ende-der-Sequenz-Token (EOS) ausgeben muss. Ist dies nicht der Fall, stoppt die Generierung, wenn eine vordefinierte Maximallänge erreicht ist.
+
+Damit sich Ihr Modell so verhält, wie Sie es für Ihre Aufgabe erwarten, müssen Sie den Schritt der Token-Auswahl und die Abbruchbedingung richtig einstellen. Aus diesem Grund haben wir zu jedem Modell eine [`~generation.GenerationConfig`]-Datei, die eine gute generative Standardparametrisierung enthält und zusammen mit Ihrem Modell geladen wird.
+
+Lassen Sie uns über Code sprechen!
+
+<Tip>
+
+Wenn Sie an der grundlegenden Verwendung von LLMs interessiert sind, ist unsere High-Level-Schnittstelle [`Pipeline`](pipeline_tutorial) ein guter Ausgangspunkt. LLMs erfordern jedoch oft fortgeschrittene Funktionen wie Quantisierung und Feinsteuerung des Token-Auswahlschritts, was am besten über [`~generation.GenerationMixin.generate`] erfolgt. Die autoregressive Generierung mit LLMs ist ebenfalls ressourcenintensiv und sollte für einen angemessenen Durchsatz auf einer GPU ausgeführt werden.
+
+</Tip>
+
+<!-- TODO: update example to llama 2 (or a newer popular baseline) when it becomes ungated -->
+Zunächst müssen Sie das Modell laden.
+
+```py
+>>> from transformers import AutoModelForCausalLM
+
+>>> model = AutoModelForCausalLM.from_pretrained(
+...     "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
+... )
+```
+
+Sie werden zwei Flags in dem Aufruf `from_pretrained` bemerken:
+
+ - `device_map` stellt sicher, dass das Modell auf Ihre GPU(s) übertragen wird
+ - `load_in_4bit` wendet [dynamische 4-Bit-Quantisierung](main_classes/quantization) an, um die Ressourcenanforderungen massiv zu reduzieren
+
+Es gibt noch andere Möglichkeiten, ein Modell zu initialisieren, aber dies ist eine gute Grundlage, um mit einem LLM zu beginnen.
+
+Als nächstes müssen Sie Ihre Texteingabe mit einem [tokenizer](tokenizer_summary) vorverarbeiten.
+
+```py
+>>> from transformers import AutoTokenizer
+
+>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b")
+>>> model_inputs = tokenizer(["A list of colors: red, blue"], return_tensors="pt").to("cuda")
+```
+
+Die Variable `model_inputs` enthält die tokenisierte Texteingabe sowie die Aufmerksamkeitsmaske. Obwohl [`~generation.GenerationMixin.generate`] sein Bestes tut, um die Aufmerksamkeitsmaske abzuleiten, wenn sie nicht übergeben wird, empfehlen wir, sie für optimale Ergebnisse wann immer möglich zu übergeben.
+
+Rufen Sie schließlich die Methode [~generation.GenerationMixin.generate] auf, um die generierten Token zurückzugeben, die vor dem Drucken in Text umgewandelt werden sollten.
+
+```py
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'A list of colors: red, blue, green, yellow, black, white, and brown'
+```
+
+Und das war's! Mit ein paar Zeilen Code können Sie sich die Macht eines LLM zunutze machen.
+
+
+## Häufige Fallstricke
+
+Es gibt viele [Generierungsstrategien](generation_strategies), und manchmal sind die Standardwerte für Ihren Anwendungsfall vielleicht nicht geeignet. Wenn Ihre Ausgaben nicht mit dem übereinstimmen, was Sie erwarten, haben wir eine Liste der häufigsten Fallstricke erstellt und wie Sie diese vermeiden können.
+
+```py
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
+
+>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b")
+>>> tokenizer.pad_token = tokenizer.eos_token  # Llama has no pad token by default
+>>> model = AutoModelForCausalLM.from_pretrained(
+...     "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
+... )
+```
+
+### Generierte Ausgabe ist zu kurz/lang
+
+Wenn in der Datei [~generation.GenerationConfig`] nichts angegeben ist, gibt `generate` standardmäßig bis zu 20 Token zurück. Wir empfehlen dringend, `max_new_tokens` in Ihrem `generate`-Aufruf manuell zu setzen, um die maximale Anzahl neuer Token zu kontrollieren, die zurückgegeben werden können. Beachten Sie, dass LLMs (genauer gesagt, [decoder-only models](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt)) auch die Eingabeaufforderung als Teil der Ausgabe zurückgeben.
+
+
+```py
+>>> model_inputs = tokenizer(["A sequence of numbers: 1, 2"], return_tensors="pt").to("cuda")
+
+>>> # By default, the output will contain up to 20 tokens
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'A sequence of numbers: 1, 2, 3, 4, 5'
+
+>>> # Setting `max_new_tokens` allows you to control the maximum length
+>>> generated_ids = model.generate(**model_inputs, max_new_tokens=50)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'A sequence of numbers: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,'
+```
+
+### Falscher Generierungsmodus
+
+Standardmäßig und sofern nicht in der Datei [~generation.GenerationConfig`] angegeben, wählt `generate` bei jeder Iteration das wahrscheinlichste Token aus (gierige Dekodierung). Je nach Aufgabe kann dies unerwünscht sein; kreative Aufgaben wie Chatbots oder das Schreiben eines Aufsatzes profitieren vom Sampling. Andererseits profitieren Aufgaben, bei denen es auf die Eingabe ankommt, wie z.B. Audiotranskription oder Übersetzung, von der gierigen Dekodierung. Aktivieren Sie das Sampling mit `do_sample=True`. Mehr zu diesem Thema erfahren Sie in diesem [Blogbeitrag] (https://huggingface.co/blog/how-to-generate).
+
+```py
+>>> # Set seed or reproducibility -- you don't need this unless you want full reproducibility
+>>> from transformers import set_seed
+>>> set_seed(0)
+
+>>> model_inputs = tokenizer(["I am a cat."], return_tensors="pt").to("cuda")
+
+>>> # LLM + greedy decoding = repetitive, boring output
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'I am a cat. I am a cat. I am a cat. I am a cat'
+
+>>> # With sampling, the output becomes more creative!
+>>> generated_ids = model.generate(**model_inputs, do_sample=True)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'I am a cat.\nI just need to be. I am always.\nEvery time'
+```
+
+### Falsche Auffüllseite
+
+LLMs sind [decoder-only](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt)-Architekturen, d.h. sie iterieren weiter über Ihre Eingabeaufforderung. Wenn Ihre Eingaben nicht die gleiche Länge haben, müssen sie aufgefüllt werden. Da LLMs nicht darauf trainiert sind, mit aufgefüllten Token fortzufahren, muss Ihre Eingabe links aufgefüllt werden. Vergessen Sie auch nicht, die Aufmerksamkeitsmaske an generate zu übergeben!
+
+```py
+>>> # The tokenizer initialized above has right-padding active by default: the 1st sequence,
+>>> # which is shorter, has padding on the right side. Generation fails.
+>>> model_inputs = tokenizer(
+...     ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
+... ).to("cuda")
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids[0], skip_special_tokens=True)[0]
+''
+
+>>> # With left-padding, it works as expected!
+>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b", padding_side="left")
+>>> tokenizer.pad_token = tokenizer.eos_token  # Llama has no pad token by default
+>>> model_inputs = tokenizer(
+...     ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
+... ).to("cuda")
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'1, 2, 3, 4, 5, 6,'
+```
+
+<!-- TODO: when the prompting guide is ready, mention the importance of setting the right prompt in this section -->
+
+## Weitere Ressourcen
+
+Während der Prozess der autoregressiven Generierung relativ einfach ist, kann die optimale Nutzung Ihres LLM ein schwieriges Unterfangen sein, da es viele bewegliche Teile gibt. Für Ihre nächsten Schritte, die Ihnen helfen, tiefer in die LLM-Nutzung und das Verständnis einzutauchen:
+
+<!-- TODO: mit neuen Anleitungen vervollständigen -->
+### Fortgeschrittene Nutzung generieren
+
+1. [Leitfaden](generation_strategies) zur Steuerung verschiedener Generierungsmethoden, zur Einrichtung der Generierungskonfigurationsdatei und zum Streaming der Ausgabe;
+2. API-Referenz zu [`~generation.GenerationConfig`], [`~generation.GenerationMixin.generate`] und [generate-bezogene Klassen](internal/generation_utils).
+
+### LLM-Ranglisten
+
+1. [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard), das sich auf die Qualität der Open-Source-Modelle konzentriert;
+2. [Open LLM-Perf Leaderboard](https://huggingface.co/spaces/optimum/llm-perf-leaderboard), das sich auf den LLM-Durchsatz konzentriert.
+
+### Latenz und Durchsatz
+
+1. [Leitfaden](main_classes/quantization) zur dynamischen Quantisierung, der Ihnen zeigt, wie Sie Ihren Speicherbedarf drastisch reduzieren können.
+
+### Verwandte Bibliotheken
+
+1. [text-generation-inference](https://github.com/huggingface/text-generation-inference), ein produktionsreifer Server für LLMs;
+2. [`optimum`](https://github.com/huggingface/optimum), eine Erweiterung von 🤗 Transformers, die für bestimmte Hardware-Geräte optimiert.

docs/source/de/peft.md

@@ -0,0 +1,216 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# Adapter mit 🤗 PEFT laden
+
+[[open-in-colab]]
+
+Die [Parameter-Efficient Fine Tuning (PEFT)](https://huggingface.co/blog/peft) Methoden frieren die vorab trainierten Modellparameter während der Feinabstimmung ein und fügen eine kleine Anzahl trainierbarer Parameter (die Adapter) hinzu. Die Adapter werden trainiert, um aufgabenspezifische Informationen zu lernen. Es hat sich gezeigt, dass dieser Ansatz sehr speichereffizient ist und weniger Rechenleistung beansprucht, während die Ergebnisse mit denen eines vollständig feinabgestimmten Modells vergleichbar sind. 
+
+Adapter, die mit PEFT trainiert wurden, sind in der Regel um eine Größenordnung kleiner als das vollständige Modell, so dass sie bequem gemeinsam genutzt, gespeichert und geladen werden können.
+
+<div class="flex flex-col justify-center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/PEFT-hub-screenshot.png"/>
+  <figcaption class="text-center">Die Adaptergewichte für ein OPTForCausalLM-Modell, die auf dem Hub gespeichert sind, sind nur ~6MB groß, verglichen mit der vollen Größe der Modellgewichte, die ~700MB betragen können.</figcaption>
+</div>
+
+Wenn Sie mehr über die 🤗 PEFT-Bibliothek erfahren möchten, sehen Sie sich die [Dokumentation](https://huggingface.co/docs/peft/index) an.
+
+## Setup
+
+Starten Sie mit der Installation von 🤗 PEFT:
+
+```bash
+pip install peft
+```
+
+Wenn Sie die brandneuen Funktionen ausprobieren möchten, sollten Sie die Bibliothek aus dem Quellcode installieren:
+
+```bash
+pip install git+https://github.com/huggingface/peft.git
+```
+
+## Unterstützte PEFT-Modelle
+
+Transformers unterstützt nativ einige PEFT-Methoden, d.h. Sie können lokal oder auf dem Hub gespeicherte Adaptergewichte laden und sie mit wenigen Zeilen Code einfach ausführen oder trainieren. Die folgenden Methoden werden unterstützt:
+
+- [Low Rank Adapters](https://huggingface.co/docs/peft/conceptual_guides/lora)
+- [IA3](https://huggingface.co/docs/peft/conceptual_guides/ia3)
+- [AdaLoRA](https://arxiv.org/abs/2303.10512)
+
+Wenn Sie andere PEFT-Methoden, wie z.B. Prompt Learning oder Prompt Tuning, verwenden möchten, oder über die 🤗 PEFT-Bibliothek im Allgemeinen, lesen Sie bitte die [Dokumentation](https://huggingface.co/docs/peft/index).
+
+
+## Laden Sie einen PEFT-Adapter
+
+Um ein PEFT-Adaptermodell von 🤗 Transformers zu laden und zu verwenden, stellen Sie sicher, dass das Hub-Repository oder das lokale Verzeichnis eine `adapter_config.json`-Datei und die Adaptergewichte enthält, wie im obigen Beispielbild gezeigt. Dann können Sie das PEFT-Adaptermodell mit der Klasse `AutoModelFor` laden. Um zum Beispiel ein PEFT-Adaptermodell für die kausale Sprachmodellierung zu laden:
+
+1. Geben Sie die PEFT-Modell-ID an.
+2. übergeben Sie es an die Klasse [`AutoModelForCausalLM`].
+
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+peft_model_id = "ybelkada/opt-350m-lora"
+model = AutoModelForCausalLM.from_pretrained(peft_model_id)
+```
+
+<Tip>
+
+Sie können einen PEFT-Adapter entweder mit einer `AutoModelFor`-Klasse oder der Basismodellklasse wie `OPTForCausalLM` oder `LlamaForCausalLM` laden.
+
+</Tip>
+
+Sie können einen PEFT-Adapter auch laden, indem Sie die Methode `load_adapter` aufrufen:
+
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_id = "facebook/opt-350m"
+peft_model_id = "ybelkada/opt-350m-lora"
+
+model = AutoModelForCausalLM.from_pretrained(model_id)
+model.load_adapter(peft_model_id)
+```
+
+## Laden in 8bit oder 4bit
+
+Die `bitsandbytes`-Integration unterstützt Datentypen mit 8bit und 4bit Genauigkeit, was für das Laden großer Modelle nützlich ist, weil es Speicher spart (lesen Sie den `bitsandbytes`-Integrations [guide](./quantization#bitsandbytes-integration), um mehr zu erfahren). Fügen Sie die Parameter `load_in_8bit` oder `load_in_4bit` zu [`~PreTrainedModel.from_pretrained`] hinzu und setzen Sie `device_map="auto"`, um das Modell effektiv auf Ihre Hardware zu verteilen:
+
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+peft_model_id = "ybelkada/opt-350m-lora"
+model = AutoModelForCausalLM.from_pretrained(peft_model_id, device_map="auto", load_in_8bit=True)
+```
+
+## Einen neuen Adapter hinzufügen
+
+Sie können [`~peft.PeftModel.add_adapter`] verwenden, um einen neuen Adapter zu einem Modell mit einem bestehenden Adapter hinzuzufügen, solange der neue Adapter vom gleichen Typ ist wie der aktuelle Adapter. Wenn Sie zum Beispiel einen bestehenden LoRA-Adapter an ein Modell angehängt haben:
+
+```py
+from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
+from peft import PeftConfig
+
+model_id = "facebook/opt-350m"
+model = AutoModelForCausalLM.from_pretrained(model_id)
+
+lora_config = LoraConfig(
+    target_modules=["q_proj", "k_proj"],
+    init_lora_weights=False
+)
+
+model.add_adapter(lora_config, adapter_name="adapter_1")
+```
+
+Um einen neuen Adapter hinzuzufügen:
+
+```py
+# attach new adapter with same config
+model.add_adapter(lora_config, adapter_name="adapter_2")
+```
+
+Jetzt können Sie mit [`~peft.PeftModel.set_adapter`] festlegen, welcher Adapter verwendet werden soll:
+
+```py
+# use adapter_1
+model.set_adapter("adapter_1")
+output = model.generate(**inputs)
+print(tokenizer.decode(output_disabled[0], skip_special_tokens=True))
+
+# use adapter_2
+model.set_adapter("adapter_2")
+output_enabled = model.generate(**inputs)
+print(tokenizer.decode(output_enabled[0], skip_special_tokens=True))
+```
+
+## Aktivieren und Deaktivieren von Adaptern
+
+Sobald Sie einen Adapter zu einem Modell hinzugefügt haben, können Sie das Adaptermodul aktivieren oder deaktivieren. So aktivieren Sie das Adaptermodul:
+
+```py
+from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
+from peft import PeftConfig
+
+model_id = "facebook/opt-350m"
+adapter_model_id = "ybelkada/opt-350m-lora"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+text = "Hello"
+inputs = tokenizer(text, return_tensors="pt")
+
+model = AutoModelForCausalLM.from_pretrained(model_id)
+peft_config = PeftConfig.from_pretrained(adapter_model_id)
+
+# to initiate with random weights
+peft_config.init_lora_weights = False
+
+model.add_adapter(peft_config)
+model.enable_adapters()
+output = model.generate(**inputs)
+```
+
+So deaktivieren Sie das Adaptermodul:
+
+```py
+model.disable_adapters()
+output = model.generate(**inputs)
+```
+
+## PEFT-Adapter trainieren
+
+PEFT-Adapter werden von der Klasse [`Trainer`] unterstützt, so dass Sie einen Adapter für Ihren speziellen Anwendungsfall trainieren können. Dazu müssen Sie nur ein paar weitere Codezeilen hinzufügen. Zum Beispiel, um einen LoRA-Adapter zu trainieren:
+
+<Tip>
+
+Wenn Sie mit der Feinabstimmung eines Modells mit [`Trainer`] noch nicht vertraut sind, werfen Sie einen Blick auf das Tutorial [Feinabstimmung eines vortrainierten Modells](Training).
+
+</Tip>
+
+1. Definieren Sie Ihre Adapterkonfiguration mit dem Aufgabentyp und den Hyperparametern (siehe [`~peft.LoraConfig`] für weitere Details darüber, was die Hyperparameter tun).
+
+```py
+from peft import LoraConfig
+
+peft_config = LoraConfig(
+    lora_alpha=16,
+    lora_dropout=0.1,
+    r=64,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+```
+
+2. Fügen Sie dem Modell einen Adapter hinzu.
+
+```py
+model.add_adapter(peft_config)
+```
+
+3. Jetzt können Sie das Modell an [`Trainer`] übergeben!
+
+```py
+trainer = Trainer(model=model, ...)
+trainer.train()
+```
+
+So speichern Sie Ihren trainierten Adapter und laden ihn wieder:
+
+```py
+model.save_pretrained(save_dir)
+model = AutoModelForCausalLM.from_pretrained(save_dir)
+```
+
+<!--
+TODO: (@younesbelkada @stevhliu)
+-   Link to PEFT docs for further details
+-   Trainer  
+-   8-bit / 4-bit examples ?
+-->

docs/source/de/pr_checks.md

@@ -0,0 +1,199 @@
+<!---
+Copyright 2020 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.
+
+-->
+
+# Überprüfungen bei einer Pull-Anfrage
+
+Wenn Sie eine Pull-Anfrage für 🤗 Transformers öffnen, wird eine ganze Reihe von Prüfungen durchgeführt, um sicherzustellen, dass der Patch, den Sie hinzufügen, nichts Bestehendes zerstört. Es gibt vier Arten von Prüfungen:
+- reguläre Tests
+- Erstellung der Dokumentation
+- Stil von Code und Dokumentation
+- allgemeine Konsistenz des Repository
+
+In diesem Dokument werden wir versuchen zu erklären, worum es sich bei diesen verschiedenen Prüfungen handelt und wie Sie sie lokal debuggen können, wenn eine der Prüfungen in Ihrer PR fehlschlägt.
+
+Beachten Sie, dass Sie im Idealfall eine Dev-Installation benötigen:
+
+```bash
+pip install transformers[dev]
+```
+
+oder für eine bearbeitbare Installation:
+
+```bash
+pip install -e .[dev]
+```
+
+innerhalb des Transformers Repo. Da die Anzahl der optionalen Abhängigkeiten von Transformers stark zugenommen hat, ist es möglich, dass Sie nicht alle davon bekommen können. Wenn die Dev-Installation fehlschlägt, stellen Sie sicher, dass Sie das Deep Learning-Framework, mit dem Sie arbeiten, installieren (PyTorch, TensorFlow und/oder Flax).
+
+```bash
+pip install transformers[quality]
+```
+
+oder für eine bearbeitbare Installation:
+
+```bash
+pip install -e .[quality]
+```
+
+
+## Tests
+
+Alle Jobs, die mit `ci/circleci: run_tests_` beginnen, führen Teile der Transformers-Testsuite aus. Jeder dieser Jobs konzentriert sich auf einen Teil der Bibliothek in einer bestimmten Umgebung: `ci/circleci: run_tests_pipelines_tf` zum Beispiel führt den Pipelines-Test in einer Umgebung aus, in der nur TensorFlow installiert ist.
+
+Beachten Sie, dass nur ein Teil der Testsuite jedes Mal ausgeführt wird, um zu vermeiden, dass Tests ausgeführt werden, wenn es keine wirkliche Änderung in den Modulen gibt, die sie testen: ein Dienstprogramm wird ausgeführt, um die Unterschiede in der Bibliothek zwischen vor und nach dem PR zu ermitteln (was GitHub Ihnen auf der Registerkarte "Files changes" anzeigt) und die Tests auszuwählen, die von diesem Unterschied betroffen sind. Dieses Dienstprogramm kann lokal mit ausgeführt werden:
+
+```bash
+python utils/tests_fetcher.py
+```
+
+aus dem Stammverzeichnis des Transformers-Repositoriums. Es wird:
+
+1. Überprüfen Sie für jede Datei im Diff, ob die Änderungen im Code oder nur in Kommentaren oder Docstrings enthalten sind. Nur die Dateien mit echten Codeänderungen werden beibehalten.
+2. Erstellen Sie eine interne Map, die für jede Datei des Quellcodes der Bibliothek alle Dateien angibt, auf die sie rekursiv Einfluss nimmt. Von Modul A wird gesagt, dass es sich auf Modul B auswirkt, wenn Modul B Modul A importiert. Für die rekursive Auswirkung benötigen wir eine Kette von Modulen, die von Modul A zu Modul B führt und in der jedes Modul das vorherige importiert.
+3. Wenden Sie diese Zuordnung auf die in Schritt 1 gesammelten Dateien an. So erhalten wir die Liste der Modelldateien, die von der PR betroffen sind.
+4. Ordnen Sie jede dieser Dateien der/den entsprechenden Testdatei(en) zu und erhalten Sie die Liste der auszuführenden Tests.
+
+Wenn Sie das Skript lokal ausführen, sollten Sie die Ergebnisse von Schritt 1, 3 und 4 ausgegeben bekommen und somit wissen, welche Tests ausgeführt werden. Das Skript erstellt außerdem eine Datei namens `test_list.txt`, die die Liste der auszuführenden Tests enthält, die Sie mit dem folgenden Befehl lokal ausführen können:
+
+```bash
+python -m pytest -n 8 --dist=loadfile -rA -s $(cat test_list.txt)
+```
+
+Für den Fall, dass Ihnen etwas entgangen ist, wird die komplette Testreihe ebenfalls täglich ausgeführt.
+
+## Dokumentation erstellen
+
+Der Job `build_pr_documentation` erstellt und generiert eine Vorschau der Dokumentation, um sicherzustellen, dass alles in Ordnung ist, wenn Ihr PR zusammengeführt wird. Ein Bot fügt einen Link zur Vorschau der Dokumentation zu Ihrem PR hinzu. Alle Änderungen, die Sie an dem PR vornehmen, werden automatisch in der Vorschau aktualisiert. Wenn die Dokumentation nicht erstellt werden kann, klicken Sie auf **Details** neben dem fehlgeschlagenen Auftrag, um zu sehen, wo der Fehler liegt. Oft ist der Fehler so einfach wie eine fehlende Datei im `toctree`.
+
+Wenn Sie daran interessiert sind, die Dokumentation lokal zu erstellen oder in der Vorschau anzusehen, werfen Sie einen Blick in die [`README.md`](https://github.com/huggingface/transformers/tree/main/docs) im Ordner docs.
+
+## Code und Dokumentationsstil
+
+Die Formatierung des Codes erfolgt für alle Quelldateien, die Beispiele und die Tests mit `black` und `ruff`. Wir haben auch ein benutzerdefiniertes Tool, das sich um die Formatierung von docstrings und `rst`-Dateien kümmert (`utils/style_doc.py`), sowie um die Reihenfolge der Lazy-Importe, die in den Transformers `__init__.py`-Dateien durchgeführt werden (`utils/custom_init_isort.py`). All dies können Sie starten, indem Sie Folgendes ausführen
+
+```bash
+make style
+```
+
+Das CI prüft, ob diese innerhalb der Prüfung `ci/circleci: check_code_quality` angewendet wurden. Es führt auch `ruff` aus, das einen grundlegenden Blick auf Ihren Code wirft und sich beschwert, wenn es eine undefinierte Variable findet oder eine, die nicht verwendet wird. Um diese Prüfung lokal auszuführen, verwenden Sie
+
+```bash
+make quality
+```
+
+Dies kann sehr viel Zeit in Anspruch nehmen. Um dasselbe nur für die Dateien zu tun, die Sie im aktuellen Zweig geändert haben, führen Sie
+
+```bash
+make fixup
+```
+
+Dieser letzte Befehl führt auch alle zusätzlichen Prüfungen für die Konsistenz des Repositorys durch. Schauen wir uns diese an.
+
+## Repository-Konsistenz
+
+Dies fasst alle Tests zusammen, die sicherstellen, dass Ihr PR das Repository in einem guten Zustand verlässt. Sie können diese Prüfung lokal durchführen, indem Sie Folgendes ausführen:
+
+```bash
+make repo-consistency
+```
+
+Dies überprüft, ob:
+
+- Alle zum Init hinzugefügten Objekte sind dokumentiert (ausgeführt von `utils/check_repo.py`)
+- Alle `__init__.py`-Dateien haben in ihren beiden Abschnitten den gleichen Inhalt (ausgeführt von `utils/check_inits.py`)
+- Der gesamte Code, der als Kopie eines anderen Moduls identifiziert wurde, stimmt mit dem Original überein (ausgeführt von `utils/check_copies.py`)
+- Alle Konfigurationsklassen haben mindestens einen gültigen Prüfpunkt, der in ihren Dokumentationen erwähnt wird (ausgeführt von `utils/check_config_docstrings.py`)
+- Alle Konfigurationsklassen enthalten nur Attribute, die in den entsprechenden Modellierungsdateien verwendet werden (ausgeführt von `utils/check_config_attributes.py`)
+- Die Übersetzungen der READMEs und der Index des Dokuments haben die gleiche Modellliste wie die Haupt-README (durchgeführt von `utils/check_copies.py`)
+- Die automatisch generierten Tabellen in der Dokumentation sind auf dem neuesten Stand (ausgeführt von `utils/check_table.py`)
+- Die Bibliothek verfügt über alle Objekte, auch wenn nicht alle optionalen Abhängigkeiten installiert sind (ausgeführt von `utils/check_dummies.py`)
+
+Sollte diese Prüfung fehlschlagen, müssen die ersten beiden Punkte manuell korrigiert werden, die letzten vier können automatisch für Sie korrigiert werden, indem Sie den Befehl
+
+```bash
+make fix-copies
+```
+
+Zusätzliche Prüfungen betreffen PRs, die neue Modelle hinzufügen, vor allem, dass:
+
+- Alle hinzugefügten Modelle befinden sich in einer Auto-Zuordnung (durchgeführt von `utils/check_repo.py`)
+<!-- TODO Sylvain, add a check that makes sure the common tests are implemented.-->
+- Alle Modelle werden ordnungsgemäß getestet (ausgeführt von `utils/check_repo.py`)
+
+<!-- TODO Sylvain, add the following
+- All models are added to the main README, inside the main doc
+- All checkpoints used actually exist on the Hub
+
+-->
+
+### Kopien prüfen
+
+Da die Transformers-Bibliothek in Bezug auf den Modellcode sehr eigenwillig ist und jedes Modell vollständig in einer einzigen Datei implementiert sein sollte, ohne sich auf andere Modelle zu stützen, haben wir einen Mechanismus hinzugefügt, der überprüft, ob eine Kopie des Codes einer Ebene eines bestimmten Modells mit dem Original übereinstimmt. Auf diese Weise können wir bei einer Fehlerbehebung alle anderen betroffenen Modelle sehen und entscheiden, ob wir die Änderung weitergeben oder die Kopie zerstören.
+
+<Tip>
+
+Wenn eine Datei eine vollständige Kopie einer anderen Datei ist, sollten Sie sie in der Konstante `FULL_COPIES` von `utils/check_copies.py` registrieren.
+
+</Tip>
+
+Dieser Mechanismus stützt sich auf Kommentare der Form `# Kopiert von xxx`. Das `xxx` sollte den gesamten Pfad zu der Klasse der Funktion enthalten, die darunter kopiert wird. Zum Beispiel ist `RobertaSelfOutput` eine direkte Kopie der Klasse `BertSelfOutput`. Sie können also [hier](https://github.com/huggingface/transformers/blob/2bd7a27a671fd1d98059124024f580f8f5c0f3b5/src/transformers/models/roberta/modeling_roberta.py#L289) sehen, dass sie einen Kommentar hat:
+
+```py
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
+```
+
+Beachten Sie, dass Sie dies nicht auf eine ganze Klasse anwenden, sondern auf die entsprechenden Methoden, von denen kopiert wird. Zum Beispiel [hier](https://github.com/huggingface/transformers/blob/2bd7a27a671fd1d98059124024f580f8f5c0f3b5/src/transformers/models/roberta/modeling_roberta.py#L598) können Sie sehen, wie `RobertaPreTrainedModel._init_weights` von der gleichen Methode in `BertPreTrainedModel` mit dem Kommentar kopiert wird:
+
+```py
+# Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
+```
+
+Manchmal ist die Kopie bis auf die Namen genau gleich: zum Beispiel verwenden wir in `RobertaAttention` `RobertaSelfAttention` anstelle von `BertSelfAttention`, aber ansonsten ist der Code genau derselbe. Aus diesem Grund unterstützt `#Copied from` einfache String-Ersetzungen mit der folgenden Syntax: `Kopiert von xxx mit foo->bar`. Das bedeutet, dass der Code kopiert wird, wobei alle Instanzen von "foo" durch "bar" ersetzt werden. Sie können sehen, wie es [hier](https://github.com/huggingface/transformers/blob/2bd7a27a671fd1d98059124024f580f8f5c0f3b5/src/transformers/models/roberta/modeling_roberta.py#L304C1-L304C86) in `RobertaAttention` mit dem Kommentar verwendet wird:
+
+```py
+# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->Roberta
+```
+
+Beachten Sie, dass um den Pfeil herum keine Leerzeichen stehen sollten (es sei denn, das Leerzeichen ist Teil des zu ersetzenden Musters, natürlich).
+
+Sie können mehrere Muster durch ein Komma getrennt hinzufügen. Zum Beispiel ist hier `CamemberForMaskedLM` eine direkte Kopie von `RobertaForMaskedLM` mit zwei Ersetzungen: `Roberta` zu `Camembert` und `ROBERTA` zu `CAMEMBERT`. Sie können [hier](https://github.com/huggingface/transformers/blob/15082a9dc6950ecae63a0d3e5060b2fc7f15050a/src/transformers/models/camembert/modeling_camembert.py#L929) sehen, wie dies mit dem Kommentar gemacht wird:
+
+```py
+# Copied from transformers.models.roberta.modeling_roberta.RobertaForMaskedLM with Roberta->Camembert, ROBERTA->CAMEMBERT
+```
+
+Wenn die Reihenfolge eine Rolle spielt (weil eine der Ersetzungen mit einer vorherigen in Konflikt geraten könnte), werden die Ersetzungen von links nach rechts ausgeführt.
+
+<Tip>
+
+Wenn die Ersetzungen die Formatierung ändern (wenn Sie z.B. einen kurzen Namen durch einen sehr langen Namen ersetzen), wird die Kopie nach Anwendung des automatischen Formats überprüft.
+
+</Tip>
+
+Eine andere Möglichkeit, wenn es sich bei den Mustern nur um verschiedene Umschreibungen derselben Ersetzung handelt (mit einer groß- und einer kleingeschriebenen Variante), besteht darin, die Option `all-casing` hinzuzufügen. [Hier](https://github.com/huggingface/transformers/blob/15082a9dc6950ecae63a0d3e5060b2fc7f15050a/src/transformers/models/mobilebert/modeling_mobilebert.py#L1237) ist ein Beispiel in `MobileBertForSequenceClassification` mit dem Kommentar:
+
+```py
+# Copied from transformers.models.bert.modeling_bert.BertForSequenceClassification with Bert->MobileBert all-casing
+```
+
+In diesem Fall wird der Code von `BertForSequenceClassification` kopiert, indem er ersetzt wird:
+- `Bert` durch `MobileBert` (zum Beispiel bei der Verwendung von `MobileBertModel` in der Init)
+- `bert` durch `mobilebert` (zum Beispiel bei der Definition von `self.mobilebert`)
+- `BERT` durch `MOBILEBERT` (in der Konstante `MOBILEBERT_INPUTS_DOCSTRING`)

docs/source/de/run_scripts.md

@@ -0,0 +1,351 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Trainieren mit einem Skript
+
+Neben den 🤗 Transformers [notebooks](./noteboks/README) gibt es auch Beispielskripte, die zeigen, wie man ein Modell für eine Aufgabe mit [PyTorch](https://github.com/huggingface/transformers/tree/main/examples/pytorch), [TensorFlow](https://github.com/huggingface/transformers/tree/main/examples/tensorflow) oder [JAX/Flax](https://github.com/huggingface/transformers/tree/main/examples/flax) trainiert.
+
+Sie werden auch Skripte finden, die wir in unseren [Forschungsprojekten](https://github.com/huggingface/transformers/tree/main/examples/research_projects) und [Legacy-Beispielen](https://github.com/huggingface/transformers/tree/main/examples/legacy) verwendet haben und die größtenteils von der Community stammen. Diese Skripte werden nicht aktiv gepflegt und erfordern eine bestimmte Version von 🤗 Transformers, die höchstwahrscheinlich nicht mit der neuesten Version der Bibliothek kompatibel ist.
+
+Es wird nicht erwartet, dass die Beispielskripte bei jedem Problem sofort funktionieren. Möglicherweise müssen Sie das Skript an das Problem anpassen, das Sie zu lösen versuchen. Um Ihnen dabei zu helfen, legen die meisten Skripte vollständig offen, wie die Daten vorverarbeitet werden, so dass Sie sie nach Bedarf für Ihren Anwendungsfall bearbeiten können.
+
+Für jede Funktion, die Sie in einem Beispielskript implementieren möchten, diskutieren Sie bitte im [Forum] (https://discuss.huggingface.co/) oder in einem [issue] (https://github.com/huggingface/transformers/issues), bevor Sie einen Pull Request einreichen. Wir freuen uns zwar über Fehlerkorrekturen, aber es ist unwahrscheinlich, dass wir einen Pull Request zusammenführen, der mehr Funktionalität auf Kosten der Lesbarkeit hinzufügt.
+
+Diese Anleitung zeigt Ihnen, wie Sie ein Beispiel für ein Trainingsskript zur Zusammenfassung in [PyTorch](https://github.com/huggingface/transformers/tree/main/examples/pytorch/summarization) und [TensorFlow](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/summarization) ausführen können. Sofern nicht anders angegeben, sollten alle Beispiele mit beiden Frameworks funktionieren.
+
+## Einrichtung
+
+Um die neueste Version der Beispielskripte erfolgreich auszuführen, **müssen Sie 🤗 Transformers aus dem Quellcode** in einer neuen virtuellen Umgebung installieren:
+
+```bash
+git clone https://github.com/huggingface/transformers
+cd transformers
+pip install .
+```
+
+Für ältere Versionen der Beispielskripte klicken Sie auf die Umschalttaste unten:
+
+<details>
+  <summary>Beispiele für ältere Versionen von 🤗 Transformers</summary>
+	<ul>
+		<li><a href="https://github.com/huggingface/transformers/tree/v4.5.1/examples">v4.5.1</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v4.4.2/examples">v4.4.2</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v4.3.3/examples">v4.3.3</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v4.2.2/examples">v4.2.2</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v4.1.1/examples">v4.1.1</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v4.0.1/examples">v4.0.1</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v3.5.1/examples">v3.5.1</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v3.4.0/examples">v3.4.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v3.3.1/examples">v3.3.1</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v3.2.0/examples">v3.2.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v3.1.0/examples">v3.1.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v3.0.2/examples">v3.0.2</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.11.0/examples">v2.11.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.10.0/examples">v2.10.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.9.1/examples">v2.9.1</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.8.0/examples">v2.8.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.7.0/examples">v2.7.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.6.0/examples">v2.6.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.5.1/examples">v2.5.1</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.4.0/examples">v2.4.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.3.0/examples">v2.3.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.2.0/examples">v2.2.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.1.0/examples">v2.1.1</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v2.0.0/examples">v2.0.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v1.2.0/examples">v1.2.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v1.1.0/examples">v1.1.0</a></li>
+		<li><a href="https://github.com/huggingface/transformers/tree/v1.0.0/examples">v1.0.0</a></li>
+	</ul>
+</details>
+
+Dann stellen Sie Ihren aktuellen Klon von 🤗 Transformers auf eine bestimmte Version um, z.B. v3.5.1:
+
+```bash
+git checkout tags/v3.5.1
+```
+
+Nachdem Sie die richtige Bibliotheksversion eingerichtet haben, navigieren Sie zu dem Beispielordner Ihrer Wahl und installieren die beispielspezifischen Anforderungen:
+
+```bash
+pip install -r requirements.txt
+```
+
+## Ein Skript ausführen
+
+<frameworkcontent>
+<pt>
+Das Beispielskript lädt einen Datensatz aus der 🤗 [Datasets](https://huggingface.co/docs/datasets/) Bibliothek herunter und verarbeitet ihn vor. Dann nimmt das Skript eine Feinabstimmung eines Datensatzes mit dem [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) auf einer Architektur vor, die eine Zusammenfassung unterstützt. Das folgende Beispiel zeigt, wie die Feinabstimmung von [T5-small](https://huggingface.co/t5-small) auf dem Datensatz [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail) durchgeführt wird. Das T5-Modell benötigt aufgrund der Art und Weise, wie es trainiert wurde, ein zusätzliches Argument `source_prefix`. Mit dieser Eingabeaufforderung weiß T5, dass es sich um eine Zusammenfassungsaufgabe handelt.
+
+```bash
+python examples/pytorch/summarization/run_summarization.py \
+    --model_name_or_path t5-small \
+    --do_train \
+    --do_eval \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --source_prefix "summarize: " \
+    --output_dir /tmp/tst-summarization \
+    --per_device_train_batch_size=4 \
+    --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
+    --predict_with_generate
+```
+</pt>
+<tf>
+Das Beispielskript lädt einen Datensatz aus der 🤗 [Datasets](https://huggingface.co/docs/datasets/) Bibliothek herunter und verarbeitet ihn vor. Anschließend nimmt das Skript die Feinabstimmung eines Datensatzes mit Keras auf einer Architektur vor, die die Zusammenfassung unterstützt. Das folgende Beispiel zeigt, wie die Feinabstimmung von [T5-small](https://huggingface.co/t5-small) auf dem [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail) Datensatz durchgeführt wird. Das T5-Modell benötigt aufgrund der Art und Weise, wie es trainiert wurde, ein zusätzliches Argument `source_prefix`. Mit dieser Eingabeaufforderung weiß T5, dass es sich um eine Zusammenfassungsaufgabe handelt.
+
+```bash
+python examples/tensorflow/summarization/run_summarization.py  \
+    --model_name_or_path t5-small \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --output_dir /tmp/tst-summarization  \
+    --per_device_train_batch_size 8 \
+    --per_device_eval_batch_size 16 \
+    --num_train_epochs 3 \
+    --do_train \
+    --do_eval
+```
+</tf>
+</frameworkcontent>
+
+## Verteiltes Training und gemischte Präzision
+
+Der [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) unterstützt verteiltes Training und gemischte Präzision, d.h. Sie können ihn auch in einem Skript verwenden. So aktivieren Sie diese beiden Funktionen:
+
+- Fügen Sie das Argument `fp16` hinzu, um gemischte Genauigkeit zu aktivieren.
+- Legen Sie die Anzahl der zu verwendenden GPUs mit dem Argument `nproc_per_node` fest.
+
+```bash
+python -m torch.distributed.launch \
+    --nproc_per_node 8 pytorch/summarization/run_summarization.py \
+    --fp16 \
+    --model_name_or_path t5-small \
+    --do_train \
+    --do_eval \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --source_prefix "summarize: " \
+    --output_dir /tmp/tst-summarization \
+    --per_device_train_batch_size=4 \
+    --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
+    --predict_with_generate
+```
+
+TensorFlow-Skripte verwenden eine [`MirroredStrategy`](https://www.tensorflow.org/guide/distributed_training#mirroredstrategy) für verteiltes Training, und Sie müssen dem Trainingsskript keine zusätzlichen Argumente hinzufügen. Das TensorFlow-Skript verwendet standardmäßig mehrere GPUs, wenn diese verfügbar sind.
+
+## Ein Skript auf einer TPU ausführen
+
+<frameworkcontent>
+<pt>
+Tensor Processing Units (TPUs) sind speziell für die Beschleunigung der Leistung konzipiert. PyTorch unterstützt TPUs mit dem [XLA](https://www.tensorflow.org/xla) Deep Learning Compiler (siehe [hier](https://github.com/pytorch/xla/blob/master/README.md) für weitere Details). Um eine TPU zu verwenden, starten Sie das Skript `xla_spawn.py` und verwenden das Argument `num_cores`, um die Anzahl der TPU-Kerne festzulegen, die Sie verwenden möchten.
+
+```bash
+python xla_spawn.py --num_cores 8 \
+    summarization/run_summarization.py \
+    --model_name_or_path t5-small \
+    --do_train \
+    --do_eval \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --source_prefix "summarize: " \
+    --output_dir /tmp/tst-summarization \
+    --per_device_train_batch_size=4 \
+    --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
+    --predict_with_generate
+```
+</pt>
+<tf>
+Tensor Processing Units (TPUs) sind speziell für die Beschleunigung der Leistung konzipiert. TensorFlow Skripte verwenden eine [`TPUStrategy`](https://www.tensorflow.org/guide/distributed_training#tpustrategy) für das Training auf TPUs. Um eine TPU zu verwenden, übergeben Sie den Namen der TPU-Ressource an das Argument `tpu`.
+
+```bash
+python run_summarization.py  \
+    --tpu name_of_tpu_resource \
+    --model_name_or_path t5-small \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --output_dir /tmp/tst-summarization  \
+    --per_device_train_batch_size 8 \
+    --per_device_eval_batch_size 16 \
+    --num_train_epochs 3 \
+    --do_train \
+    --do_eval
+```
+</tf>
+</frameworkcontent>
+
+## Führen Sie ein Skript mit 🤗 Accelerate aus.
+
+🤗 [Accelerate](https://huggingface.co/docs/accelerate) ist eine reine PyTorch-Bibliothek, die eine einheitliche Methode für das Training eines Modells auf verschiedenen Arten von Setups (nur CPU, mehrere GPUs, TPUs) bietet und dabei die vollständige Transparenz der PyTorch-Trainingsschleife beibehält. Stellen Sie sicher, dass Sie 🤗 Accelerate installiert haben, wenn Sie es nicht bereits haben:
+
+> Hinweis: Da Accelerate schnell weiterentwickelt wird, muss die Git-Version von Accelerate installiert sein, um die Skripte auszuführen.
+```bash
+pip install git+https://github.com/huggingface/accelerate
+```
+
+Anstelle des Skripts `run_summarization.py` müssen Sie das Skript `run_summarization_no_trainer.py` verwenden. Die von Accelerate unterstützten Skripte haben eine Datei `task_no_trainer.py` im Ordner. Beginnen Sie mit dem folgenden Befehl, um eine Konfigurationsdatei zu erstellen und zu speichern:
+
+```bash
+accelerate config
+```
+
+Testen Sie Ihre Einrichtung, um sicherzustellen, dass sie korrekt konfiguriert ist:
+
+```bash
+accelerate test
+```
+
+Jetzt sind Sie bereit, das Training zu starten:
+
+```bash
+accelerate launch run_summarization_no_trainer.py \
+    --model_name_or_path t5-small \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --source_prefix "summarize: " \
+    --output_dir ~/tmp/tst-summarization
+```
+
+## Verwenden Sie einen benutzerdefinierten Datensatz
+
+Das Verdichtungsskript unterstützt benutzerdefinierte Datensätze, solange es sich um eine CSV- oder JSON-Line-Datei handelt. Wenn Sie Ihren eigenen Datensatz verwenden, müssen Sie mehrere zusätzliche Argumente angeben:
+
+- `train_file` und `validation_file` geben den Pfad zu Ihren Trainings- und Validierungsdateien an.
+- text_column` ist der Eingabetext, der zusammengefasst werden soll.
+- Summary_column" ist der auszugebende Zieltext.
+
+Ein Zusammenfassungsskript, das einen benutzerdefinierten Datensatz verwendet, würde wie folgt aussehen:
+
+```bash
+python examples/pytorch/summarization/run_summarization.py \
+    --model_name_or_path t5-small \
+    --do_train \
+    --do_eval \
+    --train_file path_to_csv_or_jsonlines_file \
+    --validation_file path_to_csv_or_jsonlines_file \
+    --text_column text_column_name \
+    --summary_column summary_column_name \
+    --source_prefix "summarize: " \
+    --output_dir /tmp/tst-summarization \
+    --overwrite_output_dir \
+    --per_device_train_batch_size=4 \
+    --per_device_eval_batch_size=4 \
+    --predict_with_generate
+```
+
+## Testen Sie ein Skript
+
+Es ist oft eine gute Idee, Ihr Skript an einer kleineren Anzahl von Beispielen für Datensätze auszuführen, um sicherzustellen, dass alles wie erwartet funktioniert, bevor Sie sich auf einen ganzen Datensatz festlegen, dessen Fertigstellung Stunden dauern kann. Verwenden Sie die folgenden Argumente, um den Datensatz auf eine maximale Anzahl von Stichproben zu beschränken:
+
+- `max_train_samples`
+- `max_eval_samples`
+- `max_predict_samples`
+
+```bash
+python examples/pytorch/summarization/run_summarization.py \
+    --model_name_or_path t5-small \
+    --max_train_samples 50 \
+    --max_eval_samples 50 \
+    --max_predict_samples 50 \
+    --do_train \
+    --do_eval \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --source_prefix "summarize: " \
+    --output_dir /tmp/tst-summarization \
+    --per_device_train_batch_size=4 \
+    --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
+    --predict_with_generate
+```
+
+Nicht alle Beispielskripte unterstützen das Argument `max_predict_samples`. Wenn Sie sich nicht sicher sind, ob Ihr Skript dieses Argument unterstützt, fügen Sie das Argument `-h` hinzu, um dies zu überprüfen:
+
+```bash
+examples/pytorch/summarization/run_summarization.py -h
+```
+
+## Training vom Kontrollpunkt fortsetzen
+
+Eine weitere hilfreiche Option, die Sie aktivieren können, ist die Wiederaufnahme des Trainings von einem früheren Kontrollpunkt aus. Auf diese Weise können Sie im Falle einer Unterbrechung Ihres Trainings dort weitermachen, wo Sie aufgehört haben, ohne von vorne beginnen zu müssen. Es gibt zwei Methoden, um das Training von einem Kontrollpunkt aus wieder aufzunehmen.
+
+Die erste Methode verwendet das Argument `output_dir previous_output_dir`, um das Training ab dem letzten in `output_dir` gespeicherten Kontrollpunkt wieder aufzunehmen. In diesem Fall sollten Sie `overwrite_output_dir` entfernen:
+
+```bash
+python examples/pytorch/summarization/run_summarization.py
+    --model_name_or_path t5-small \
+    --do_train \
+    --do_eval \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --source_prefix "summarize: " \
+    --output_dir /tmp/tst-summarization \
+    --per_device_train_batch_size=4 \
+    --per_device_eval_batch_size=4 \
+    --output_dir previous_output_dir \
+    --predict_with_generate
+```
+
+Die zweite Methode verwendet das Argument `Resume_from_checkpoint path_to_specific_checkpoint`, um das Training ab einem bestimmten Checkpoint-Ordner wieder aufzunehmen.
+
+```bash
+python examples/pytorch/summarization/run_summarization.py
+    --model_name_or_path t5-small \
+    --do_train \
+    --do_eval \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --source_prefix "summarize: " \
+    --output_dir /tmp/tst-summarization \
+    --per_device_train_batch_size=4 \
+    --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
+    --resume_from_checkpoint path_to_specific_checkpoint \
+    --predict_with_generate
+```
+
+## Teilen Sie Ihr Modell
+
+Alle Skripte können Ihr endgültiges Modell in den [Model Hub](https://huggingface.co/models) hochladen. Stellen Sie sicher, dass Sie bei Hugging Face angemeldet sind, bevor Sie beginnen:
+
+```bash
+huggingface-cli login
+```
+
+Dann fügen Sie dem Skript das Argument `push_to_hub` hinzu. Mit diesem Argument wird ein Repository mit Ihrem Hugging Face-Benutzernamen und dem in `output_dir` angegebenen Ordnernamen erstellt.
+
+Wenn Sie Ihrem Repository einen bestimmten Namen geben möchten, fügen Sie ihn mit dem Argument `push_to_hub_model_id` hinzu. Das Repository wird automatisch unter Ihrem Namensraum aufgeführt.
+
+Das folgende Beispiel zeigt, wie Sie ein Modell mit einem bestimmten Repository-Namen hochladen können:
+
+```bash
+python examples/pytorch/summarization/run_summarization.py
+    --model_name_or_path t5-small \
+    --do_train \
+    --do_eval \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --source_prefix "summarize: " \
+    --push_to_hub \
+    --push_to_hub_model_id finetuned-t5-cnn_dailymail \
+    --output_dir /tmp/tst-summarization \
+    --per_device_train_batch_size=4 \
+    --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
+    --predict_with_generate
+```

docs/source/de/transformers_agents.md

@@ -0,0 +1,323 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Transformers Agents
+
+<Tip warning={true}>
+
+Transformers Agents ist eine experimentelle API, die jederzeit geändert werden kann. Die von den Agenten zurückgegebenen Ergebnisse
+zurückgegeben werden, können variieren, da sich die APIs oder die zugrunde liegenden Modelle ändern können.
+
+</Tip>
+
+Transformers Version v4.29.0, die auf dem Konzept von *Tools* und *Agenten* aufbaut. Sie können damit spielen in
+[dieses Colab](https://colab.research.google.com/drive/1c7MHD-T1forUPGcC_jlwsIptOzpG3hSj).
+
+Kurz gesagt, es bietet eine API für natürliche Sprache auf der Grundlage von Transformers: Wir definieren eine Reihe von kuratierten Tools und entwerfen einen 
+Agenten, um natürliche Sprache zu interpretieren und diese Werkzeuge zu verwenden. Es ist von vornherein erweiterbar; wir haben einige relevante Tools kuratiert, 
+aber wir werden Ihnen zeigen, wie das System einfach erweitert werden kann, um jedes von der Community entwickelte Tool zu verwenden.
+
+Beginnen wir mit einigen Beispielen dafür, was mit dieser neuen API erreicht werden kann. Sie ist besonders leistungsfähig, wenn es um 
+Sie ist besonders leistungsstark, wenn es um multimodale Aufgaben geht. Lassen Sie uns also eine Runde drehen, um Bilder zu erzeugen und Text vorzulesen.
+
+```py
+agent.run("Caption the following image", image=image)
+```
+
+| **Input**                                                                                                                   | **Output**                        |
+|-----------------------------------------------------------------------------------------------------------------------------|-----------------------------------|
+| <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/beaver.png" width=200> | A beaver is swimming in the water |
+
+---
+
+```py
+agent.run("Read the following text out loud", text=text)
+```
+| **Input**                                                                                                               | **Output**                                   |
+|-------------------------------------------------------------------------------------------------------------------------|----------------------------------------------|
+| A beaver is swimming in the water | <audio controls><source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tts_example.wav" type="audio/wav"> your browser does not support the audio element. </audio>
+
+---
+
+```py
+agent.run(
+    "In the following `document`, where will the TRRF Scientific Advisory Council Meeting take place?",
+    document=document,
+)
+```
+| **Input**                                                                                                                   | **Output**     |
+|-----------------------------------------------------------------------------------------------------------------------------|----------------|
+| <img src="https://datasets-server.huggingface.co/assets/hf-internal-testing/example-documents/--/hf-internal-testing--example-documents/test/0/image/image.jpg" width=200> | ballroom foyer |
+
+## Schnellstart
+
+Bevor Sie `agent.run` verwenden können, müssen Sie einen Agenten instanziieren, der ein großes Sprachmodell (LLM) ist. 
+Wir bieten Unterstützung für openAI-Modelle sowie für OpenSource-Alternativen von BigCode und OpenAssistant. Die openAI
+Modelle sind leistungsfähiger (erfordern aber einen openAI-API-Schlüssel, können also nicht kostenlos verwendet werden); Hugging Face
+bietet kostenlosen Zugang zu Endpunkten für BigCode- und OpenAssistant-Modelle.
+
+To start with, please install the `agents` extras in order to install all default dependencies.
+```bash
+pip install transformers[agents]
+```
+
+Um openAI-Modelle zu verwenden, instanziieren Sie einen [`OpenAiAgent`], nachdem Sie die `openai`-Abhängigkeit installiert haben:
+
+```bash
+pip install openai
+```
+
+
+```py
+from transformers import OpenAiAgent
+
+agent = OpenAiAgent(model="text-davinci-003", api_key="<your_api_key>")
+```
+
+Um BigCode oder OpenAssistant zu verwenden, melden Sie sich zunächst an, um Zugriff auf die Inference API zu erhalten:
+
+```py
+from huggingface_hub import login
+
+login("<YOUR_TOKEN>")
+```
+
+Dann instanziieren Sie den Agenten
+
+```py
+from transformers import HfAgent
+
+# Starcoder
+agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder")
+# StarcoderBase
+# agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoderbase")
+# OpenAssistant
+# agent = HfAgent(url_endpoint="https://api-inference.huggingface.co/models/OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5")
+```
+
+Dies geschieht mit der Inferenz-API, die Hugging Face derzeit kostenlos zur Verfügung stellt. Wenn Sie Ihren eigenen Inferenz
+Endpunkt für dieses Modell (oder einen anderen) haben, können Sie die obige URL durch Ihren URL-Endpunkt ersetzen.
+
+<Tip>
+
+StarCoder und OpenAssistant sind kostenlos und leisten bei einfachen Aufgaben bewundernswert gute Arbeit. Allerdings halten die Kontrollpunkte
+nicht, wenn es um komplexere Aufforderungen geht. Wenn Sie mit einem solchen Problem konfrontiert sind, empfehlen wir Ihnen, das OpenAI
+Modell auszuprobieren, das zwar leider nicht quelloffen ist, aber zur Zeit eine bessere Leistung erbringt.
+
+</Tip>
+
+Sie sind jetzt startklar! Lassen Sie uns in die beiden APIs eintauchen, die Ihnen jetzt zur Verfügung stehen.
+
+### Einzelne Ausführung (run)
+
+Die Methode der einmaligen Ausführung ist die Verwendung der [`~Agent.run`] Methode des Agenten:
+
+```py
+agent.run("Draw me a picture of rivers and lakes.")
+```
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" width=200>
+
+Es wählt automatisch das (oder die) Werkzeug(e) aus, das (die) für die von Ihnen gewünschte Aufgabe geeignet ist (sind) und führt es (sie) entsprechend aus. Es
+kann eine oder mehrere Aufgaben in der gleichen Anweisung ausführen (je komplexer Ihre Anweisung ist, desto wahrscheinlicher ist ein
+der Agent scheitern).
+
+```py
+agent.run("Draw me a picture of the sea then transform the picture to add an island")
+```
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/sea_and_island.png" width=200>
+
+<br/>
+
+
+Jede [`~Agent.run`] Operation ist unabhängig, so dass Sie sie mehrmals hintereinander mit unterschiedlichen Aufgaben ausführen können.
+
+Beachten Sie, dass Ihr `Agent` nur ein großsprachiges Modell ist, so dass kleine Variationen in Ihrer Eingabeaufforderung völlig unterschiedliche Ergebnisse liefern können.
+unterschiedliche Ergebnisse liefern. Es ist wichtig, dass Sie die Aufgabe, die Sie ausführen möchten, so genau wie möglich erklären. Wir gehen noch weiter ins Detail
+wie man gute Prompts schreibt [hier](custom_tools#writing-good-user-inputs).
+
+Wenn Sie einen Status über Ausführungszeiten hinweg beibehalten oder dem Agenten Nicht-Text-Objekte übergeben möchten, können Sie dies tun, indem Sie
+Variablen, die der Agent verwenden soll. Sie könnten zum Beispiel das erste Bild von Flüssen und Seen erzeugen, 
+und das Modell bitten, dieses Bild zu aktualisieren und eine Insel hinzuzufügen, indem Sie Folgendes tun:
+
+```python
+picture = agent.run("Generate a picture of rivers and lakes.")
+updated_picture = agent.run("Transform the image in `picture` to add an island to it.", picture=picture)
+```
+
+<Tip>
+
+Dies kann hilfreich sein, wenn das Modell Ihre Anfrage nicht verstehen kann und die Werkzeuge verwechselt. Ein Beispiel wäre:
+
+```py
+agent.run("Draw me the picture of a capybara swimming in the sea")
+```
+
+Hier könnte das Modell auf zwei Arten interpretieren:
+- Die Funktion `Text-zu-Bild` erzeugt ein Wasserschwein, das im Meer schwimmt.
+- Oder Sie lassen das `Text-zu-Bild` ein Wasserschwein erzeugen und verwenden dann das Werkzeug `Bildtransformation`, um es im Meer schwimmen zu lassen.
+
+Falls Sie das erste Szenario erzwingen möchten, können Sie dies tun, indem Sie die Eingabeaufforderung als Argument übergeben:
+
+```py
+agent.run("Draw me a picture of the `prompt`", prompt="a capybara swimming in the sea")
+```
+
+</Tip>
+
+
+### Chat-basierte Ausführung (Chat)
+
+Der Agent verfügt auch über einen Chat-basierten Ansatz, der die Methode [`~Agent.chat`] verwendet:
+
+```py
+agent.chat("Generate a picture of rivers and lakes")
+```
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" width=200> 
+
+```py
+agent.chat("Transform the picture so that there is a rock in there")
+```
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes_and_beaver.png" width=200>
+
+<br/>
+
+Dies ist ein interessanter Ansatz, wenn Sie den Zustand über Anweisungen hinweg beibehalten möchten. Er ist besser für Experimente geeignet, 
+eignet sich aber eher für einzelne Anweisungen als für komplexe Anweisungen (die die [`~Agent.run`]
+Methode besser verarbeiten kann).
+
+Diese Methode kann auch Argumente entgegennehmen, wenn Sie Nicht-Text-Typen oder bestimmte Aufforderungen übergeben möchten.
+
+### ⚠️ Fernausführung
+
+Zu Demonstrationszwecken und damit es mit allen Setups verwendet werden kann, haben wir Remote-Executors für mehrere 
+der Standard-Tools erstellt, auf die der Agent in dieser Version Zugriff hat. Diese werden erstellt mit 
+[inference endpoints](https://huggingface.co/inference-endpoints).
+
+Wir haben diese vorerst deaktiviert, aber um zu sehen, wie Sie selbst Remote Executors Tools einrichten können,
+empfehlen wir die Lektüre des [custom tool guide](./custom_tools).
+
+### Was passiert hier? Was sind Tools und was sind Agenten?
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/diagram.png">
+
+#### Agenten
+
+Der "Agent" ist hier ein großes Sprachmodell, das wir auffordern, Zugang zu einem bestimmten Satz von Tools zu erhalten.
+
+LLMs sind ziemlich gut darin, kleine Codeproben zu erzeugen. Diese API macht sich das zunutze, indem sie das 
+LLM ein kleines Codebeispiel gibt, das eine Aufgabe mit einer Reihe von Werkzeugen ausführt. Diese Aufforderung wird dann ergänzt durch die 
+Aufgabe, die Sie Ihrem Agenten geben, und die Beschreibung der Werkzeuge, die Sie ihm geben. Auf diese Weise erhält er Zugriff auf die Dokumentation der 
+Tools, insbesondere die erwarteten Eingaben und Ausgaben, und kann den entsprechenden Code generieren.
+
+#### Tools
+
+Tools sind sehr einfach: Sie bestehen aus einer einzigen Funktion mit einem Namen und einer Beschreibung. Wir verwenden dann die Beschreibungen dieser Tools 
+um den Agenten aufzufordern. Anhand der Eingabeaufforderung zeigen wir dem Agenten, wie er die Tools nutzen kann, um das zu tun, was in der 
+in der Abfrage angefordert wurde.
+
+Dies geschieht mit brandneuen Tools und nicht mit Pipelines, denn der Agent schreibt besseren Code mit sehr atomaren Tools. 
+Pipelines sind stärker refaktorisiert und fassen oft mehrere Aufgaben in einer einzigen zusammen. Tools sind dafür gedacht, sich auf
+eine einzige, sehr einfache Aufgabe konzentrieren.
+
+#### Code-Ausführung?!
+
+Dieser Code wird dann mit unserem kleinen Python-Interpreter auf den mit Ihren Tools übergebenen Eingaben ausgeführt. 
+Wir hören Sie schon schreien "Willkürliche Codeausführung!", aber lassen Sie uns erklären, warum das nicht der Fall ist.
+
+Die einzigen Funktionen, die aufgerufen werden können, sind die von Ihnen zur Verfügung gestellten Tools und die Druckfunktion, so dass Sie bereits eingeschränkt sind 
+eingeschränkt, was ausgeführt werden kann. Sie sollten sicher sein, wenn es sich auf die Werkzeuge für das Umarmungsgesicht beschränkt. 
+
+Dann lassen wir keine Attributsuche oder Importe zu (die ohnehin nicht benötigt werden, um die 
+Inputs/Outputs an eine kleine Gruppe von Funktionen), so dass alle offensichtlichen Angriffe (und Sie müssten den LLM 
+dazu auffordern, sie auszugeben) kein Problem darstellen sollten. Wenn Sie auf Nummer sicher gehen wollen, können Sie die 
+run()-Methode mit dem zusätzlichen Argument return_code=True ausführen. In diesem Fall gibt der Agent nur den auszuführenden Code 
+zur Ausführung zurück und Sie können entscheiden, ob Sie ihn ausführen möchten oder nicht.
+
+Die Ausführung bricht bei jeder Zeile ab, in der versucht wird, eine illegale Operation auszuführen, oder wenn ein regulärer Python-Fehler 
+mit dem vom Agenten generierten Code.
+
+### Ein kuratierter Satz von Tools
+
+Wir haben eine Reihe von Tools identifiziert, die solche Agenten unterstützen können. Hier ist eine aktualisierte Liste der Tools, die wir integriert haben 
+in `transformers` integriert haben:
+
+- **Beantwortung von Fragen zu Dokumenten**: Beantworten Sie anhand eines Dokuments (z.B. PDF) im Bildformat eine Frage zu diesem Dokument ([Donut](./model_doc/donut))
+- Beantworten von Textfragen**: Geben Sie einen langen Text und eine Frage an, beantworten Sie die Frage im Text ([Flan-T5](./model_doc/flan-t5))
+- **Unbedingte Bildunterschriften**: Beschriften Sie das Bild! ([BLIP](./model_doc/blip))
+- **Bildfragebeantwortung**: Beantworten Sie bei einem Bild eine Frage zu diesem Bild ([VILT](./model_doc/vilt))
+- **Bildsegmentierung**: Geben Sie ein Bild und einen Prompt an und geben Sie die Segmentierungsmaske dieses Prompts aus ([CLIPSeg](./model_doc/clipseg))
+- **Sprache in Text**: Geben Sie eine Audioaufnahme einer sprechenden Person an und transkribieren Sie die Sprache in Text ([Whisper](./model_doc/whisper))
+- **Text in Sprache**: wandelt Text in Sprache um ([SpeechT5](./model_doc/speecht5))
+- **Zero-Shot-Textklassifizierung**: Ermitteln Sie anhand eines Textes und einer Liste von Bezeichnungen, welcher Bezeichnung der Text am ehesten entspricht ([BART](./model_doc/bart))
+- **Textzusammenfassung**: fassen Sie einen langen Text in einem oder wenigen Sätzen zusammen ([BART](./model_doc/bart))
+- **Übersetzung**: Übersetzen des Textes in eine bestimmte Sprache ([NLLB](./model_doc/nllb))
+
+Diese Tools sind in Transformatoren integriert und können auch manuell verwendet werden, zum Beispiel:
+
+```py
+from transformers import load_tool
+
+tool = load_tool("text-to-speech")
+audio = tool("This is a text to speech tool")
+```
+
+### Benutzerdefinierte Tools
+
+Wir haben zwar eine Reihe von Tools identifiziert, sind aber der festen Überzeugung, dass der Hauptwert dieser Implementierung darin besteht 
+die Möglichkeit, benutzerdefinierte Tools schnell zu erstellen und weiterzugeben.
+
+Indem Sie den Code eines Tools in einen Hugging Face Space oder ein Modell-Repository stellen, können Sie das Tool 
+direkt mit dem Agenten nutzen. Wir haben ein paar neue Funktionen hinzugefügt 
+**transformers-agnostic** Tools zur [`huggingface-tools` Organisation](https://huggingface.co/huggingface-tools) hinzugefügt:
+
+- **Text-Downloader**: zum Herunterladen eines Textes von einer Web-URL
+- **Text zu Bild**: erzeugt ein Bild nach einer Eingabeaufforderung und nutzt dabei stabile Diffusion
+- **Bildtransformation**: verändert ein Bild anhand eines Ausgangsbildes und einer Eingabeaufforderung, unter Ausnutzung der stabilen pix2pix-Diffusion
+- **Text zu Video**: Erzeugen eines kleinen Videos nach einer Eingabeaufforderung, unter Verwendung von damo-vilab
+
+Das Text-zu-Bild-Tool, das wir von Anfang an verwendet haben, ist ein Remote-Tool, das sich in 
+[*huggingface-tools/text-to-image*](https://huggingface.co/spaces/huggingface-tools/text-to-image)! Wir werden
+weiterhin solche Tools für diese und andere Organisationen veröffentlichen, um diese Implementierung weiter zu verbessern.
+
+Die Agenten haben standardmäßig Zugriff auf die Tools, die sich auf [*huggingface-tools*](https://huggingface.co/huggingface-tools) befinden.
+Wie Sie Ihre eigenen Tools schreiben und freigeben können und wie Sie jedes benutzerdefinierte Tool, das sich auf dem Hub befindet, nutzen können, erklären wir in [folgender Anleitung](custom_tools).
+
+### Code-Erzeugung
+
+Bisher haben wir gezeigt, wie Sie die Agenten nutzen können, um Aktionen für Sie durchzuführen. Der Agent generiert jedoch nur Code
+den wir dann mit einem sehr eingeschränkten Python-Interpreter ausführen. Falls Sie den generierten Code in einer anderen Umgebung verwenden möchten 
+einer anderen Umgebung verwenden möchten, können Sie den Agenten auffordern, den Code zusammen mit einer Tooldefinition und genauen Importen zurückzugeben.
+
+Zum Beispiel die folgende Anweisung
+```python
+agent.run("Draw me a picture of rivers and lakes", return_code=True)
+```
+
+gibt den folgenden Code zurück
+
+```python
+from transformers import load_tool
+
+image_generator = load_tool("huggingface-tools/text-to-image")
+
+image = image_generator(prompt="rivers and lakes")
+```
+
+die Sie dann selbst ändern und ausführen können.

docs/source/en/_toctree.yml

@@ -88,6 +88,11 @@
     - local: generation_strategies
       title: Customize the generation strategy
     title: Generation
+  - isExpanded: false
+    sections:
+    - local: tasks/idefics
+      title: Image tasks with IDEFICS
+    title: Prompting
   title: Task Guides
 - sections:
   - local: fast_tokenizers
@@ -98,6 +103,8 @@
     title: Use model-specific APIs
   - local: custom_models
     title: Share a custom model
+  - local: chat_templating
+    title: Templates for chat models
   - local: sagemaker
     title: Run training on Amazon SageMaker
   - local: serialization
@@ -377,6 +384,8 @@
         title: MegatronBERT
       - local: model_doc/megatron_gpt2
         title: MegatronGPT2
+      - local: model_doc/mistral
+        title: Mistral
       - local: model_doc/mluke
         title: mLUKE
       - local: model_doc/mobilebert
@@ -407,6 +416,8 @@
         title: Pegasus
       - local: model_doc/pegasus_x
         title: PEGASUS-X
+      - local: model_doc/persimmon
+        title: Persimmon
       - local: model_doc/phobert
         title: PhoBERT
       - local: model_doc/plbart
@@ -564,6 +575,8 @@
         title: ViTDet
       - local: model_doc/vit_mae
         title: ViTMAE
+      - local: model_doc/vitmatte
+        title: ViTMatte
       - local: model_doc/vit_msn
         title: ViTMSN
       - local: model_doc/vivit
@@ -634,6 +647,8 @@
         title: BLIP-2
       - local: model_doc/bridgetower
         title: BridgeTower
+      - local: model_doc/bros
+        title: BROS
       - local: model_doc/chinese_clip
         title: Chinese-CLIP
       - local: model_doc/clip
@@ -672,6 +687,8 @@
         title: MatCha
       - local: model_doc/mgp-str
         title: MGP-STR
+      - local: model_doc/nougat
+        title: Nougat
       - local: model_doc/oneformer
         title: OneFormer
       - local: model_doc/owlvit

docs/source/en/add_new_model.md

@@ -52,7 +52,7 @@ A good first starting point to better understand the library is to read the [doc
 
 In our opinion, the library's code is not just a means to provide a product, *e.g.* the ability to use BERT for
 inference, but also as the very product that we want to improve. Hence, when adding a model, the user is not only the
-person that will use your model, but also everybody that will read, try to understand, and possibly tweak your code.
+person who will use your model, but also everybody who will read, try to understand, and possibly tweak your code.
 
 With this in mind, let's go a bit deeper into the general library design.
 
@@ -131,9 +131,9 @@ From experience, we can tell you that the most important things to keep in mind
   friends. Note that it might very well happen that your model's tokenizer is based on one model implementation, and
   your model's modeling code on another one. *E.g.* FSMT's modeling code is based on BART, while FSMT's tokenizer code
   is based on XLM.
--  It's more of an engineering challenge than a scientific challenge. You should spend more time on creating an
-  efficient debugging environment than trying to understand all theoretical aspects of the model in the paper.
--  Ask for help, when you're stuck! Models are the core component of 🤗 Transformers so that we at Hugging Face are more
+-  It's more of an engineering challenge than a scientific challenge. You should spend more time creating an
+  efficient debugging environment rather than trying to understand all theoretical aspects of the model in the paper.
+-  Ask for help, when you're stuck! Models are the core component of 🤗 Transformers so we at Hugging Face are more
   than happy to help you at every step to add your model. Don't hesitate to ask if you notice you are not making
   progress.
 
@@ -157,9 +157,9 @@ List:
 ☐ Submitted the pull request<br>
 ☐ (Optional) Added a demo notebook
 
-To begin with, we usually recommend to start by getting a good theoretical understanding of `BrandNewBert`. However,
+To begin with, we usually recommend starting by getting a good theoretical understanding of `BrandNewBert`. However,
 if you prefer to understand the theoretical aspects of the model *on-the-job*, then it is totally fine to directly dive
-into the `BrandNewBert`'s code-base. This option might suit you better, if your engineering skills are better than
+into the `BrandNewBert`'s code-base. This option might suit you better if your engineering skills are better than
 your theoretical skill, if you have trouble understanding `BrandNewBert`'s paper, or if you just enjoy programming
 much more than reading scientific papers.
 
@@ -175,7 +175,7 @@ theoretical aspects, but rather focus on the practical ones, namely:
   encoder-decoder model? Look at the [model_summary](model_summary) if you're not familiar with the differences between those.
 -  What are the applications of *brand_new_bert*? Text classification? Text generation? Seq2Seq tasks, *e.g.,*
   summarization?
--  What is the novel feature of the model making it different from BERT/GPT-2/BART?
+-  What is the novel feature of the model that makes it different from BERT/GPT-2/BART?
 -  Which of the already existing [🤗 Transformers models](https://huggingface.co/transformers/#contents) is most
   similar to *brand_new_bert*?
 -  What type of tokenizer is used? A sentencepiece tokenizer? Word piece tokenizer? Is it the same tokenizer as used
@@ -261,7 +261,7 @@ figure out the following:
 - How can you debug the model in the original environment of the repo? Do you have to add *print* statements, can you
   work with an interactive debugger like *ipdb*, or should you use an efficient IDE to debug the model, like PyCharm?
 
-It is very important that before you start the porting process, that you can **efficiently** debug code in the original
+It is very important that before you start the porting process, you can **efficiently** debug code in the original
 repository! Also, remember that you are working with an open-source library, so do not hesitate to open an issue, or
 even a pull request in the original repository. The maintainers of this repository are most likely very happy about
 someone looking into their code!
@@ -280,10 +280,10 @@ In general, there are two possible debugging environments for running the origin
 Jupyter notebooks have the advantage that they allow for cell-by-cell execution which can be helpful to better split
 logical components from one another and to have faster debugging cycles as intermediate results can be stored. Also,
 notebooks are often easier to share with other contributors, which might be very helpful if you want to ask the Hugging
-Face team for help. If you are familiar with Jupyter notebooks, we strongly recommend you to work with them.
+Face team for help. If you are familiar with Jupyter notebooks, we strongly recommend you work with them.
 
 The obvious disadvantage of Jupyter notebooks is that if you are not used to working with them you will have to spend
-some time adjusting to the new programming environment and that you might not be able to use your known debugging tools
+some time adjusting to the new programming environment and you might not be able to use your known debugging tools
 anymore, like `ipdb`.
 
 For each code-base, a good first step is always to load a **small** pretrained checkpoint and to be able to reproduce a
@@ -329,7 +329,7 @@ example is [T5's MeshTensorFlow](https://github.com/tensorflow/mesh/tree/master/
 very complex and does not offer a simple way to decompose the model into its sub-components. For such libraries, one
 often relies on verifying print statements.
 
-No matter which strategy you choose, the recommended procedure is often the same in that you should start to debug the
+No matter which strategy you choose, the recommended procedure is often the same that you should start to debug the
 starting layers first and the ending layers last.
 
 It is recommended that you retrieve the output, either by print statements or sub-component functions, of the following
@@ -364,7 +364,7 @@ depending on the library framework, we accept an error tolerance of 1e-3 (0.001)
 nearly the same output, they have to be almost identical. Therefore, you will certainly compare the intermediate
 outputs of the 🤗 Transformers version multiple times against the intermediate outputs of the original implementation of
 *brand_new_bert* in which case an **efficient** debugging environment of the original repository is absolutely
-important. Here is some advice is to make your debugging environment as efficient as possible.
+important. Here is some advice to make your debugging environment as efficient as possible.
 
 - Find the best way of debugging intermediate results. Is the original repository written in PyTorch? Then you should
   probably take the time to write a longer script that decomposes the original model into smaller sub-components to
@@ -409,7 +409,7 @@ Otherwise, let's start generating a new model. You have two choices here:
 - `transformers-cli add-new-model-like` to add a new model like an existing one
 - `transformers-cli add-new-model` to add a new model from our template (will look like BERT or Bart depending on the type of model you select)
 
-In both cases, you will be prompted with a questionnaire to fill the basic information of your model. The second command requires to install `cookiecutter`, you can find more information on it [here](https://github.com/huggingface/transformers/tree/main/templates/adding_a_new_model).
+In both cases, you will be prompted with a questionnaire to fill in the basic information of your model. The second command requires to install `cookiecutter`, you can find more information on it [here](https://github.com/huggingface/transformers/tree/main/templates/adding_a_new_model).
 
 **Open a Pull Request on the main huggingface/transformers repo**
 
@@ -451,7 +451,7 @@ git push -u origin a-descriptive-name-for-my-changes
 
 6. Change the PR into a draft by clicking on “Convert to draft” on the right of the GitHub pull request web page.
 
-In the following, whenever you have done some progress, don't forget to commit your work and push it to your account so
+In the following, whenever you have made some progress, don't forget to commit your work and push it to your account so
 that it shows in the pull request. Additionally, you should make sure to update your work with the current main from
 time to time by doing:
 
@@ -483,7 +483,7 @@ Now you can finally start coding :). The generated code in
 `src/transformers/models/brand_new_bert/modeling_brand_new_bert.py` will either have the same architecture as BERT if
 it's an encoder-only model or BART if it's an encoder-decoder model. At this point, you should remind yourself what
 you've learned in the beginning about the theoretical aspects of the model: *How is the model different from BERT or
-BART?*". Implement those changes which often means to change the *self-attention* layer, the order of the normalization
+BART?*". Implement those changes which often means changing the *self-attention* layer, the order of the normalization
 layer, etc… Again, it is often useful to look at the similar architecture of already existing models in Transformers to
 get a better feeling of how your model should be implemented.
 
@@ -665,7 +665,7 @@ PyTorch's implementation of a layer requires the weight to be transposed beforeh
 
 Finally, you should also check that **all** required weights are initialized and print out all checkpoint weights that
 were not used for initialization to make sure the model is correctly converted. It is completely normal, that the
-conversion trials fail with either a wrong shape statement or wrong name assignment. This is most likely because either
+conversion trials fail with either a wrong shape statement or a wrong name assignment. This is most likely because either
 you used incorrect parameters in `BrandNewBertConfig()`, have a wrong architecture in the 🤗 Transformers
 implementation, you have a bug in the `init()` functions of one of the components of the 🤗 Transformers
 implementation or you need to transpose one of the checkpoint weights.
@@ -722,7 +722,7 @@ in the 🤗 Transformers implementation. From our experience, a simple and effic
 in both the original implementation and 🤗 Transformers implementation, at the same positions in the network
 respectively, and to successively remove print statements showing the same values for intermediate presentations.
 
-When you're confident that both implementations yield the same output, verifying the outputs with
+When you're confident that both implementations yield the same output, verify the outputs with
 `torch.allclose(original_output, output, atol=1e-3)`, you're done with the most difficult part! Congratulations - the
 work left to be done should be a cakewalk 😊.
 
@@ -744,7 +744,7 @@ Having fixed all common tests, it is now crucial to ensure that all the nice wor
 - b) Future changes to your model will not break any important feature of the model.
 
 At first, integration tests should be added. Those integration tests essentially do the same as the debugging scripts
-you used earlier to implement the model to 🤗 Transformers. A template of those model tests is already added by the
+you used earlier to implement the model to 🤗 Transformers. A template of those model tests has already added by the
 Cookiecutter, called `BrandNewBertModelIntegrationTests` and only has to be filled out by you. To ensure that those
 tests are passing, run
 
@@ -769,7 +769,7 @@ ways:
 
 **9. Implement the tokenizer**
 
-Next, we should add the tokenizer of *brand_new_bert*. Usually, the tokenizer is equivalent or very similar to an
+Next, we should add the tokenizer of *brand_new_bert*. Usually, the tokenizer is equivalent to or very similar to an
 already existing tokenizer of 🤗 Transformers.
 
 It is very important to find/extract the original tokenizer file and to manage to load this file into the 🤗
@@ -890,6 +890,6 @@ reviewer.
 Now, it's time to get some credit from the community for your work! Having completed a model addition is a major
 contribution to Transformers and the whole NLP community. Your code and the ported pre-trained models will certainly be
 used by hundreds and possibly even thousands of developers and researchers. You should be proud of your work and share
-your achievement with the community.
+your achievements with the community.
 
 **You have made another model that is super easy to access for everyone in the community! 🤯**

docs/source/en/add_new_pipeline.md

@@ -111,8 +111,8 @@ def _sanitize_parameters(self, **kwargs):
 ```
 
 Try to keep the inputs/outputs very simple and ideally JSON-serializable as it makes the pipeline usage very easy
-without requiring users to understand new kind of objects. It's also relatively common to support many different types
-of arguments for ease of use (audio files, can be filenames, URLs or pure bytes)
+without requiring users to understand new kinds of objects. It's also relatively common to support many different types
+of arguments for ease of use (audio files, which can be filenames, URLs or pure bytes)
 
 
 
@@ -219,8 +219,8 @@ repo.push_to_hub()
 ```
 
 This will copy the file where you defined `PairClassificationPipeline` inside the folder `"test-dynamic-pipeline"`,
-along with saving the model and tokenizer of the pipeline, before pushing everything in the repository
-`{your_username}/test-dynamic-pipeline`. After that anyone can use it as long as they provide the option
+along with saving the model and tokenizer of the pipeline, before pushing everything into the repository
+`{your_username}/test-dynamic-pipeline`. After that, anyone can use it as long as they provide the option
 `trust_remote_code=True`:
 
 ```py
@@ -232,9 +232,9 @@ classifier = pipeline(model="{your_username}/test-dynamic-pipeline", trust_remot
 ## Add the pipeline to 🤗 Transformers
 
 If you want to contribute your pipeline to 🤗 Transformers, you will need to add a new module in the `pipelines` submodule
-with the code of your pipeline, then add it in the list of tasks defined in `pipelines/__init__.py`.
+with the code of your pipeline, then add it to the list of tasks defined in `pipelines/__init__.py`.
 
-Then you will need to add tests. Create a new file `tests/test_pipelines_MY_PIPELINE.py` with example with the other tests.
+Then you will need to add tests. Create a new file `tests/test_pipelines_MY_PIPELINE.py` with examples of the other tests.
 
 The `run_pipeline_test` function will be very generic and run on small random models on every possible
 architecture as defined by `model_mapping` and `tf_model_mapping`.

docs/source/en/add_tensorflow_model.md

@@ -229,7 +229,6 @@ documentation pages. You can complete this part entirely following the patterns
 changes:
 - Include all public classes of *BrandNewBert* in `src/transformers/__init__.py`
 - Add *BrandNewBert* classes to the corresponding Auto classes in `src/transformers/models/auto/modeling_tf_auto.py`
-- Include the modeling file in the documentation test file list in `utils/documentation_tests.txt`
 - Add the lazy loading classes related to *BrandNewBert* in `src/transformers/utils/dummy_tf_objects.py`
 - Update the import structures for the public classes in `src/transformers/models/brand_new_bert/__init__.py`
 - Add the documentation pointers to the public methods of *BrandNewBert* in `docs/source/en/model_doc/brand_new_bert.md`

docs/source/en/chat_templating.md

@@ -0,0 +1,255 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Templates for Chat Models
+
+## Introduction
+
+An increasingly common use case for LLMs is **chat**. In a chat context, rather than continuing a single string
+of text (as is the case with a standard language model), the model instead continues a conversation that consists
+of one or more **messages**, each of which includes a **role** as well as message text.
+
+Most commonly, these roles are "user" for messages sent by the user, and "assistant" for messages sent by the model.
+Some models also support a "system" role. System messages are usually sent at the beginning of the conversation
+and include directives about how the model should behave in the subsequent chat.
+
+All language models, including models fine-tuned for chat, operate on linear sequences of tokens and do not intrinsically
+have special handling for roles. This means that role information is usually injected by adding control tokens
+between messages, to indicate both the message boundary and the relevant roles.
+
+Unfortunately, there isn't (yet!) a standard for which tokens to use, and so different models have been trained
+with wildly different formatting and control tokens for chat. This can be a real problem for users - if you use the
+wrong format, then the model will be confused by your input, and your performance will be a lot worse than it should be.
+This is the problem that **chat templates** aim to resolve. 
+
+Chat conversations are typically represented as a list of dictionaries, where each dictionary contains `role`
+and `content` keys, and represents a single chat message. Chat templates are strings containing a Jinja template that
+specifies how to format a conversation for a given model into a single tokenizable sequence. By storing this information
+with the tokenizer, we can ensure that models get input data in the format they expect.
+
+Let's make this concrete with a quick example using the `BlenderBot` model. BlenderBot has an extremely simple default 
+template, which mostly just adds whitespace between rounds of dialogue:
+
+```python
+>>> from transformers import AutoTokenizer
+>>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+
+>>> chat = [
+...   {"role": "user", "content": "Hello, how are you?"},
+...   {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
+...   {"role": "user", "content": "I'd like to show off how chat templating works!"},
+... ]
+
+>>> tokenizer.apply_chat_template(chat, tokenize=False)
+" Hello, how are you?  I'm doing great. How can I help you today?   I'd like to show off how chat templating works!</s>"
+```
+
+Notice how the entire chat is condensed into a single string. If we use `tokenize=True`, which is the default setting,
+that string will also be tokenized for us. To see a more complex template in action, though, let's use the 
+`meta-llama/Llama-2-7b-chat-hf` model. Note that this model has gated access, so you will have to
+[request access on the repo](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) if you want to run this code yourself:
+
+```python
+>> from transformers import AutoTokenizer
+>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
+
+>> chat = [
+...   {"role": "user", "content": "Hello, how are you?"},
+...   {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
+...   {"role": "user", "content": "I'd like to show off how chat templating works!"},
+... ]
+
+>> tokenizer.use_default_system_prompt = False
+>> tokenizer.apply_chat_template(chat, tokenize=False)
+"<s>[INST] Hello, how are you? [/INST] I'm doing great. How can I help you today? </s><s>[INST] I'd like to show off how chat templating works! [/INST]"
+```
+
+Note that this time, the tokenizer has added the control tokens [INST] and [/INST] to indicate the start and end of 
+user messages (but not assistant messages!)
+
+## How do chat templates work?
+
+The chat template for a model is stored on the `tokenizer.chat_template` attribute. If no chat template is set, the
+default template for that model class is used instead. Let's take a look at the template for `BlenderBot`:
+
+```python
+
+>>> from transformers import AutoTokenizer
+>>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+
+>>> tokenizer.default_chat_template
+"{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ '  ' }}{% endif %}{% endfor %}{{ eos_token }}"
+```
+
+That's kind of intimidating. Let's add some newlines and indentation to make it more readable. Note that
+we remove the first newline after each block as well as any preceding whitespace before a block by default, using the 
+Jinja `trim_blocks` and `lstrip_blocks` flags. This means that you can write your templates with indentations and 
+newlines and still have them function correctly!
+
+```
+{% for message in messages %}
+    {% if message['role'] == 'user' %}
+        {{ ' ' }}
+    {% endif %}
+    {{ message['content'] }}
+    {% if not loop.last %}
+        {{ '  ' }}
+    {% endif %}
+{% endfor %}
+{{ eos_token }}
+```
+
+If you've never seen one of these before, this is a [Jinja template](https://jinja.palletsprojects.com/en/3.1.x/templates/).
+Jinja is a templating language that allows you to write simple code that generates text. In many ways, the code and
+syntax resembles Python. In pure Python, this template would look something like this:
+
+```python
+for idx, message in enumerate(messages):
+    if message['role'] == 'user':
+        print(' ')
+    print(message['content'])
+    if not idx == len(messages) - 1:  # Check for the last message in the conversation
+        print('  ')
+print(eos_token)
+```
+
+Effectively, the template does three things:
+1. For each message, if the message is a user message, add a blank space before it, otherwise print nothing.
+2. Add the message content
+3. If the message is not the last message, add two spaces after it. After the final message, print the EOS token.
+
+This is a pretty simple template - it doesn't add any control tokens, and it doesn't support "system" messages, which 
+are a common way to give the model directives about how it should behave in the subsequent conversation.
+But Jinja gives you a lot of flexibility to do those things! Let's see a Jinja template that can format inputs
+similarly to the way LLaMA formats them (note that the real LLaMA template includes handling for default system
+messages and slightly different system message handling in general - don't use this one in your actual code!)
+
+```
+{% for message in messages %}
+    {% if message['role'] == 'user' %}
+        {{ bos_token + '[INST] ' + message['content'] + ' [/INST]' }}
+    {% elif message['role'] == 'system' %}
+        {{ '<<SYS>>\\n' + message['content'] + '\\n<</SYS>>\\n\\n' }}
+    {% elif message['role'] == 'assistant' %}
+        {{ ' '  + message['content'] + ' ' + eos_token }}
+    {% endif %}
+{% endfor %}
+```
+
+Hopefully if you stare at this for a little bit you can see what this template is doing - it adds specific tokens based
+on the "role" of each message, which represents who sent it. User, assistant and system messages are clearly
+distinguishable to the model because of the tokens they're wrapped in.
+
+## How do I create a chat template?
+
+Simple, just write a jinja template and set `tokenizer.chat_template`. You may find it easier to start with an 
+existing template from another model and simply edit it for your needs! For example, we could take the LLaMA template
+above and add "[ASST]" and "[/ASST]" to assistant messages:
+
+```
+{% for message in messages %}
+    {% if message['role'] == 'user' %}
+        {{ bos_token + '[INST] ' + message['content'].strip() + ' [/INST]' }}
+    {% elif message['role'] == 'system' %}
+        {{ '<<SYS>>\\n' + message['content'].strip() + '\\n<</SYS>>\\n\\n' }}
+    {% elif message['role'] == 'assistant' %}
+        {{ '[ASST] '  + message['content'] + ' [/ASST]' + eos_token }}
+    {% endif %}
+{% endfor %}
+```
+
+Now, simply set the `tokenizer.chat_template` attribute. Next time you use [`~PreTrainedTokenizer.apply_chat_template`], it will
+use your new template! This attribute will be saved in the `tokenizer_config.json` file, so you can use
+[`~utils.PushToHubMixin.push_to_hub`] to upload your new template to the Hub and make sure everyone's using the right
+template for your model!
+
+```python
+template = tokenizer.chat_template
+template = template.replace("SYS", "SYSTEM")  # Change the system token
+tokenizer.chat_template = template  # Set the new template
+tokenizer.push_to_hub("model_name")  # Upload your new template to the Hub!
+```
+
+The method [`~PreTrainedTokenizer.apply_chat_template`] which uses your chat template is called by the [`ConversationalPipeline`] class, so 
+once you set the correct chat template, your model will automatically become compatible with [`ConversationalPipeline`].
+
+## What are "default" templates?
+
+Before the introduction of chat templates, chat handling was hardcoded at the model class level. For backwards 
+compatibility, we have retained this class-specific handling as default templates, also set at the class level. If a
+model does not have a chat template set, but there is a default template for its model class, the `ConversationalPipeline`
+class and methods like `apply_chat_template` will use the class template instead. You can find out what the default
+template for your tokenizer is by checking the `tokenizer.default_chat_template` attribute.
+
+This is something we do purely for backward compatibility reasons, to avoid breaking any existing workflows. Even when
+the class template is appropriate for your model, we strongly recommend overriding the default template by
+setting the `chat_template` attribute explicitly to make it clear to users that your model has been correctly configured
+for chat, and to future-proof in case the default templates are ever altered or deprecated.
+
+## What template should I use?
+
+When setting the template for a model that's already been trained for chat, you should ensure that the template
+exactly matches the message formatting that the model saw during training, or else you will probably experience
+performance degradation. This is true even if you're training the model further - you will probably get the best 
+performance if you keep the chat tokens constant. This is very analogous to tokenization - you generally get the
+best performance for inference or fine-tuning when you precisely match the tokenization used during training.
+
+If you're training a model from scratch, or fine-tuning a base language model for chat, on the other hand,
+you have a lot of freedom to choose an appropriate template! LLMs are smart enough to learn to handle lots of different
+input formats. Our default template for models that don't have a class-specific template follows the 
+[ChatML format](https://github.com/openai/openai-python/blob/main/chatml.md), and this is a good, flexible choice for many use-cases. It looks like this:
+
+```
+{% for message in messages %}
+    {{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}
+{% endfor %}
+```
+
+If you like this one, here it is in one-liner form, ready to copy into your code:
+
+```
+tokenizer.chat_template = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}"
+```
+
+This template wraps each message in `<|im_start|>` and `<|im_end|>` tokens, and simply writes the role as a string, which
+allows for flexibility in the roles you train with. The output looks like this:
+
+```
+<|im_start|>system
+You are a helpful chatbot that will do its best not to say anything so stupid that people tweet about it.<|im_end|>
+<|im_start|>user
+How are you?<|im_end|>
+<|im_start|>assistant
+I'm doing great!<|im_end|>
+```
+
+The "user", "system" and "assistant" roles are the standard for chat, and we recommend using them when it makes sense,
+particularly if you want your model to operate well with [`ConversationalPipeline`]. However, you are not limited
+to these roles - templating is extremely flexible, and any string can be a role.
+
+## I want to use chat templates! How should I get started?
+
+If you have any chat models, you should set their `tokenizer.chat_template` attribute and test it using
+[`~PreTrainedTokenizer.apply_chat_template`]. This applies even if you're not the model owner - if you're using a model
+with an empty chat template, or one that's still using the default class template, please open a [pull request](https://huggingface.co/docs/hub/repositories-pull-requests-discussions) to
+the model repository so that this attribute can be set properly!
+
+Once the attribute is set, that's it, you're done! `tokenizer.apply_chat_template` will now work correctly for that
+model, which means it is also automatically supported in places like `ConversationalPipeline`!
+
+By ensuring that models have this attribute, we can make sure that the whole community gets to use the full power of
+open-source models. Formatting mismatches have been haunting the field and silently harming performance for too long - 
+it's time to put an end to them!

docs/source/en/custom_tools.md

@@ -25,7 +25,7 @@ If you are not aware of what tools and agents are in the context of transformers
 
 <Tip warning={true}>
 
-Transformers Agent is an experimental API that is subject to change at any time. Results returned by the agents
+Transformers Agents is an experimental API that is subject to change at any time. Results returned by the agents
 can vary as the APIs or underlying models are prone to change.
 
 </Tip>

docs/source/en/generation_strategies.md

@@ -55,12 +55,10 @@ When you load a model explicitly, you can inspect the generation configuration t
 >>> from transformers import AutoModelForCausalLM
 
 >>> model = AutoModelForCausalLM.from_pretrained("distilgpt2")
->>> model.generation_config  # doctest: +IGNORE_RESULT
+>>> model.generation_config
 GenerationConfig {
-    "_from_model_config": true,
     "bos_token_id": 50256,
     "eos_token_id": 50256,
-    "transformers_version": "4.26.0.dev0"
 }
 ```
 
@@ -84,7 +82,8 @@ Even if the default decoding strategy mostly works for your task, you can still
 commonly adjusted parameters include:
 
 - `max_new_tokens`: the maximum number of tokens to generate. In other words, the size of the output sequence, not
-including the tokens in the prompt.
+including the tokens in the prompt. As an alternative to using the output's length as a stopping criteria, you can choose 
+to stop generation whenever the full generation exceeds some amount of time. To learn more, check [`StoppingCriteria`].
 - `num_beams`: by specifying a number of beams higher than 1, you are effectively switching from greedy search to
 beam search. This strategy evaluates several hypotheses at each time step and eventually chooses the hypothesis that
 has the overall highest probability for the entire sequence. This has the advantage of identifying high-probability

docs/source/en/hpo_train.md

@@ -54,6 +54,18 @@ For optuna, see optuna [object_parameter](https://optuna.readthedocs.io/en/stabl
 ...     }
 ```
 
+Optuna provides multi-objective HPO. You can pass `direction` in `hyperparameter_search` and define your own compute_objective to return multiple objective values. The Pareto Front (`List[BestRun]`) will be returned in hyperparameter_search, you should refer to the test case `TrainerHyperParameterMultiObjectOptunaIntegrationTest` in [test_trainer](https://github.com/huggingface/transformers/blob/main/tests/trainer/test_trainer.py). It's like following
+
+```py
+>>> best_trials = trainer.hyperparameter_search(
+...     direction=["minimize", "maximize"],
+...     backend="optuna",
+...     hp_space=optuna_hp_space,
+...     n_trials=20,
+...     compute_objective=compute_objective,
+... )
+```
+
 For raytune, see raytune [object_parameter](https://docs.ray.io/en/latest/tune/api/search_space.html), it's like following:
 
 ```py

docs/source/en/index.md

@@ -76,6 +76,7 @@ The documentation is organized into five sections:
 1. **[BLOOM](model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](model_doc/bros)** (from NAVER CLOVA) released with the paper [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
 1. **[ByT5](model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
 1. **[CamemBERT](model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
@@ -139,6 +140,7 @@ The documentation is organized into five sections:
 1. **[GPTSAN-japanese](model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
 1. **[Graphormer](model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
+1. **[HerBERT](model_doc/herbert)** (from Allegro.pl, AGH University of Science and Technology) released with the paper [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) by Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.
 1. **[Hubert](model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
 1. **[IDEFICS](model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh.
@@ -172,6 +174,7 @@ The documentation is organized into five sections:
 1. **[Megatron-BERT](model_doc/megatron-bert)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
 1. **[Megatron-GPT2](model_doc/megatron_gpt2)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
 1. **[MGP-STR](model_doc/mgp-str)** (from Alibaba Research) released with the paper [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592) by Peng Wang, Cheng Da, and Cong Yao.
+1. **[Mistral](model_doc/mistral)** (from Mistral AI) by The [Mistral AI](https://mistral.ai) team: Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.
 1. **[mLUKE](model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
 1. **[MMS](model_doc/mms)** (from Facebook) released with the paper [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516) by Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli.
 1. **[MobileBERT](model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
@@ -189,6 +192,7 @@ The documentation is organized into five sections:
 1. **[Nezha](model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu.
 1. **[NLLB](model_doc/nllb)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
 1. **[NLLB-MOE](model_doc/nllb-moe)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
+1. **[Nougat](model_doc/nougat)** (from Meta AI) released with the paper [Nougat: Neural Optical Understanding for Academic Documents](https://arxiv.org/abs/2308.13418) by Lukas Blecher, Guillem Cucurull, Thomas Scialom, Robert Stojnic.
 1. **[Nyströmformer](model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
 1. **[OneFormer](model_doc/oneformer)** (from SHI Labs) released with the paper [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) by Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi.
 1. **[OpenLlama](model_doc/open-llama)** (from [s-JoL](https://huggingface.co/s-JoL)) released in [Open-Llama](https://github.com/s-JoL/Open-Llama).
@@ -197,6 +201,7 @@ The documentation is organized into five sections:
 1. **[Pegasus](model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[PEGASUS-X](model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, and Peter J. Liu.
 1. **[Perceiver IO](model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
+1. **[Persimmon](model_doc/persimmon)** (from ADEPT) released in a [blog post](https://www.adept.ai/blog/persimmon-8b) by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
 1. **[PhoBERT](model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
 1. **[Pix2Struct](model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
 1. **[PLBart](model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
@@ -254,6 +259,7 @@ The documentation is organized into five sections:
 1. **[ViT Hybrid](model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VitDet](model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
+1. **[ViTMatte](model_doc/vitmatte)** (from HUST-VL) rreleased with the paper [ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers](https://arxiv.org/abs/2305.15272) by Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang.
 1. **[ViTMSN](model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
 1. **[VITS](model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
@@ -307,6 +313,7 @@ Flax), PyTorch, and/or TensorFlow.
 |            BLIP-2             |       ✅        |         ❌         |      ❌      |
 |             BLOOM             |       ✅        |         ❌         |      ✅      |
 |          BridgeTower          |       ✅        |         ❌         |      ❌      |
+|             BROS              |       ✅        |         ❌         |      ❌      |
 |           CamemBERT           |       ✅        |         ✅         |      ❌      |
 |            CANINE             |       ✅        |         ❌         |      ❌      |
 |         Chinese-CLIP          |       ✅        |         ❌         |      ❌      |
@@ -388,11 +395,11 @@ Flax), PyTorch, and/or TensorFlow.
 |           MarkupLM            |       ✅        |         ❌         |      ❌      |
 |          Mask2Former          |       ✅        |         ❌         |      ❌      |
 |          MaskFormer           |       ✅        |         ❌         |      ❌      |
-|        MaskFormerSwin         |       ❌        |         ❌         |      ❌      |
 |             mBART             |       ✅        |         ✅         |      ✅      |
 |             MEGA              |       ✅        |         ❌         |      ❌      |
 |         Megatron-BERT         |       ✅        |         ❌         |      ❌      |
 |            MGP-STR            |       ✅        |         ❌         |      ❌      |
+|            Mistral            |       ✅        |         ❌         |      ❌      |
 |          MobileBERT           |       ✅        |         ✅         |      ❌      |
 |          MobileNetV1          |       ✅        |         ❌         |      ❌      |
 |          MobileNetV2          |       ✅        |         ❌         |      ❌      |
@@ -407,6 +414,7 @@ Flax), PyTorch, and/or TensorFlow.
 |              NAT              |       ✅        |         ❌         |      ❌      |
 |             Nezha             |       ✅        |         ❌         |      ❌      |
 |           NLLB-MOE            |       ✅        |         ❌         |      ❌      |
+|            Nougat             |       ✅        |         ✅         |      ✅      |
 |         Nyströmformer         |       ✅        |         ❌         |      ❌      |
 |           OneFormer           |       ✅        |         ❌         |      ❌      |
 |          OpenAI GPT           |       ✅        |         ✅         |      ❌      |
@@ -417,6 +425,7 @@ Flax), PyTorch, and/or TensorFlow.
 |            Pegasus            |       ✅        |         ✅         |      ✅      |
 |           PEGASUS-X           |       ✅        |         ❌         |      ❌      |
 |           Perceiver           |       ✅        |         ❌         |      ❌      |
+|           Persimmon           |       ✅        |         ❌         |      ❌      |
 |          Pix2Struct           |       ✅        |         ❌         |      ❌      |
 |            PLBart             |       ✅        |         ❌         |      ❌      |
 |          PoolFormer           |       ✅        |         ❌         |      ❌      |
@@ -456,7 +465,6 @@ Flax), PyTorch, and/or TensorFlow.
 |             TAPAS             |       ✅        |         ✅         |      ❌      |
 |    Time Series Transformer    |       ✅        |         ❌         |      ❌      |
 |          TimeSformer          |       ✅        |         ❌         |      ❌      |
-|         TimmBackbone          |       ❌        |         ❌         |      ❌      |
 |    Trajectory Transformer     |       ✅        |         ❌         |      ❌      |
 |        Transformer-XL         |       ✅        |         ✅         |      ❌      |
 |             TrOCR             |       ✅        |         ❌         |      ❌      |
@@ -475,6 +483,7 @@ Flax), PyTorch, and/or TensorFlow.
 |          ViT Hybrid           |       ✅        |         ❌         |      ❌      |
 |            VitDet             |       ✅        |         ❌         |      ❌      |
 |            ViTMAE             |       ✅        |         ✅         |      ❌      |
+|           ViTMatte            |       ✅        |         ❌         |      ❌      |
 |            ViTMSN             |       ✅        |         ❌         |      ❌      |
 |             VITS              |       ✅        |         ❌         |      ❌      |
 |             ViViT             |       ✅        |         ❌         |      ❌      |

docs/source/en/installation.md

@@ -169,28 +169,28 @@ Pretrained models are downloaded and locally cached at: `~/.cache/huggingface/hu
 
 ## Offline mode
 
-🤗 Transformers is able to run in a firewalled or offline environment by only using local files. Set the environment variable `TRANSFORMERS_OFFLINE=1` to enable this behavior.
+Run 🤗 Transformers in a firewalled or offline environment with locally cached files by setting the environment variable `TRANSFORMERS_OFFLINE=1`.
 
 <Tip>
 
-Add [🤗 Datasets](https://huggingface.co/docs/datasets/) to your offline training workflow by setting the environment variable `HF_DATASETS_OFFLINE=1`.
+Add [🤗 Datasets](https://huggingface.co/docs/datasets/) to your offline training workflow with the environment variable `HF_DATASETS_OFFLINE=1`.
 
 </Tip>
 
-For example, you would typically run a program on a normal network firewalled to external instances with the following command:
-
-```bash
-python examples/pytorch/translation/run_translation.py --model_name_or_path t5-small --dataset_name wmt16 --dataset_config ro-en ...
-```
-
-Run this same program in an offline instance with:
-
 ```bash
 HF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1 \
 python examples/pytorch/translation/run_translation.py --model_name_or_path t5-small --dataset_name wmt16 --dataset_config ro-en ...
 ```
 
-The script should now run without hanging or waiting to timeout because it knows it should only look for local files.
+This script should run without hanging or waiting to timeout because it won't attempt to download the model from the Hub.
+
+You can also bypass loading a model from the Hub from each [`~PreTrainedModel.from_pretrained`] call with the [`local_files_only`] parameter. When set to `True`, only local files are loaded:
+
+```py
+from transformers import T5Model
+
+model = T5Model.from_pretrained("./path/to/local/directory", local_files_only=True)
+```
 
 ### Fetch models and tokenizers to use offline
 

docs/source/en/main_classes/agent.md

@@ -18,7 +18,7 @@ rendered properly in your Markdown viewer.
 
 <Tip warning={true}>
 
-Transformers Agent is an experimental API which is subject to change at any time. Results returned by the agents
+Transformers Agents is an experimental API which is subject to change at any time. Results returned by the agents
 can vary as the APIs or underlying models are prone to change.
 
 </Tip>

docs/source/en/main_classes/output.md

@@ -40,6 +40,13 @@ an optional `attentions` attribute. Here we have the `loss` since we passed alon
 `hidden_states` and `attentions` because we didn't pass `output_hidden_states=True` or
 `output_attentions=True`.
 
+<Tip>
+
+When passing `output_hidden_states=True` you may expect the `outputs.hidden_states[-1]` to match `outputs.last_hidden_states` exactly.
+However, this is not always the case. Some models apply normalization or subsequent process to the last hidden state when it's returned.
+</Tip>
+
+
 You can access each attribute as you would usually do, and if that attribute has not been returned by the model, you
 will get `None`. Here for instance `outputs.loss` is the loss computed by the model, and `outputs.attentions` is
 `None`.

docs/source/en/main_classes/pipelines.md

@@ -352,6 +352,12 @@ Pipelines available for computer vision tasks include the following.
     - __call__
     - all
 
+### ImageToImagePipeline
+
+[[autodoc]] ImageToImagePipeline
+    - __call__
+    - all
+
 ### ObjectDetectionPipeline
 
 [[autodoc]] ObjectDetectionPipeline

docs/source/en/main_classes/tokenizer.md

@@ -55,6 +55,7 @@ to a given token).
 
 [[autodoc]] PreTrainedTokenizer
     - __call__
+    - apply_chat_template
     - batch_decode
     - decode
     - encode
@@ -68,6 +69,7 @@ loaded very simply into 🤗 transformers. Take a look at the [Using tokenizers
 
 [[autodoc]] PreTrainedTokenizerFast
     - __call__
+    - apply_chat_template
     - batch_decode
     - decode
     - encode

docs/source/en/model_doc/auto.md

@@ -266,6 +266,10 @@ The following auto classes are available for the following computer vision tasks
 
 [[autodoc]] AutoModelForImageSegmentation
 
+### AutoModelForImageToImage
+
+[[autodoc]] AutoModelForImageToImage
+
 ### AutoModelForSemanticSegmentation
 
 [[autodoc]] AutoModelForSemanticSegmentation

docs/source/en/model_doc/bros.md

@@ -0,0 +1,115 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# BROS
+
+## Overview
+
+The BROS model was proposed in [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
+
+BROS stands for *BERT Relying On Spatiality*. It is an encoder-only Transformer model that takes a sequence of tokens and their bounding boxes as inputs and outputs a sequence of hidden states. BROS encode relative spatial information instead of using absolute spatial information.
+
+It is pre-trained with two objectives: a token-masked language modeling objective (TMLM) used in BERT, and a novel area-masked language modeling objective (AMLM)
+In TMLM, tokens are randomly masked, and the model predicts the masked tokens using spatial information and other unmasked tokens.
+AMLM is a 2D version of TMLM. It randomly masks text tokens and predicts with the same information as TMLM, but it masks text blocks (areas).
+
+`BrosForTokenClassification` has a simple linear layer on top of BrosModel. It predicts the label of each token.
+`BrosSpadeEEForTokenClassification` has an `initial_token_classifier` and `subsequent_token_classifier` on top of BrosModel. `initial_token_classifier` is used to predict the first token of each entity, and `subsequent_token_classifier` is used to predict the next token of within entity. `BrosSpadeELForTokenClassification` has an `entity_linker` on top of BrosModel. `entity_linker` is used to predict the relation between two entities.
+
+`BrosForTokenClassification` and `BrosSpadeEEForTokenClassification` essentially perform the same job. However, `BrosForTokenClassification` assumes input tokens are perfectly serialized (which is very challenging task since they exist in a 2D space), while `BrosSpadeEEForTokenClassification` allows for more flexibility in handling serialization errors as it predicts next connection tokens from one token.
+
+`BrosSpadeELForTokenClassification` perform the intra-entity linking task. It predicts relation from one token (of one entity) to another token (of another entity) if these two entities share some relation.
+
+BROS achieves comparable or better result on Key Information Extraction (KIE) benchmarks such as FUNSD, SROIE, CORD and SciTSR, without relying on explicit visual features.
+
+
+The abstract from the paper is the following:
+
+*Key information extraction (KIE) from document images requires understanding the contextual and spatial semantics of texts in two-dimensional (2D) space. Many recent studies try to solve the task by developing pre-trained language models focusing on combining visual features from document images with texts and their layout. On the other hand, this paper tackles the problem by going back to the basic: effective combination of text and layout. Specifically, we propose a pre-trained language model, named BROS (BERT Relying On Spatiality), that encodes relative positions of texts in 2D space and learns from unlabeled documents with area-masking strategy. With this optimized training scheme for understanding texts in 2D space, BROS shows comparable or better performance compared to previous methods on four KIE benchmarks (FUNSD, SROIE*, CORD, and SciTSR) without relying on visual features. This paper also reveals two real-world challenges in KIE tasks-(1) minimizing the error from incorrect text ordering and (2) efficient learning from fewer downstream examples-and demonstrates the superiority of BROS over previous methods.*
+
+Tips:
+
+- [`~transformers.BrosModel.forward`] requires `input_ids` and `bbox` (bounding box). Each bounding box should be in (x0, y0, x1, y1) format (top-left corner, bottom-right corner). Obtaining of Bounding boxes depends on external OCR system. The `x` coordinate should be normalized by document image width, and the `y` coordinate should be normalized by document image height.
+
+```python
+def expand_and_normalize_bbox(bboxes, doc_width, doc_height):
+    # here, bboxes are numpy array
+
+    # Normalize bbox -> 0 ~ 1
+    bboxes[:, [0, 2]] = bboxes[:, [0, 2]] / width
+    bboxes[:, [1, 3]] = bboxes[:, [1, 3]] / height
+```
+
+- [`~transformers.BrosForTokenClassification.forward`, `~transformers.BrosSpadeEEForTokenClassification.forward`, `~transformers.BrosSpadeEEForTokenClassification.forward`] require not only `input_ids` and `bbox` but also `box_first_token_mask` for loss calculation. It is a mask to filter out non-first tokens of each box. You can obtain this mask by saving start token indices of bounding boxes when creating `input_ids` from words. You can make `box_first_token_mask` with following code,
+
+
+```python
+def make_box_first_token_mask(bboxes, words, tokenizer, max_seq_length=512):
+
+    box_first_token_mask = np.zeros(max_seq_length, dtype=np.bool_)
+
+    # encode(tokenize) each word from words (List[str])
+    input_ids_list: List[List[int]] = [tokenizer.encode(e, add_special_tokens=False) for e in words]
+
+    # get the length of each box
+    tokens_length_list: List[int] = [len(l) for l in input_ids_list]
+
+    box_end_token_indices = np.array(list(itertools.accumulate(tokens_length_list)))
+    box_start_token_indices = box_end_token_indices - np.array(tokens_length_list)
+
+    # filter out the indices that are out of max_seq_length
+    box_end_token_indices = box_end_token_indices[box_end_token_indices < max_seq_length - 1]
+    if len(box_start_token_indices) > len(box_end_token_indices):
+        box_start_token_indices = box_start_token_indices[: len(box_end_token_indices)]
+
+    # set box_start_token_indices to True
+    box_first_token_mask[box_start_token_indices] = True
+
+    return box_first_token_mask
+
+```
+
+- Demo scripts can be found [here](https://github.com/clovaai/bros).
+
+This model was contributed by [jinho8345](https://huggingface.co/jinho8345). The original code can be found [here](https://github.com/clovaai/bros).
+
+## BrosConfig
+
+[[autodoc]] BrosConfig
+
+## BrosProcessor
+
+[[autodoc]] BrosProcessor
+    - __call__
+
+## BrosModel
+
+[[autodoc]] BrosModel
+    - forward
+
+
+## BrosForTokenClassification
+
+[[autodoc]] BrosForTokenClassification
+    - forward
+
+
+## BrosSpadeEEForTokenClassification
+
+[[autodoc]] BrosSpadeEEForTokenClassification
+    - forward
+
+
+## BrosSpadeELForTokenClassification
+
+[[autodoc]] BrosSpadeELForTokenClassification
+    - forward

docs/source/en/model_doc/deberta-v2.md

@@ -152,3 +152,8 @@ contributed by [kamalkraj](https://huggingface.co/kamalkraj). The original code
 
 [[autodoc]] TFDebertaV2ForQuestionAnswering
     - call
+
+## TFDebertaV2ForMultipleChoice
+
+[[autodoc]] TFDebertaV2ForMultipleChoice
+    - call

docs/source/en/model_doc/llama.md

@@ -55,6 +55,28 @@ Based on the original LLaMA model, Meta AI has released some follow-up works:
 
 - **Llama2**: Llama2 is an improved version of Llama with some architectural tweaks (Grouped Query Attention), and is pre-trained on 2Trillion tokens. Refer to the documentation of Llama2 which can be found [here](llama2).
 
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with LLaMA. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
+
+<PipelineTag pipeline="text-classification"/>
+
+- A [notebook](https://colab.research.google.com/github/bigscience-workshop/petals/blob/main/examples/prompt-tuning-sst2.ipynb#scrollTo=f04ba4d2) on how to use prompt tuning to adapt the LLaMA model for text classification task. 🌎
+
+<PipelineTag pipeline="question-answering"/>
+
+- [StackLLaMA: A hands-on guide to train LLaMA with RLHF](https://huggingface.co/blog/stackllama#stackllama-a-hands-on-guide-to-train-llama-with-rlhf), a blog post about how to train LLaMA to answer questions on [Stack Exchange](https://stackexchange.com/) with RLHF.
+
+⚗️ Optimization
+- A [notebook](https://colab.research.google.com/drive/1SQUXq1AMZPSLD4mk3A3swUIc6Y2dclme?usp=sharing) on how to fine-tune LLaMA model using xturing library on GPU which has limited memory. 🌎 
+
+⚡️ Inference
+- A [notebook](https://colab.research.google.com/github/DominguesM/alpaca-lora-ptbr-7b/blob/main/notebooks/02%20-%20Evaluate.ipynb) on how to run the LLaMA Model using PeftModel from the 🤗 PEFT library. 🌎 
+- A [notebook](https://colab.research.google.com/drive/1l2GiSSPbajVyp2Nk3CFT4t3uH6-5TiBe?usp=sharing) on how to load a PEFT adapter LLaMA model with LangChain. 🌎
+
+🚀 Deploy
+- A [notebook](https://colab.research.google.com/github/lxe/simple-llama-finetuner/blob/master/Simple_LLaMA_FineTuner.ipynb#scrollTo=3PM_DilAZD8T) on how to fine-tune LLaMA model using LoRA method via the 🤗 PEFT library with intuitive UI. 🌎 
+- A [notebook](https://github.com/aws/amazon-sagemaker-examples/blob/main/introduction_to_amazon_algorithms/jumpstart-foundation-models/text-generation-open-llama.ipynb) on how to deploy Open-LLaMA model for text generation on Amazon SageMaker. 🌎 
 
 ## LlamaConfig
 

docs/source/en/model_doc/llama2.md

@@ -78,6 +78,10 @@ A list of official Hugging Face and community (indicated by 🌎) resources to h
 - A [notebook](https://colab.research.google.com/drive/1PEQyJO1-f6j0S_XJ8DV50NkpzasXkrzd?usp=sharing) on how to fine-tune Llama 2 in Google Colab using QLoRA and 4-bit precision. 🌎
 - A [notebook](https://colab.research.google.com/drive/134o_cXcMe_lsvl15ZE_4Y75Kstepsntu?usp=sharing) on how to fine-tune the "Llama-v2-7b-guanaco" model with 4-bit QLoRA and generate Q&A datasets from PDFs. 🌎
 
+<PipelineTag pipeline="text-classification"/>
+
+- A [notebook](https://colab.research.google.com/drive/1ggaa2oRFphdBmqIjSEbnb_HGkcIRC2ZB?usp=sharing) on how to fine-tune the Llama 2 model with QLoRa, TRL, and Korean text classification dataset. 🌎🇰🇷
+
 ⚗️ Optimization
 - [Fine-tune Llama 2 with DPO](https://huggingface.co/blog/dpo-trl), a guide to using the TRL library's DPO method to fine tune Llama 2 on a specific dataset.
 - [Extended Guide: Instruction-tune Llama 2](https://www.philschmid.de/instruction-tune-llama-2), a guide to training Llama 2 to generate instructions from inputs, transforming the model from instruction-following to instruction-giving.

docs/source/en/model_doc/mistral.md

@@ -0,0 +1,151 @@
+<!--Copyright 2023 Mistral AI and 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.
+
+-->
+
+# Mistral
+
+## Overview
+
+Mistral-7B-v0.1 is Mistral AI’s first Large Language Model (LLM). 
+
+## Model Details
+
+Mistral-7B-v0.1 is a decoder-based LM with the following architectural choices:
+* Sliding Window Attention - Trained with 8k context length and fixed cache size, with a theoretical attention span of 128K tokens
+* GQA (Grouped Query Attention) - allowing faster inference and lower cache size.
+* Byte-fallback BPE tokenizer - ensures that characters are never mapped to out of vocabulary tokens.
+
+We also provide an instruction fine-tuned model: `Mistral-7B-Instruct-v0.1` which can be used for chat-based inference.
+
+For more details please read our [release blog post](https://mistral.ai/news/announcing-mistral-7b-v0.1/)
+
+## License
+
+Both `Mistral-7B-v0.1` and `Mistral-7B-Instruct-v0.1` are released under the Apache 2.0 license.
+
+## Usage
+
+`Mistral-7B-v0.1` and `Mistral-7B-Instruct-v0.1` can be found on the [Huggingface Hub](https://huggingface.co/mistralai)
+
+These ready-to-use checkpoints can be downloaded and used via the HuggingFace Hub:
+
+```python
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
+>>> device = "cuda" # the device to load the model onto
+
+>>> model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
+>>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
+
+>>> prompt = "My favourite condiment is"
+
+>>> model_inputs = tokenizer([prompt], return_tensors="pt").to(device)
+>>> model.to(device)
+
+>>> generated_ids = model.generate(**model_inputs, max_new_tokens=100, do_sample=True)
+>>> tokenizer.batch_decode(generated_ids)[0]
+"The expected outupt"
+```
+
+Raw weights for `Mistral-7B-v0.1` and `Mistral-7B-Instruct-v0.1` can be downloaded from:
+
+| Model Name                 | Checkpoint                                                                              |
+|----------------------------|-----------------------------------------------------------------------------------------|
+| `Mistral-7B-v0.1`          | [Raw Checkpoint](https://files.mistral-7b-v0-1.mistral.ai/mistral-7B-v0.1.tar)          |
+| `Mistral-7B-Instruct-v0.1` | [Raw Checkpoint](https://files.mistral-7b-v0-1.mistral.ai/mistral-7B-instruct-v0.1.tar) |
+
+
+To use these raw checkpoints with HuggingFace you can use the `convert_mistral_weights_to_hf.py` script to convert them to the HuggingFace format:
+
+```bash
+python src/transformers/models/mistral/convert_mistral_weights_to_hf.py \
+    --input_dir /path/to/downloaded/mistral/weights --model_size 7B --output_dir /output/path
+```
+
+You can then load the converted model from the `output/path`:
+
+```python
+from transformers import MistralForCausalLM, LlamaTokenzier
+
+tokenizer = LlamaTokenizer.from_pretrained("/output/path")
+model = MistralForCausalLM.from_pretrained("/output/path")
+```
+
+## Combining Mistral and Flash Attention 2
+
+First, make sure to install the latest version of Flash Attention 2 to include the sliding window attention feature.
+
+```bash
+pip install -U flash-attn --no-build-isolation
+```
+
+Make also sure that you have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of [`flash-attn`](https://github.com/Dao-AILab/flash-attention) repository. Make also sure to load your model in half-precision (e.g. `torch.float16`)
+
+To load and run a model using Flash Attention 2, refer to the snippet below:
+
+```python
+>>> import torch
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
+>>> device = "cuda" # the device to load the model onto
+
+>>> model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1", torch_dtype=torch.float16, use_flash_attention_2=True)
+>>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
+
+>>> prompt = "My favourite condiment is"
+
+>>> model_inputs = tokenizer([prompt], return_tensors="pt").to(device)
+>>> model.to(device)
+
+>>> generated_ids = model.generate(**model_inputs, max_new_tokens=100, do_sample=True)
+>>> tokenizer.batch_decode(generated_ids)[0]
+"The expected outupt"
+```
+
+### Expected speedups
+
+Below is a expected speedup diagram that compares pure inference time between the native implementation in transformers using `mistralai/Mistral-7B-v0.1` checkpoint and the Flash Attention 2 version of the model.
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/mistral-7b-inference-large-seqlen.png">
+</div>
+
+### Sliding window Attention
+
+The current implementation supports the sliding window attention mechanism and memory efficient cache management. 
+To enable sliding window attention, just make sure to have a `flash-attn` version that is compatible with sliding window attention (`>=2.3.0`). 
+
+The Flash Attention-2 model uses also a more memory efficient cache slicing mechanism - as recommended per the official implementation of Mistral model that use rolling cache mechanism we keep the cache size fixed (`self.config.sliding_window`), support batched generation only for `padding_side="left"` and use the absolute position of the current token to compute the positional embedding.
+
+## The Mistral Team
+
+Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.
+
+## MistralConfig
+
+[[autodoc]] MistralConfig
+
+## MistralModel
+
+[[autodoc]] MistralModel
+    - forward
+
+## MistralForCausalLM
+
+[[autodoc]] MistralForCausalLM
+    - forward
+
+## MistralForSequenceClassification
+
+[[autodoc]] MistralForSequenceClassification
+    - forward

docs/source/en/model_doc/nougat.md

@@ -0,0 +1,109 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the
+License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an
+"AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+
+⚠️ 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.
+
+specific language governing permissions and limitations under the License. -->
+
+# Nougat
+
+## Overview
+
+The Nougat model was proposed in [Nougat: Neural Optical Understanding for Academic Documents](https://arxiv.org/abs/2308.13418) by
+Lukas Blecher, Guillem Cucurull, Thomas Scialom, Robert Stojnic. Nougat uses the same architecture as [Donut](donut), meaning an image Transformer
+encoder and an autoregressive text Transformer decoder to translate scientific PDFs to markdown, enabling easier access to them.
+
+The abstract from the paper is the following:
+
+*Scientific knowledge is predominantly stored in books and scientific journals, often in the form of PDFs. However, the PDF format leads to a loss of semantic information, particularly for mathematical expressions. We propose Nougat (Neural Optical Understanding for Academic Documents), a Visual Transformer model that performs an Optical Character Recognition (OCR) task for processing scientific documents into a markup language, and demonstrate the effectiveness of our model on a new dataset of scientific documents. The proposed approach offers a promising solution to enhance the accessibility of scientific knowledge in the digital age, by bridging the gap between human-readable documents and machine-readable text. We release the models and code to accelerate future work on scientific text recognition.*
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/nougat_architecture.jpg"
+alt="drawing" width="600"/>
+
+<small> Nougat high-level overview. Taken from the <a href="https://arxiv.org/abs/2308.13418">original paper</a>. </small>
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr). The original code can be found
+[here](https://github.com/facebookresearch/nougat).
+
+Tips:
+
+- The quickest way to get started with Nougat is by checking the [tutorial
+  notebooks](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/Nougat), which show how to use the model
+  at inference time as well as fine-tuning on custom data.
+- Nougat is always used within the [VisionEncoderDecoder](vision-encoder-decoder) framework. The model is identical to [Donut](donut) in terms of architecture.
+
+## Inference
+
+Nougat's [`VisionEncoderDecoder`] model accepts images as input and makes use of
+[`~generation.GenerationMixin.generate`] to autoregressively generate text given the input image.
+
+The [`NougatImageProcessor`] class is responsible for preprocessing the input image and
+[`NougatTokenizerFast`] decodes the generated target tokens to the target string. The
+[`NougatProcessor`] wraps [`NougatImageProcessor`] and [`NougatTokenizerFast`] classes
+into a single instance to both extract the input features and decode the predicted token ids.
+
+- Step-by-step PDF transcription
+
+```py
+>>> from huggingface_hub import hf_hub_download
+>>> import re
+>>> from PIL import Image
+
+>>> from transformers import NougatProcessor, VisionEncoderDecoderModel
+>>> from datasets import load_dataset
+>>> import torch
+
+>>> processor = NougatProcessor.from_pretrained("facebook/nougat-base")
+>>> model = VisionEncoderDecoderModel.from_pretrained("facebook/nougat-base")
+
+>>> device = "cuda" if torch.cuda.is_available() else "cpu"
+>>> model.to(device)  # doctest: +IGNORE_RESULT
+
+>>> # prepare PDF image for the model
+>>> filepath = hf_hub_download(repo_id="hf-internal-testing/fixtures_docvqa", filename="nougat_paper.png", repo_type="dataset")
+>>> image = Image.open(filepath)
+>>> pixel_values = processor(image, return_tensors="pt").pixel_values
+
+>>> # generate transcription (here we only generate 30 tokens)
+>>> outputs = model.generate(
+...     pixel_values.to(device),
+...     min_length=1,
+...     max_new_tokens=30,
+...     bad_words_ids=[[processor.tokenizer.unk_token_id]],
+... )
+
+>>> sequence = processor.batch_decode(outputs, skip_special_tokens=True)[0]
+>>> sequence = processor.post_process_generation(sequence, fix_markdown=False)
+>>> # note: we're using repr here such for the sake of printing the \n characters, feel free to just print the sequence
+>>> print(repr(sequence))
+'\n\n# Nougat: Neural Optical Understanding for Academic Documents\n\n Lukas Blecher\n\nCorrespondence to: lblecher@'
+```
+
+See the [model hub](https://huggingface.co/models?filter=nougat) to look for Nougat checkpoints.
+
+## NougatImageProcessor
+
+[[autodoc]] NougatImageProcessor
+    - preprocess
+
+## NougatTokenizerFast
+
+[[autodoc]] NougatTokenizerFast
+
+## NougatProcessor
+
+[[autodoc]] NougatProcessor
+    - __call__
+    - from_pretrained
+    - save_pretrained
+    - batch_decode
+    - decode
+    - post_process_generation

docs/source/en/model_doc/persimmon.md

@@ -0,0 +1,96 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Persimmon
+
+## Overview
+
+The Persimmon model was created by [ADEPT](https://www.adept.ai/blog/persimmon-8b), and authored by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
+
+The authors introduced Persimmon-8B, a decoder model based on the classic transformers architecture, with query and key normalization. Persimmon-8B is a fully permissively-licensed model with approximately 8 billion parameters, released under the Apache license.  Some of the key attributes of Persimmon-8B are long context size (16K), performance, and capabilities for multimodal extensions.
+
+The authors showcase their approach to model evaluation, focusing on practical text generation, mirroring how users interact with language models. The work also includes a comparative analysis, pitting Persimmon-8B against other prominent models (MPT 7B Instruct and Llama 2 Base 7B 1-Shot), across various evaluation tasks. The results demonstrate Persimmon-8B's competitive performance, even with limited training data.
+
+In terms of model details, the work outlines the architecture and training methodology of Persimmon-8B, providing insights into its design choices, sequence length, and dataset composition. The authors present a fast inference code that outperforms traditional implementations through operator fusion and CUDA graph utilization while maintaining code coherence. They express their anticipation of how the community will leverage this contribution to drive innovation, hinting at further upcoming releases as part of an ongoing series of developments.
+
+
+<Tip warning={true}>
+
+The `Persimmon` models were trained using `bfloat16`, but the original inference uses `float16` The checkpoints uploaded on the hub use `torch_dtype = 'float16'` which will be
+used by the `AutoModel` API to cast the checkpoints from `torch.float32` to `torch.float16`. 
+
+The `dtype` of the online weights is mostly irrelevant, unless you are using `torch_dtype="auto"` when initializing a model using `model = AutoModelForCausalLM.from_pretrained("path", torch_dtype = "auto")`. The reason is that the model will first be downloaded ( using the `dtype` of the checkpoints online) then it will be cast to the default `dtype` of `torch` (becomes `torch.float32`). Users should specify the `torch_dtype` they want, and if they don't it will be `torch.float32`.
+
+Finetuning the model in `float16` is not recommended and known to produce `nan`, as such the model should be fine-tuned in `bfloat16`.
+
+</Tip>
+
+
+Tips:
+
+- To convert the model, you need to clone the original repository using `git clone https://github.com/persimmon-ai-labs/adept-inference`, then get the checkpoints:
+
+```bash
+git clone https://github.com/persimmon-ai-labs/adept-inference
+wget https://axtkn4xl5cip.objectstorage.us-phoenix-1.oci.customer-oci.com/n/axtkn4xl5cip/b/adept-public-data/o/8b_base_model_release.tar
+tar -xvf 8b_base_model_release.tar
+python src/transformers/models/persimmon/convert_persimmon_weights_to_hf.py  --input_dir /path/to/downloaded/persimmon/weights/ --output_dir /output/path \
+    --pt_model_path /path/to/8b_chat_model_release/iter_0001251/mp_rank_00/model_optim_rng.pt
+    --ada_lib_path /path/to/adept-inference
+```
+
+For the chat model:
+```bash
+wget https://axtkn4xl5cip.objectstorage.us-phoenix-1.oci.customer-oci.com/n/axtkn4xl5cip/b/adept-public-data/o/8b_chat_model_release.tar
+tar -xvf 8b_base_model_release.tar
+```
+
+Thereafter, models can be loaded via:
+
+```py
+from transformers import PersimmonForCausalLM, PersimmonTokenizer
+
+model = PersimmonForCausalLM.from_pretrained("/output/path")
+tokenizer = PersimmonTokenizer.from_pretrained("/output/path")
+```
+
+This model was contributed by [ArthurZ](https://huggingface.co/ArthurZ).
+The original code can be found [here](https://github.com/persimmon-ai-labs/adept-inference).
+
+- Perismmon uses a `sentencepiece` based tokenizer, with a `Unigram` model. It supports bytefallback, which is only available in `tokenizers==0.14.0` for the fast tokenizer.
+The `LlamaTokenizer` is used as it is a standard wrapper around sentencepiece. The `chat` template will be updated with the templating functions in a follow up PR!
+
+- The authors suggest to use the following prompt format for the chat mode: `f"human: {prompt}\n\nadept:"`
+
+
+## PersimmonConfig
+
+[[autodoc]] PersimmonConfig
+
+## PersimmonModel
+
+[[autodoc]] PersimmonModel
+    - forward
+
+## PersimmonForCausalLM
+
+[[autodoc]] PersimmonForCausalLM
+    - forward
+
+## PersimmonForSequenceClassification
+
+[[autodoc]] PersimmonForSequenceClassification
+    - forward

docs/source/en/model_doc/pop2piano.md

@@ -12,6 +12,12 @@ specific language governing permissions and limitations under the License.
 
 # Pop2Piano
 
+<div class="flex flex-wrap space-x-1">
+<a href="https://huggingface.co/spaces/sweetcocoa/pop2piano">
+<img alt="Spaces" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue">
+</a>
+</div>
+
 ## Overview
 
 The Pop2Piano model was proposed in [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi and Kyogu Lee.

docs/source/en/model_doc/videomae.md

@@ -47,7 +47,7 @@ review it! The resource should ideally demonstrate something new instead of dupl
 **Video classification**
 - [A notebook](https://github.com/huggingface/notebooks/blob/main/examples/video_classification.ipynb) that shows how
 to fine-tune a VideoMAE model on a custom dataset.
-- [Video classification task guide](../tasks/video-classification)
+- [Video classification task guide](../tasks/video_classification)
 - [A 🤗 Space](https://huggingface.co/spaces/sayakpaul/video-classification-ucf101-subset) showing how to perform inference with a video classification model.
 
 

docs/source/en/model_doc/vitmatte.md

@@ -0,0 +1,55 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# ViTMatte
+
+## Overview
+
+The ViTMatte model was proposed in [Boosting Image Matting with Pretrained Plain Vision Transformers](https://arxiv.org/abs/2305.15272) by Jingfeng Yao, Xinggang Wang, Shusheng Yang, Baoyuan Wang.
+ViTMatte leverages plain [Vision Transformers](vit) for the task of image matting, which is the process of accurately estimating the foreground object in images and videos.
+
+The abstract from the paper is the following:
+
+*Recently, plain vision Transformers (ViTs) have shown impressive performance on various computer vision tasks, thanks to their strong modeling capacity and large-scale pretraining. However, they have not yet conquered the problem of image matting. We hypothesize that image matting could also be boosted by ViTs and present a new efficient and robust ViT-based matting system, named ViTMatte. Our method utilizes (i) a hybrid attention mechanism combined with a convolution neck to help ViTs achieve an excellent performance-computation trade-off in matting tasks. (ii) Additionally, we introduce the detail capture module, which just consists of simple lightweight convolutions to complement the detailed information required by matting. To the best of our knowledge, ViTMatte is the first work to unleash the potential of ViT on image matting with concise adaptation. It inherits many superior properties from ViT to matting, including various pretraining strategies, concise architecture design, and flexible inference strategies. We evaluate ViTMatte on Composition-1k and Distinctions-646, the most commonly used benchmark for image matting, our method achieves state-of-the-art performance and outperforms prior matting works by a large margin.*
+
+Tips:
+
+- The model expects both the image and trimap (concatenated) as input. One can use [`ViTMatteImageProcessor`] for this purpose.
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr).
+The original code can be found [here](https://github.com/hustvl/ViTMatte).
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/vitmatte_architecture.png"
+alt="drawing" width="600"/>
+
+<small> ViTMatte high-level overview. Taken from the <a href="https://arxiv.org/abs/2305.15272">original paper.</a> </small>
+
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with ViTMatte.
+
+- A demo notebook regarding inference with [`VitMatteForImageMatting`], including background replacement, can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/ViTMatte).
+
+
+## VitMatteConfig
+
+[[autodoc]] VitMatteConfig
+
+## VitMatteImageProcessor
+
+[[autodoc]] VitMatteImageProcessor
+    - preprocess
+
+## VitMatteForImageMatting
+
+[[autodoc]] VitMatteForImageMatting
+    - forward

docs/source/en/perf_infer_gpu_many.md

@@ -22,6 +22,10 @@ Note: A multi GPU setup can use the majority of the strategies described in the
 
 </Tip>
 
+## Flash Attention 2
+
+Flash Attention 2 integration also works in a multi-GPU setup, check out the appropriate section in the [single GPU section](./perf_infer_gpu_one#Flash-Attention-2)
+
 ## BetterTransformer
 
 [BetterTransformer](https://huggingface.co/docs/optimum/bettertransformer/overview) converts 🤗 Transformers models to use the PyTorch-native fastpath execution, which calls optimized kernels like Flash Attention under the hood.  

docs/source/en/perf_infer_gpu_one.md

@@ -17,6 +17,153 @@ rendered properly in your Markdown viewer.
 
 In addition to this guide, relevant information can be found as well in [the guide for training on a single GPU](perf_train_gpu_one) and [the guide for inference on CPUs](perf_infer_cpu).
 
+## Flash Attention 2
+
+<Tip>
+
+Note that this feature is experimental and might considerably change in future versions. For instance, the Flash Attention 2 API might migrate to `BetterTransformer` API in the near future.
+
+</Tip>
+
+Flash Attention 2 can considerably speed up transformer-based models' training and inference speed. Flash Attention 2 has been introduced in the [official Flash Attention repository](https://github.com/Dao-AILab/flash-attention) by Tri Dao et al. The scientific paper on Flash Attention can be found [here](https://arxiv.org/abs/2205.14135).
+
+Make sure to follow the installation guide on the repository mentioned above to properly install Flash Attention 2. Once that package is installed, you can benefit from this feature.
+
+We natively support Flash Attention 2 for the following models:
+
+- Llama
+- Mistral
+- Falcon
+
+You can request to add Flash Attention 2 support for more models by opening an issue on GitHub, and even open a Pull Request to integrate the changes. The supported models can be used for inference and training, including training with padding tokens - *which is currently not supported for `BetterTransformer` API below.*
+
+<Tip>
+
+Flash Attention 2 can only be used when the models' dtype is  `fp16` or `bf16` and runs only on NVIDIA-GPU devices. Make sure to cast your model to the appropriate dtype and load them on a supported device before using that feature.
+  
+</Tip>
+
+### Quick usage
+
+To enable Flash Attention 2 in your model, add `use_flash_attention_2` in the `from_pretrained` arguments:
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, LlamaForCausalLM
+
+model_id = "tiiuae/falcon-7b"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_id, 
+    torch_dtype=torch.bfloat16, 
+    use_flash_attention_2=True,
+)
+```
+
+And use it for generation or fine-tuning.
+
+### Expected speedups
+
+You can benefit from considerable speedups for fine-tuning and inference, especially for long sequences. However, since Flash Attention does not support computing attention scores with padding tokens under the hood, we must manually pad / unpad the attention scores for batched inference when the sequence contains padding tokens. This leads to a significant slowdown for batched generations with padding tokens. 
+
+To overcome this, one should use Flash Attention without padding tokens in the sequence for training (e.g., by packing a dataset, i.e., concatenating sequences until reaching the maximum sequence length. An example is provided [here](https://github.com/huggingface/transformers/blob/main/examples/pytorch/language-modeling/run_clm.py#L516).
+
+Below is the expected speedup you can get for a simple forward pass on [tiiuae/falcon-7b](https://hf.co/tiiuae/falcon-7b) with a sequence length of 4096 and various batch sizes, without padding tokens:
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/falcon-7b-inference-large-seqlen.png">
+</div>
+
+Below is the expected speedup you can get for a simple forward pass on [`meta-llama/Llama-7b-hf`](https://hf.co/meta-llama/Llama-7b-hf) with a sequence length of 4096 and various batch sizes, without padding tokens:
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/llama-7b-inference-large-seqlen.png">
+</div>
+
+For sequences with padding tokens (training with padding tokens or generating with padding tokens), we need to unpad / pad the input sequences to compute correctly the attention scores. For relatively small sequence length, on pure forward pass, this creates an overhead leading to a small speedup (below 30% of the input has been filled with padding tokens). 
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/llama-2-small-seqlen-padding.png">
+</div>
+
+But for large sequence length you can benefit from interesting speedup for pure inference (also training)
+
+Note that Flash Attention makes the attention computation more memory efficient, meaning you can train with much larger sequence lengths without facing CUDA OOM issues. It can lead up to memory reduction up to 20 for large sequence length. Check out [the official flash attention repository](https://github.com/Dao-AILab/flash-attention) for more details.
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/llama-2-large-seqlen-padding.png">
+</div>
+
+
+### Advanced usage
+
+You can combine this feature with many exisiting feature for model optimization. Check out few examples below:
+
+### Combining Flash Attention 2 and 8-bit models
+
+You can combine this feature together with 8-bit quantization:
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, LlamaForCausalLM
+
+model_id = "tiiuae/falcon-7b"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_id, 
+    load_in_8bit=True,
+    use_flash_attention_2=True,
+)
+```
+
+### Combining Flash Attention 2 and 4-bit models
+
+You can combine this feature together with 4-bit quantization:
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, LlamaForCausalLM
+
+model_id = "tiiuae/falcon-7b"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_id, 
+    load_in_4bit=True,
+    use_flash_attention_2=True,
+)
+```
+
+### Combining Flash Attention 2 and PEFT
+
+You can combine this feature together with PEFT for training adapters using Flash Attention 2 under the hood:
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, LlamaForCausalLM
+from peft import LoraConfig
+
+model_id = "tiiuae/falcon-7b"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_id, 
+    load_in_4bit=True,
+    use_flash_attention_2=True,
+)
+
+lora_config = LoraConfig(
+    r=8,
+    task_type="CAUSAL_LM"
+)
+
+model.add_adapter(lora_config)
+
+... # train your model
+```
+
 ## BetterTransformer
 
 [BetterTransformer](https://huggingface.co/docs/optimum/bettertransformer/overview) converts 🤗 Transformers models to use the PyTorch-native fastpath execution, which calls optimized kernels like Flash Attention under the hood.  

docs/source/en/perf_train_gpu_one.md

@@ -228,6 +228,10 @@ For additional information on tf32 vs other precisions, please refer to the foll
 [RTX-3090](https://github.com/huggingface/transformers/issues/14608#issuecomment-1004390803) and
 [A100](https://github.com/huggingface/transformers/issues/15026#issuecomment-1004543189).
 
+## Flash Attention 2
+
+You can speedup the training throughput by using Flash Attention 2 integration in transformers. Check out the appropriate section in the [single GPU section](./perf_infer_gpu_one#Flash-Attention-2) to learn more about how to load a model with Flash Attention 2 modules. 
+
 ## Optimizer choice
 
 The most common optimizer used to train transformer models is Adam or AdamW (Adam with weight decay). Adam achieves 

docs/source/en/pipeline_tutorial.md

@@ -30,33 +30,44 @@ Take a look at the [`pipeline`] documentation for a complete list of supported t
 
 ## Pipeline usage
 
-While each task has an associated [`pipeline`], it is simpler to use the general [`pipeline`] abstraction which contains all the task-specific pipelines. The [`pipeline`] automatically loads a default model and a preprocessing class capable of inference for your task.
+While each task has an associated [`pipeline`], it is simpler to use the general [`pipeline`] abstraction which contains 
+all the task-specific pipelines. The [`pipeline`] automatically loads a default model and a preprocessing class capable 
+of inference for your task. Let's take the example of using the [`pipeline`] for automatic speech recognition (ASR), or
+speech-to-text.
 
-1. Start by creating a [`pipeline`] and specify an inference task:
+
+1. Start by creating a [`pipeline`] and specify the inference task:
 
 ```py
 >>> from transformers import pipeline
 
->>> generator = pipeline(task="automatic-speech-recognition")
+>>> transcriber = pipeline(task="automatic-speech-recognition")
 ```
 
-2. Pass your input text to the [`pipeline`]:
+2. Pass your input to the [`pipeline`]. In the case of speech recognition, this is an audio input file:
 
 ```py
->>> generator("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac")
+>>> transcriber("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac")
 {'text': 'I HAVE A DREAM BUT ONE DAY THIS NATION WILL RISE UP LIVE UP THE TRUE MEANING OF ITS TREES'}
 ```
 
-Not the result you had in mind? Check out some of the [most downloaded automatic speech recognition models](https://huggingface.co/models?pipeline_tag=automatic-speech-recognition&sort=downloads) on the Hub to see if you can get a better transcription.
-Let's try [openai/whisper-large](https://huggingface.co/openai/whisper-large):
+Not the result you had in mind? Check out some of the [most downloaded automatic speech recognition models](https://huggingface.co/models?pipeline_tag=automatic-speech-recognition&sort=trending) 
+on the Hub to see if you can get a better transcription.
+
+Let's try the [Whisper large-v2](https://huggingface.co/openai/whisper-large) model from OpenAI. Whisper was released 
+2 years later than Wav2Vec2, and was trained on close to 10x more data. As such, it beats Wav2Vec2 on most downstream 
+benchmarks. It also has the added benefit of predicting punctuation and casing, neither of which are possible with  
+Wav2Vec2.
+
+Let's give it a try here to see how it performs:
 
 ```py
->>> generator = pipeline(model="openai/whisper-large")
->>> generator("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac")
+>>> transcriber = pipeline(model="openai/whisper-large-v2")
+>>> transcriber("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac")
 {'text': ' I have a dream that one day this nation will rise up and live out the true meaning of its creed.'}
 ```
 
-Now this result looks more accurate!
+Now this result looks more accurate! For a deep-dive comparison on Wav2Vec2 vs Whisper, refer to the [Audio Transformers Course](https://huggingface.co/learn/audio-course/chapter5/asr_models).
 We really encourage you to check out the Hub for models in different languages, models specialized in your field, and more.
 You can check out and compare model results directly from your browser on the Hub to see if it fits or 
 handles corner cases better than other ones.
@@ -65,7 +76,7 @@ And if you don't find a model for your use case, you can always start [training]
 If you have several inputs, you can pass your input as a list:
 
 ```py
-generator(
+transcriber(
     [
         "https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac",
         "https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/1.flac",
@@ -73,22 +84,22 @@ generator(
 )
 ```
 
-If you want to iterate over a whole dataset, or want to use it for inference in a webserver, check out dedicated parts
-
-[Using pipelines on a dataset](#using-pipelines-on-a-dataset)
-
-[Using pipelines for a webserver](./pipeline_webserver)
+Pipelines are great for experimentation as switching from one model to another is trivial; however, there are some ways to optimize them for larger workloads than experimentation. See the following guides that dive into iterating over whole datasets or using pipelines in a webserver:
+of the docs:
+* [Using pipelines on a dataset](#using-pipelines-on-a-dataset)
+* [Using pipelines for a webserver](./pipeline_webserver)
 
 ## Parameters
 
 [`pipeline`] supports many parameters; some are task specific, and some are general to all pipelines.
-In general you can specify parameters anywhere you want:
+In general, you can specify parameters anywhere you want:
 
 ```py
-generator = pipeline(model="openai/whisper-large", my_parameter=1)
-out = generator(...)  # This will use `my_parameter=1`.
-out = generator(..., my_parameter=2)  # This will override and use `my_parameter=2`.
-out = generator(...)  # This will go back to using `my_parameter=1`.
+transcriber = pipeline(model="openai/whisper-large-v2", my_parameter=1)
+
+out = transcriber(...)  # This will use `my_parameter=1`.
+out = transcriber(..., my_parameter=2)  # This will override and use `my_parameter=2`.
+out = transcriber(...)  # This will go back to using `my_parameter=1`.
 ```
 
 Let's check out 3 important ones:
@@ -99,14 +110,21 @@ If you use `device=n`, the pipeline automatically puts the model on the specifie
 This will work regardless of whether you are using PyTorch or Tensorflow.
 
 ```py
-generator = pipeline(model="openai/whisper-large", device=0)
+transcriber = pipeline(model="openai/whisper-large-v2", device=0)
 ```
 
-If the model is too large for a single GPU, you can set `device_map="auto"` to allow 🤗 [Accelerate](https://huggingface.co/docs/accelerate) to automatically determine how to load and store the model weights.
+If the model is too large for a single GPU and you are using PyTorch, you can set `device_map="auto"` to automatically 
+determine how to load and store the model weights. Using the `device_map` argument requires the 🤗 [Accelerate](https://huggingface.co/docs/accelerate)
+package:
+
+```bash
+pip install --upgrade accelerate
+```
+
+The following code automatically loads and stores model weights across devices:
 
 ```py
-#!pip install accelerate
-generator = pipeline(model="openai/whisper-large", device_map="auto")
+transcriber = pipeline(model="openai/whisper-large-v2", device_map="auto")
 ```
 
 Note that if  `device_map="auto"` is passed, there is no need to add the argument `device=device` when instantiating your `pipeline` as you may encounter some unexpected behavior!
@@ -118,12 +136,12 @@ By default, pipelines will not batch inference for reasons explained in detail [
 But if it works in your use case, you can use:
 
 ```py
-generator = pipeline(model="openai/whisper-large", device=0, batch_size=2)
-audio_filenames = [f"audio_{i}.flac" for i in range(10)]
-texts = generator(audio_filenames)
+transcriber = pipeline(model="openai/whisper-large-v2", device=0, batch_size=2)
+audio_filenames = [f"https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/{i}.flac" for i in range(1, 5)]
+texts = transcriber(audio_filenames)
 ```
 
-This runs the pipeline on the 10 provided audio files, but it will pass them in batches of 2
+This runs the pipeline on the 4 provided audio files, but it will pass them in batches of 2
 to the model (which is on a GPU, where batching is more likely to help) without requiring any further code from you. 
 The output should always match what you would have received without batching. It is only meant as a way to help you get more speed out of a pipeline.
 
@@ -136,18 +154,23 @@ For instance, the [`transformers.AutomaticSpeechRecognitionPipeline.__call__`] m
 
 
 ```py
->>> # Not using whisper, as it cannot provide timestamps.
->>> generator = pipeline(model="facebook/wav2vec2-large-960h-lv60-self", return_timestamps="word")
->>> generator("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac")
-{'text': 'I HAVE A DREAM BUT ONE DAY THIS NATION WILL RISE UP AND LIVE OUT THE TRUE MEANING OF ITS CREED', 'chunks': [{'text': 'I', 'timestamp': (1.22, 1.24)}, {'text': 'HAVE', 'timestamp': (1.42, 1.58)}, {'text': 'A', 'timestamp': (1.66, 1.68)}, {'text': 'DREAM', 'timestamp': (1.76, 2.14)}, {'text': 'BUT', 'timestamp': (3.68, 3.8)}, {'text': 'ONE', 'timestamp': (3.94, 4.06)}, {'text': 'DAY', 'timestamp': (4.16, 4.3)}, {'text': 'THIS', 'timestamp': (6.36, 6.54)}, {'text': 'NATION', 'timestamp': (6.68, 7.1)}, {'text': 'WILL', 'timestamp': (7.32, 7.56)}, {'text': 'RISE', 'timestamp': (7.8, 8.26)}, {'text': 'UP', 'timestamp': (8.38, 8.48)}, {'text': 'AND', 'timestamp': (10.08, 10.18)}, {'text': 'LIVE', 'timestamp': (10.26, 10.48)}, {'text': 'OUT', 'timestamp': (10.58, 10.7)}, {'text': 'THE', 'timestamp': (10.82, 10.9)}, {'text': 'TRUE', 'timestamp': (10.98, 11.18)}, {'text': 'MEANING', 'timestamp': (11.26, 11.58)}, {'text': 'OF', 'timestamp': (11.66, 11.7)}, {'text': 'ITS', 'timestamp': (11.76, 11.88)}, {'text': 'CREED', 'timestamp': (12.0, 12.38)}]}
+>>> transcriber = pipeline(model="openai/whisper-large-v2", return_timestamps=True)
+>>> transcriber("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac")
+{'text': ' I have a dream that one day this nation will rise up and live out the true meaning of its creed.', 'chunks': [{'timestamp': (0.0, 11.88), 'text': ' I have a dream that one day this nation will rise up and live out the true meaning of its'}, {'timestamp': (11.88, 12.38), 'text': ' creed.'}]}
 ```
 
-As you can see, the model inferred the text and also outputted **when** the various words were pronounced
-in the sentence.
+As you can see, the model inferred the text and also outputted **when** the various sentences were pronounced.
 
 There are many parameters available for each task, so check out each task's API reference to see what you can tinker with!
-For instance, the [`~transformers.AutomaticSpeechRecognitionPipeline`] has a `chunk_length_s` parameter which is helpful for working on really long audio files (for example, subtitling entire movies or hour-long videos) that a model typically cannot handle on its own.
+For instance, the [`~transformers.AutomaticSpeechRecognitionPipeline`] has a `chunk_length_s` parameter which is helpful 
+for working on really long audio files (for example, subtitling entire movies or hour-long videos) that a model typically 
+cannot handle on its own:
 
+```python
+>>> transcriber = pipeline(model="openai/whisper-large-v2", chunk_length_s=30, return_timestamps=True)
+>>> transcriber("https://huggingface.co/datasets/sanchit-gandhi/librispeech_long/resolve/main/audio.wav")
+{'text': " Chapter 16. I might have told you of the beginning of this liaison in a few lines, but I wanted you to see every step by which we came.  I, too, agree to whatever Marguerite wished, Marguerite to be unable to live apart from me. It was the day after the evening...
+```
 
 If you can't find a parameter that would really help you out, feel free to [request it](https://github.com/huggingface/transformers/issues/new?assignees=&labels=feature&template=feature-request.yml)!
 

docs/source/en/tasks/idefics.md

@@ -0,0 +1,426 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Image tasks with IDEFICS
+
+[[open-in-colab]]
+
+While individual tasks can be tackled by fine-tuning specialized models, an alternative approach 
+that has recently emerged and gained popularity is to use large models for a diverse set of tasks without fine-tuning. 
+For instance, large language models can handle such NLP tasks as summarization, translation, classification, and more. 
+This approach is no longer limited to a single modality, such as text, and in this guide, we will illustrate how you can 
+solve image-text tasks with a large multimodal model called IDEFICS. 
+
+[IDEFICS](../model_doc/idefics) is an open-access vision and language model based on [Flamingo](https://huggingface.co/papers/2204.14198), 
+a state-of-the-art visual language model initially developed by DeepMind. The model accepts arbitrary sequences of image 
+and text inputs and generates coherent text as output. It can answer questions about images, describe visual content, 
+create stories grounded in multiple images, and so on. IDEFICS comes in two variants - [80 billion parameters](https://huggingface.co/HuggingFaceM4/idefics-80b) 
+and [9 billion parameters](https://huggingface.co/HuggingFaceM4/idefics-9b), both of which are available on the 🤗 Hub. For each variant, you can also find fine-tuned instructed 
+versions of the model adapted for conversational use cases.
+
+This model is exceptionally versatile and can be used for a wide range of image and multimodal tasks. However, 
+being a large model means it requires significant computational resources and infrastructure. It is up to you to decide whether 
+this approach suits your use case better than fine-tuning specialized models for each individual task. 
+
+In this guide, you'll learn how to: 
+- [Load IDEFICS](#loading-the-model) and [load the quantized version of the model](#loading-the-quantized-version-of-the-model)
+- Use IDEFICS for: 
+  - [Image captioning](#image-captioning)
+  - [Prompted image captioning](#prompted-image-captioning)
+  - [Few-shot prompting](#few-shot-prompting)
+  - [Visual question answering](#visual-question-answering)
+  - [Image classificaiton](#image-classification)
+  - [Image-guided text generation](#image-guided-text-generation)
+- [Run inference in batch mode](#running-inference-in-batch-mode)
+- [Run IDEFICS instruct for conversational use](#idefics-instruct-for-conversational-use)
+
+Before you begin, make sure you have all the necessary libraries installed. 
+
+```bash
+pip install -q bitsandbytes sentencepiece accelerate transformers
+```
+
+<Tip>
+To run the following examples with a non-quantized version of the model checkpoint you will need at least 20GB of GPU memory.
+</Tip>
+
+## Loading the model
+
+Let's start by loading the model's 9 billion parameters checkpoint: 
+
+```py
+>>> checkpoint = "HuggingFaceM4/idefics-9b"
+```
+
+Just like for other Transformers models, you need to load a processor and the model itself from the checkpoint. 
+The IDEFICS processor wraps a [`LlamaTokenizer`] and IDEFICS image processor into a single processor to take care of 
+preparing text and image inputs for the model.
+
+```py
+>>> import torch
+
+>>> from transformers import IdeficsForVisionText2Text, AutoProcessor
+
+>>> processor = AutoProcessor.from_pretrained(checkpoint)
+
+>>> model = IdeficsForVisionText2Text.from_pretrained(checkpoint, torch_dtype=torch.bfloat16, device_map="auto")
+```
+
+Setting `device_map` to `"auto"` will automatically determine how to load and store the model weights in the most optimized 
+manner given existing devices.
+
+### Quantized model
+
+If high-memory GPU availability is an issue, you can load the quantized version of the model. To load the model and the 
+processor in 4bit precision, pass a `BitsAndBytesConfig` to the `from_pretrained` method and the model will be compressed 
+on the fly while loading.
+
+```py
+>>> import torch
+>>> from transformers import IdeficsForVisionText2Text, AutoProcessor, BitsAndBytesConfig
+
+>>> quantization_config = BitsAndBytesConfig(
+...     load_in_4bit=True,
+...     bnb_4bit_compute_dtype=torch.float16,
+... )
+
+>>> processor = AutoProcessor.from_pretrained(checkpoint)
+
+>>> model = IdeficsForVisionText2Text.from_pretrained(
+...     checkpoint,
+...     quantization_config=quantization_config,
+...     device_map="auto"
+... )
+```
+
+Now that you have the model loaded in one of the suggested ways, let's move on to exploring tasks that you can use IDEFICS for.
+
+## Image captioning
+
+Image captioning is the task of predicting a caption for a given image. A common application is to aid visually impaired 
+people navigate through different situations, for instance, explore image content online. 
+
+To illustrate the task, get an image to be captioned, e.g.:
+
+<div class="flex justify-center">
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/idefics-im-captioning.jpg" alt="Image of a puppy in a flower bed"/>
+</div>
+
+Photo by [Hendo Wang](https://unsplash.com/@hendoo). 
+
+IDEFICS accepts text and image prompts. However, to caption an image, you do not have to provide a text prompt to the 
+model, only the preprocessed input image. Without a text prompt, the model will start generating text from the 
+BOS (beginning-of-sequence) token thus creating a caption.
+
+As image input to the model, you can use either an image object (`PIL.Image`) or a url from which the image can be retrieved.
+
+```py
+>>> prompt = [
+...     "https://images.unsplash.com/photo-1583160247711-2191776b4b91?ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D&auto=format&fit=crop&w=3542&q=80",
+... ]
+
+>>> inputs = processor(prompt, return_tensors="pt").to("cuda")
+>>> bad_words_ids = processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
+
+>>> generated_ids = model.generate(**inputs, max_new_tokens=10, bad_words_ids=bad_words_ids)
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+>>> print(generated_text[0])
+A puppy in a flower bed
+```
+
+<Tip>
+
+It is a good idea to include the `bad_words_ids` in the call to `generate` to avoid errors arising when increasing 
+the `max_new_tokens`: the model will want to generate a new `<image>` or `<fake_token_around_image>` token when there 
+is no image being generated by the model.
+You can set it on-the-fly as in this guide, or store in the `GenerationConfig` as described in the [Text generation strategies](../generation_strategies) guide.
+</Tip>
+
+## Prompted image captioning
+
+You can extend image captioning by providing a text prompt, which the model will continue given the image. Let's take 
+another image to illustrate:
+
+<div class="flex justify-center">
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/idefics-prompted-im-captioning.jpg" alt="Image of the Eiffel Tower at night"/>
+</div>
+
+Photo by [Denys Nevozhai](https://unsplash.com/@dnevozhai).
+   
+Textual and image prompts can be passed to the model's processor as a single list to create appropriate inputs.
+
+```py
+>>> prompt = [
+...     "https://images.unsplash.com/photo-1543349689-9a4d426bee8e?ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D&auto=format&fit=crop&w=3501&q=80",
+...     "This is an image of ",
+... ]
+
+>>> inputs = processor(prompt, return_tensors="pt").to("cuda")
+>>> bad_words_ids = processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
+
+>>> generated_ids = model.generate(**inputs, max_new_tokens=10, bad_words_ids=bad_words_ids)
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+>>> print(generated_text[0])
+This is an image of the Eiffel Tower in Paris, France.
+```
+
+## Few-shot prompting
+
+While IDEFICS demonstrates great zero-shot results, your task may require a certain format of the caption, or come with 
+other restrictions or requirements that increase task's complexity. Few-shot prompting can be used to enable in-context learning.
+By providing examples in the prompt, you can steer the model to generate results that mimic the format of given examples. 
+
+Let's use the previous image of the Eiffel Tower as an example for the model and build a prompt that demonstrates to the model 
+that in addition to learning what the object in an image is, we would also like to get some interesting information about it. 
+Then, let's see, if we can get the same response format for an image of the Statue of Liberty:
+
+<div class="flex justify-center">
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/idefics-few-shot.jpg" alt="Image of the Statue of Liberty"/>
+</div>
+
+Photo by [Juan Mayobre](https://unsplash.com/@jmayobres).
+  
+```py
+>>> prompt = ["User:",
+...            "https://images.unsplash.com/photo-1543349689-9a4d426bee8e?ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D&auto=format&fit=crop&w=3501&q=80",
+...            "Describe this image.\nAssistant: An image of the Eiffel Tower at night. Fun fact: the Eiffel Tower is the same height as an 81-storey building.\n",
+...            "User:",
+...            "https://images.unsplash.com/photo-1524099163253-32b7f0256868?ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D&auto=format&fit=crop&w=3387&q=80",
+...            "Describe this image.\nAssistant:"
+...            ]
+
+>>> inputs = processor(prompt, return_tensors="pt").to("cuda")
+>>> bad_words_ids = processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
+
+>>> generated_ids = model.generate(**inputs, max_new_tokens=30, bad_words_ids=bad_words_ids)
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+>>> print(generated_text[0])
+User: Describe this image.
+Assistant: An image of the Eiffel Tower at night. Fun fact: the Eiffel Tower is the same height as an 81-storey building. 
+User: Describe this image.
+Assistant: An image of the Statue of Liberty. Fun fact: the Statue of Liberty is 151 feet tall.
+```
+
+Notice that just from a single example (i.e., 1-shot) the model has learned how to perform the task. For more complex tasks, 
+feel free to experiment with a larger number of examples (e.g., 3-shot, 5-shot, etc.).
+
+## Visual question answering
+
+Visual Question Answering (VQA) is the task of answering open-ended questions based on an image. Similar to image 
+captioning it can be used in accessibility applications, but also in education (reasoning about visual materials), customer 
+service (questions about products based on images), and image retrieval.
+
+Let's get a new image for this task: 
+
+<div class="flex justify-center">
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/idefics-vqa.jpg" alt="Image of a couple having a picnic"/>
+</div>
+
+Photo by [Jarritos Mexican Soda](https://unsplash.com/@jarritos).
+
+You can steer the model from image captioning to visual question answering by prompting it with appropriate instructions: 
+
+```py
+>>> prompt = [
+...     "Instruction: Provide an answer to the question. Use the image to answer.\n",
+...     "https://images.unsplash.com/photo-1623944889288-cd147dbb517c?ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D&auto=format&fit=crop&w=3540&q=80",
+...     "Question: Where are these people and what's the weather like? Answer:"
+... ]
+
+>>> inputs = processor(prompt, return_tensors="pt").to("cuda")
+>>> bad_words_ids = processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
+
+>>> generated_ids = model.generate(**inputs, max_new_tokens=20, bad_words_ids=bad_words_ids)
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+>>> print(generated_text[0])
+Instruction: Provide an answer to the question. Use the image to answer.
+ Question: Where are these people and what's the weather like? Answer: They're in a park in New York City, and it's a beautiful day.
+```
+
+## Image classification
+
+IDEFICS is capable of classifying images into different categories without being explicitly trained on data containing 
+labeled examples from those specific categories. Given a list of categories and using its image and text understanding 
+capabilities, the model can infer which category the image likely belongs to. 
+
+Say, we have this image of a vegetable stand: 
+
+<div class="flex justify-center">
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/idefics-classification.jpg" alt="Image of a vegetable stand"/>
+</div>
+
+Photo by [Peter Wendt](https://unsplash.com/@peterwendt).
+
+We can instruct the model to classify the image into one of the categories that we have:
+
+```py
+>>> categories = ['animals','vegetables', 'city landscape', 'cars', 'office']
+>>> prompt = [f"Instruction: Classify the following image into a single category from the following list: {categories}.\n",
+...     "https://images.unsplash.com/photo-1471193945509-9ad0617afabf?ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D&auto=format&fit=crop&w=3540&q=80",    
+...     "Category: "
+... ]
+
+>>> inputs = processor(prompt, return_tensors="pt").to("cuda")
+>>> bad_words_ids = processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
+
+>>> generated_ids = model.generate(**inputs, max_new_tokens=4, bad_words_ids=bad_words_ids)
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+>>> print(generated_text[0])
+Instruction: Classify the following image into a single category from the following list: ['animals', 'vegetables', 'city landscape', 'cars', 'office'].
+Category: Vegetables
+```  
+
+In the example above we instruct the model to classify the image into a single category, however, you can also prompt the model to do rank classification.
+
+## Image-guided text generation
+
+For more creative applications, you can use image-guided text generation to generate text based on an image. This can be 
+useful to create descriptions of products, ads, descriptions of a scene, etc. 
+
+Let's prompt IDEFICS to write a story based on a simple image of a red door: 
+
+<div class="flex justify-center">
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/idefics-story-generation.jpg" alt="Image of a red door with a pumpkin on the steps"/>
+</div>
+
+Photo by [Craig Tidball](https://unsplash.com/@devonshiremedia).
+  
+```py
+>>> prompt = ["Instruction: Use the image to write a story. \n",
+...     "https://images.unsplash.com/photo-1517086822157-2b0358e7684a?ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D&auto=format&fit=crop&w=2203&q=80",
+...     "Story: \n"]
+
+>>> inputs = processor(prompt, return_tensors="pt").to("cuda")
+>>> bad_words_ids = processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
+
+>>> generated_ids = model.generate(**inputs, num_beams=2, max_new_tokens=200, bad_words_ids=bad_words_ids)
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+>>> print(generated_text[0]) 
+Instruction: Use the image to write a story. 
+ Story: 
+Once upon a time, there was a little girl who lived in a house with a red door.  She loved her red door.  It was the prettiest door in the whole world.
+
+One day, the little girl was playing in her yard when she noticed a man standing on her doorstep.  He was wearing a long black coat and a top hat.
+
+The little girl ran inside and told her mother about the man.
+
+Her mother said, “Don’t worry, honey.  He’s just a friendly ghost.”
+
+The little girl wasn’t sure if she believed her mother, but she went outside anyway.
+
+When she got to the door, the man was gone.
+
+The next day, the little girl was playing in her yard again when she noticed the man standing on her doorstep.
+
+He was wearing a long black coat and a top hat.
+
+The little girl ran
+```
+
+Looks like IDEFICS noticed the pumpkin on the doorstep and went with a spooky Halloween story about a ghost.
+
+<Tip>
+
+For longer outputs like this, you will greatly benefit from tweaking the text generation strategy. This can help 
+you significantly improve the quality of the generated output. Check out [Text generation strategies](../generation_strategies) 
+to learn more. 
+</Tip>
+
+## Running inference in batch mode
+
+All of the earlier sections illustrated IDEFICS for a single example. In a very similar fashion, you can run inference 
+for a batch of examples by passing a list of prompts:
+
+```py
+>>> prompts = [
+...     [   "https://images.unsplash.com/photo-1543349689-9a4d426bee8e?ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D&auto=format&fit=crop&w=3501&q=80",
+...         "This is an image of ",
+...     ],
+...     [   "https://images.unsplash.com/photo-1623944889288-cd147dbb517c?ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D&auto=format&fit=crop&w=3540&q=80",
+...         "This is an image of ",
+...     ],
+...     [   "https://images.unsplash.com/photo-1471193945509-9ad0617afabf?ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D&auto=format&fit=crop&w=3540&q=80",
+...         "This is an image of ",
+...     ],
+... ]
+
+>>> inputs = processor(prompts, return_tensors="pt")
+>>> bad_words_ids = processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
+
+>>> generated_ids = model.generate(**inputs, max_new_tokens=10, bad_words_ids=bad_words_ids)
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+>>> for i,t in enumerate(generated_text):
+...     print(f"{i}:\n{t}\n") 
+0:
+This is an image of the Eiffel Tower in Paris, France.
+
+1:
+This is an image of a couple on a picnic blanket.
+
+2:
+This is an image of a vegetable stand.
+```
+
+## IDEFICS instruct for conversational use
+
+For conversational use cases, you can find fine-tuned instructed versions of the model on the 🤗 Hub: 
+`HuggingFaceM4/idefics-80b-instruct` and `HuggingFaceM4/idefics-9b-instruct`.
+
+These checkpoints are the result of fine-tuning the respective base models on a mixture of supervised and instruction 
+fine-tuning datasets, which boosts the downstream performance while making the models more usable in conversational settings.
+
+The use and prompting for the conversational use is very similar to using the base models: 
+
+```py
+>>> import torch
+>>> from transformers import IdeficsForVisionText2Text, AutoProcessor
+
+>>> device = "cuda" if torch.cuda.is_available() else "cpu"
+
+>>> checkpoint = "HuggingFaceM4/idefics-9b-instruct"
+>>> model = IdeficsForVisionText2Text.from_pretrained(checkpoint, torch_dtype=torch.bfloat16).to(device)
+>>> processor = AutoProcessor.from_pretrained(checkpoint)
+
+>>> prompts = [
+...     [
+...         "User: What is in this image?",
+...         "https://upload.wikimedia.org/wikipedia/commons/8/86/Id%C3%A9fix.JPG",
+...         "<end_of_utterance>",
+
+...         "\nAssistant: This picture depicts Idefix, the dog of Obelix in Asterix and Obelix. Idefix is running on the ground.<end_of_utterance>",
+
+...         "\nUser:",
+...         "https://static.wikia.nocookie.net/asterix/images/2/25/R22b.gif/revision/latest?cb=20110815073052",
+...         "And who is that?<end_of_utterance>",
+
+...         "\nAssistant:",
+...     ],
+... ]
+
+>>> # --batched mode
+>>> inputs = processor(prompts, add_end_of_utterance_token=False, return_tensors="pt").to(device)
+>>> # --single sample mode
+>>> # inputs = processor(prompts[0], return_tensors="pt").to(device)
+
+>>> # Generation args
+>>> exit_condition = processor.tokenizer("<end_of_utterance>", add_special_tokens=False).input_ids
+>>> bad_words_ids = processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
+
+>>> generated_ids = model.generate(**inputs, eos_token_id=exit_condition, bad_words_ids=bad_words_ids, max_length=100)
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+>>> for i, t in enumerate(generated_text):
+...     print(f"{i}:\n{t}\n")
+```

docs/source/en/tasks/language_modeling.md

@@ -37,7 +37,7 @@ You can finetune other architectures for causal language modeling following the
 Choose one of the following architectures:
 
 <!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
-[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeLlama](../model_doc/code_llama), [CodeGen](../model_doc/codegen), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [Falcon](../model_doc/falcon), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MPT](../model_doc/mpt), [MusicGen](../model_doc/musicgen), [MVP](../model_doc/mvp), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [RWKV](../model_doc/rwkv), [Speech2Text2](../model_doc/speech_to_text_2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod)
+[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeLlama](../model_doc/code_llama), [CodeGen](../model_doc/codegen), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [Falcon](../model_doc/falcon), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [Mistral](../model_doc/mistral), [MPT](../model_doc/mpt), [MusicGen](../model_doc/musicgen), [MVP](../model_doc/mvp), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [Persimmon](../model_doc/persimmon), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [RWKV](../model_doc/rwkv), [Speech2Text2](../model_doc/speech_to_text_2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod)
 
 
 

docs/source/en/tasks/object_detection.md

@@ -136,8 +136,8 @@ To get an even better understanding of the data, visualize an example in the dat
 >>> label2id = {v: k for k, v in id2label.items()}
 
 >>> for i in range(len(annotations["id"])):
-...     box = annotations["bbox"][i - 1]
-...     class_idx = annotations["category"][i - 1]
+...     box = annotations["bbox"][i]
+...     class_idx = annotations["category"][i]
 ...     x, y, w, h = tuple(box)
 ...     draw.rectangle((x, y, x + w, y + h), outline="red", width=1)
 ...     draw.text((x, y), id2label[class_idx], fill="white")

docs/source/en/tasks/semantic_segmentation.md

@@ -206,7 +206,7 @@ The transform is applied on the fly which is faster and consumes less disk space
 
 ## Evaluate
 
-Including a metric during training is often helpful for evaluating your model's performance. You can quickly load a evaluation method with the 🤗 [Evaluate](https://huggingface.co/docs/evaluate/index) library. For this task, load the [mean Intersection over Union](https://huggingface.co/spaces/evaluate-metric/accuracy) (IoU) metric (see the 🤗 Evaluate [quick tour](https://huggingface.co/docs/evaluate/a_quick_tour) to learn more about how to load and compute a metric):
+Including a metric during training is often helpful for evaluating your model's performance. You can quickly load an evaluation method with the 🤗 [Evaluate](https://huggingface.co/docs/evaluate/index) library. For this task, load the [mean Intersection over Union](https://huggingface.co/spaces/evaluate-metric/accuracy) (IoU) metric (see the 🤗 Evaluate [quick tour](https://huggingface.co/docs/evaluate/a_quick_tour) to learn more about how to load and compute a metric):
 
 ```py
 >>> import evaluate

docs/source/en/tasks/sequence_classification.md

@@ -33,7 +33,7 @@ The task illustrated in this tutorial is supported by the following model archit
 <!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
 
 
-[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [CodeLlama](../model_doc/code_llama), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MPT](../model_doc/mpt), [MRA](../model_doc/mra), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [T5](../model_doc/t5), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [UMT5](../model_doc/umt5), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
+[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [CodeLlama](../model_doc/code_llama), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [Mistral](../model_doc/mistral), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MPT](../model_doc/mpt), [MRA](../model_doc/mra), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [Persimmon](../model_doc/persimmon), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [T5](../model_doc/t5), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [UMT5](../model_doc/umt5), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
 
 
 

docs/source/en/tasks/text-to-speech.md

@@ -19,16 +19,40 @@ rendered properly in your Markdown viewer.
 [[open-in-colab]]
 
 Text-to-speech (TTS) is the task of creating natural-sounding speech from text, where the speech can be generated in multiple 
-languages and for multiple speakers. The only text-to-speech model currently available in 🤗 Transformers 
-is [SpeechT5](model_doc/speecht5), though more will be added in the future. SpeechT5 is pre-trained on a combination of 
+languages and for multiple speakers. Several text-to-speech models are currently available in 🤗 Transformers, such as 
+[Bark](../model_doc/bark), [MMS](../model_doc/mms), [VITS](../model_doc/vits) and [SpeechT5](../model_doc/speecht5). 
+
+You can easily generate audio using the `"text-to-audio"` pipeline (or its alias - `"text-to-speech"`). Some models, like Bark, 
+can also be conditioned to generate non-verbal communications such as laughing, sighing and crying, or even add music.
+Here's an example of how you would use the `"text-to-speech"` pipeline with Bark: 
+
+```py
+>>> from transformers import pipeline
+
+>>> pipe = pipeline("text-to-speech", model="suno/bark-small")
+>>> text = "[clears throat] This is a test ... and I just took a long pause."
+>>> output = pipe(text)
+```
+
+Here's a code snippet you can use to listen to the resulting audio in a notebook: 
+
+```python
+>>> from IPython.display import Audio
+>>> Audio(output["audio"], rate=output["sampling_rate"])
+```
+
+For more examples on what Bark and other pretrained TTS models can do, refer to our 
+[Audio course](https://huggingface.co/learn/audio-course/chapter6/pre-trained_models). 
+
+If you are looking to fine-tune a TTS model, you can currently fine-tune SpeechT5 only. SpeechT5 is pre-trained on a combination of 
 speech-to-text and text-to-speech data, allowing it to learn a unified space of hidden representations shared by both text 
 and speech. This means that the same pre-trained model can be fine-tuned for different tasks. Furthermore, SpeechT5 
 supports multiple speakers through x-vector speaker embeddings. 
 
-This guide illustrates how to:
+The remainder of this guide illustrates how to:
 
-1. Fine-tune [SpeechT5](model_doc/speecht5) that was originally trained on English speech on the Dutch (`nl`) language subset of the [VoxPopuli](https://huggingface.co/datasets/facebook/voxpopuli) dataset.
-2. Use your fine-tuned model for inference.
+1. Fine-tune [SpeechT5](../model_doc/speecht5) that was originally trained on English speech on the Dutch (`nl`) language subset of the [VoxPopuli](https://huggingface.co/datasets/facebook/voxpopuli) dataset.
+2. Use your refined model for inference in one of two ways: using a pipeline or directly.
 
 Before you begin, make sure you have all the necessary libraries installed:
 
@@ -485,6 +509,12 @@ the `per_device_train_batch_size` incrementally by factors of 2 and increase `gr
 >>> trainer.train()
 ```
 
+To be able to use your checkpoint with a pipeline, make sure to save the processor with the checkpoint: 
+
+```py
+>>> processor.save_pretrained("YOUR_ACCOUNT_NAME/speecht5_finetuned_voxpopuli_nl")
+```
+
 Push the final model to the 🤗 Hub:
 
 ```py
@@ -493,29 +523,70 @@ Push the final model to the 🤗 Hub:
 
 ## Inference
 
+### Inference with a pipeline
+
 Great, now that you've fine-tuned a model, you can use it for inference!
-Load the model from the 🤗 Hub (make sure to use your account name in the following code snippet): 
+First, let's see how you can use it with a corresponding pipeline. Let's create a `"text-to-speech"` pipeline with your 
+checkpoint: 
+
+```py
+>>> from transformers import pipeline
+
+>>> pipe = pipeline("text-to-speech", model="YOUR_ACCOUNT_NAME/speecht5_finetuned_voxpopuli_nl")
+```
+
+Pick a piece of text in Dutch you'd like narrated, e.g.:
+
+```py
+>>> text = "hallo allemaal, ik praat nederlands. groetjes aan iedereen!"
+```
+
+To use SpeechT5 with the pipeline, you'll need a speaker embedding. Let's get it from an example in the test dataset: 
+
+```py
+>>> example = dataset["test"][304]
+>>> speaker_embeddings = torch.tensor(example["speaker_embeddings"]).unsqueeze(0)
+```
+
+Now you can pass the text and speaker embeddings to the pipeline, and it will take care of the rest: 
+
+```py
+>>> forward_params = {"speaker_embeddings": speaker_embeddings}
+>>> output = pipe(text, forward_params=forward_params)
+>>> output
+{'audio': array([-6.82714235e-05, -4.26525949e-04,  1.06134125e-04, ...,
+        -1.22392643e-03, -7.76011671e-04,  3.29112721e-04], dtype=float32),
+ 'sampling_rate': 16000}
+```
+
+You can then listen to the result:
+
+```py
+>>> from IPython.display import Audio
+>>> Audio(output['audio'], rate=output['sampling_rate']) 
+```
+
+### Run inference manually
+
+You can achieve the same inference results without using the pipeline, however, more steps will be required. 
+
+Load the model from the 🤗 Hub: 
 
 ```py
 >>> model = SpeechT5ForTextToSpeech.from_pretrained("YOUR_ACCOUNT/speecht5_finetuned_voxpopuli_nl")
 ```
 
-Pick an example, here we'll take one from the test dataset. Obtain a speaker embedding. 
+Pick an example from the test dataset obtain a speaker embedding. 
 
 ```py 
 >>> example = dataset["test"][304]
 >>> speaker_embeddings = torch.tensor(example["speaker_embeddings"]).unsqueeze(0)
 ```
 
-Define some input text and tokenize it.
+Define the input text and tokenize it.
 
 ```py 
 >>> text = "hallo allemaal, ik praat nederlands. groetjes aan iedereen!"
-```
-
-Preprocess the input text: 
-
-```py
 >>> inputs = processor(text=text, return_tensors="pt")
 ```
 

docs/source/en/tasks/token_classification.md

@@ -32,8 +32,7 @@ The task illustrated in this tutorial is supported by the following model archit
 
 <!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
 
-[ALBERT](../model_doc/albert), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LiLT](../model_doc/lilt), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MPT](../model_doc/mpt), [MRA](../model_doc/mra), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [QDQBert](../model_doc/qdqbert), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
-
+[ALBERT](../model_doc/albert), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [BROS](../model_doc/bros), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LiLT](../model_doc/lilt), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MPT](../model_doc/mpt), [MRA](../model_doc/mra), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [QDQBert](../model_doc/qdqbert), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
 
 <!--End of the generated tip-->
 

docs/source/en/tasks/zero_shot_image_classification.md

@@ -75,7 +75,7 @@ include a local path to an image or an image url.
 The candidate labels can be simple words like in this example, or more descriptive.
 
 ```py
->>> predictions = classifier(image, candidate_labels=["fox", "bear", "seagull", "owl"])
+>>> predictions = detector(image, candidate_labels=["fox", "bear", "seagull", "owl"])
 >>> predictions
 [{'score': 0.9996670484542847, 'label': 'owl'},
  {'score': 0.000199399160919711, 'label': 'seagull'},

docs/source/en/transformers_agents.md

@@ -14,11 +14,11 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# Transformers Agent
+# Transformers Agents
 
 <Tip warning={true}>
 
-Transformers Agent is an experimental API which is subject to change at any time. Results returned by the agents
+Transformers Agents is an experimental API which is subject to change at any time. Results returned by the agents
 can vary as the APIs or underlying models are prone to change.
 
 </Tip>

docs/source/es/tasks/language_modeling.md

@@ -122,7 +122,7 @@ Así es como puedes crear una función de preprocesamiento para convertir la lis
 ...     return tokenizer([" ".join(x) for x in examples["answers.text"]], truncation=True)
 ```
 
-Usa de 🤗 Datasets la función [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) para aplicar la función de preprocesamiento sobre el dataset en su totalidad. Puedes acelerar la función `map` configurando el argumento `batched=True` para procesar múltiples elementos del dataset a la vez y aumentar la cantidad de procesos con `num_proc`. Elimina las columnas que no necesitas:
+Usa de 🤗 Datasets la función [`map`](https://huggingface.co/docs/datasets/process#map) para aplicar la función de preprocesamiento sobre el dataset en su totalidad. Puedes acelerar la función `map` configurando el argumento `batched=True` para procesar múltiples elementos del dataset a la vez y aumentar la cantidad de procesos con `num_proc`. Elimina las columnas que no necesitas:
 
 ```py
 >>> tokenized_eli5 = eli5.map(

docs/source/ko/_toctree.yml

@@ -25,6 +25,8 @@
     title: 만든 모델 공유하기
   - local: transformers_agents
     title: 에이전트
+  - local: llm_tutorial
+    title: 대규모 언어 모델로 생성하기
   title: 튜토리얼
 - sections:
   - sections:
@@ -47,11 +49,11 @@
     title: 자연어처리
     isExpanded: false
   - sections:
-      - local: in_translation
-        title: (번역중) Audio classification
+      - local: tasks/audio_classification
+        title: 오디오 분류
       - local: tasks/asr
         title: 자동 음성 인식
-    title: (번역중) 오디오
+    title: 오디오
     isExpanded: false
   - sections:
       - local: tasks/image_classification
@@ -115,8 +117,8 @@
       title: 성능 및 확장성
     - local: in_translation
       title: (번역중) Training on one GPU
-    - local: in_translation
-      title: (번역중) Training on many GPUs
+    - local: perf_train_gpu_many
+      title: 다중 GPU에서 훈련 진행하기
     - local: perf_train_cpu
       title: CPU에서 훈련
     - local: perf_train_cpu_many
@@ -139,18 +141,18 @@
       title: 훈련용 사용자 맞춤형 하드웨어
     - local: in_translation
       title: (번역중) Instantiating a big model
-    - local: in_translation
-      title: (번역중) Debugging
+    - local: debugging
+      title: 디버깅
     - local: hpo_train
       title: Trainer API를 사용한 하이퍼파라미터 탐색
     - local: tf_xla
       title: TensorFlow 모델을 위한 XLA 통합
   title: (번역중) 성능 및 확장성
 - sections:
-    - local: in_translation
-      title: (번역중) How to contribute to transformers?
+    - local: contributing
+      title: 🤗 Transformers에 기여하는 방법
     - local: add_new_model
-      title: 🤗 Transformers에 새로운 모델을 추가하는 방법 
+      title: 🤗 Transformers에 새로운 모델을 추가하는 방법
     - local: add_tensorflow_model
       title: 어떻게 🤗 Transformers 모델을 TensorFlow로 변환하나요?
     - local: add_new_pipeline
@@ -172,8 +174,8 @@
     title: 🤗 Transformers로 작업을 해결하는 방법
   - local: model_summary
     title: Transformer 모델군
-  - local: in_translation
-    title: (번역중) Summary of the tokenizers
+  - local: tokenizer_summary
+    title: 토크나이저 요약
   - local: attention
     title: 어텐션 매커니즘
   - local: pad_truncation
@@ -335,8 +337,10 @@
         title: (번역중) Jukebox
       - local: in_translation
         title: (번역중) LED
-      - local: in_translation
-        title: (번역중) LLaMA
+      - local: model_doc/llama
+        title: LLaMA
+      - local: model_doc/llama2
+        title: LLaMA2
       - local: in_translation
         title: (번역중) Longformer
       - local: in_translation
@@ -563,8 +567,8 @@
         title: (번역중) Wav2Vec2Phoneme
       - local: in_translation
         title: (번역중) WavLM
-      - local: in_translation
-        title: (번역중) Whisper
+      - local: model_doc/whisper
+        title: Whisper
       - local: in_translation
         title: (번역중) XLS-R
       - local: in_translation

docs/source/ko/add_tensorflow_model.md

@@ -175,7 +175,6 @@ git push -u origin add_tf_brand_new_bert
 
 - `src/transformers/__init__.py`에 *BrandNewBert*의 모든 공개 클래스를 포함합니다.
 - `src/transformers/models/auto/modeling_tf_auto.py`에서 *BrandNewBert* 클래스를 해당 Auto 클래스에 추가합니다.
-- `utils/documentation_tests.txt`에 모델 파일을 문서화하는 테스트 파일 목록을 추가합니다.
 - `src/transformers/utils/dummy_tf_objects.py`에 *BrandNewBert*와 관련된 레이지 로딩 클래스를 추가합니다.
 - `src/transformers/models/brand_new_bert/__init__.py`에서 공개 클래스에 대한 import 구조를 업데이트합니다.
 - `docs/source/en/model_doc/brand_new_bert.md`에서 *BrandNewBert*의 공개 메서드에 대한 문서 포인터를 추가합니다.

docs/source/ko/contributing.md

@@ -0,0 +1,332 @@
+<!---
+Copyright 2020 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.
+-->
+
+# 🤗 Transformers에 기여하기 [[contribute-to-transformers]]
+
+누구나 🤗 Transformers에 기여할 수 있으며, 우리는 모든 사람의 기여를 소중히 생각합니다. 코드 기여는 커뮤니티를 돕는 유일한 방법이 아닙니다. 질문에 답하거나 다른 사람을 도와 문서를 개선하는 것도 매우 가치가 있습니다.
+
+🤗 Transformers를 널리 알리는 것도 큰 도움이 됩니다! 멋진 프로젝트들을 가능하게 한 🤗 Transformers 라이브러리에 대해 블로그 게시글에 언급하거나, 도움이 되었을 때마다 Twitter에 알리거나, 저장소에 ⭐️ 를 표시하여 감사 인사를 전해주세요.
+
+어떤 방식으로 기여하든 [행동 규칙](https://github.com/huggingface/transformers/blob/main/CODE_OF_CONDUCT.md)을 숙지하고 존중해주세요.
+
+**이 안내서는 멋진 [scikit-learn 기여 안내서](https://github.com/scikit-learn/scikit-learn/blob/main/CONTRIBUTING.md)에서 큰 영감을 받았습니다.**
+
+## 기여하는 방법 [[ways-to-contribute]]
+
+여러 가지 방법으로 🤗 Transformers에 기여할 수 있습니다:
+
+* 기존 코드의 미해결된 문제를 수정합니다.
+* 버그 또는 새로 추가되길 원하는 기능과 관련된 이슈를 제출합니다.
+* 새로운 모델을 구현합니다.
+* 예제나 문서에 기여합니다.
+
+어디서부터 시작할지 모르겠다면, [Good First Issue](https://github.com/huggingface/transformers/contribute) 목록을 확인해보세요. 이 목록은 초보자도 참여하기 쉬운 오픈 이슈 목록을 제공하며, 당신이 오픈소스에 처음으로 기여하는 데 큰 도움이 될 것입니다. 그저 작업하고 싶은 이슈에 댓글만 달아주면 됩니다.
+
+조금 더 도전적인 작업을 원한다면, [Good Second Issue](https://github.com/huggingface/transformers/labels/Good%20Second%20Issue) 목록도 확인해보세요. 이미 당신이 잘 하고 있다고 생각되더라도, 한 번 시도해보세요! 우리도 여러분을 도울 것입니다. 🚀
+
+> 커뮤니티에 이루어지는 모든 기여는 똑같이 소중합니다. 🥰
+
+## 미해결된 문제 수정하기 [[fixing-outstanding-issues]]
+
+기존 코드에서 발견한 문제점에 대한 해결책이 떠오른 경우, 언제든지 [기여를 시작](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md/#create-a-pull-request)하고 Pull Request를 생성해주세요!
+
+## 버그 관련 이슈를 제기하거나 새로운 기능 요청하기 [[submitting-a-bugrelated-issue-or-feature-request]]
+
+버그 관련 이슈를 제기하거나 새로운 기능을 요청할 때는 다음 가이드라인을 최대한 준수해주세요. 이렇게 하면 좋은 피드백과 함께 빠르게 답변해 드릴 수 있습니다.
+
+### 버그를 발견하셨나요? [[did-you-find-a-bug]]
+
+🤗 Transformers 라이브러리는 사용 중에 겪는 문제를 보고해주는 사용자들 덕분에 더욱 견고해지고 신뢰할 수 있게 되었습니다.
+
+이슈를 보고하기 전에, 버그가 이미 **보고되지 않았는지** 확인해주세요. (GitHub의 이슈 탭 아래의 검색 바를 사용하세요). 이슈는 라이브러리 자체에서 발생한 버그어야 하며, 코드의 다른 부분과 관련된 것이 아니어야 합니다. 버그가 라이브러리의 문제로 발생하였는지 확실하지 않은 경우 먼저 [포럼](https://discuss.huggingface.co/)에서 질문해 주세요. 이렇게 하면 일반적인 질문보다 라이브러리와 관련된 문제를 더 빠르게 해결할 수 있습니다.
+
+버그가 이미 보고되지 않았다는 것을 확인했다면, 다음 정보를 포함하여 이슈를 제출해 주세요. 그러면 우리가 빠르게 해결할 수 있습니다:
+
+* 사용 중인 **운영체제 종류와 버전**, 그리고 **Python**, **PyTorch** 또는 **TensorFlow** 버전.
+* 버그를 30초 이내로 재현할 수 있는 간단하고 독립적인 코드 스니펫.
+* 예외가 발생한 경우 *전체* 트레이스백.
+* 스크린샷과 같이 도움이 될 것으로 생각되는 추가 정보를 첨부해 주세요.
+
+운영체제와 소프트웨어 버전을 자동으로 가져오려면 다음 명령을 실행하세요:
+
+```bash
+transformers-cli env
+```
+
+저장소의 루트 디렉터리에서도 같은 명령을 실행할 수 있습니다:
+
+```bash
+python src/transformers/commands/transformers_cli.py env
+```
+
+
+### 새로운 기능을 원하시나요? [[do-you-want-a-new-feature]]
+
+🤗 Transformers에서 사용하고 싶은 새로운 기능이 있다면, 다음 내용을 포함하여 이슈를 제출해 주세요:
+
+1. 이 기능이 필요한 *이유*는 무엇인가요? 라이브러리에 대한 문제나 불만과 관련이 있나요? 프로젝트에 필요한 기능인가요? 커뮤니티에 도움이 될 만한 기능인가요?
+
+   어떤 내용이든 여러분의 이야기를 듣고 싶습니다!
+
+2. 요청하는 기능을 최대한 자세히 설명해 주세요. 더 많은 정보를 제공할수록 더 나은 도움을 드릴 수 있습니다.
+3. 해당 기능의 사용법을 보여주는 *코드 스니펫*을 제공해 주세요.
+4. 기능과 관련된 논문이 있는 경우 링크를 포함해 주세요.
+
+이슈가 잘 작성되었다면 이슈가 생성된 순간, 이미 80% 정도의 작업이 완료된 것입니다. 
+
+이슈를 제기하는 데 도움이 될 만한 [템플릿](https://github.com/huggingface/transformers/tree/main/templates)도 준비되어 있습니다.
+
+## 새로운 모델을 구현하고 싶으신가요? [[do-you-want-to-implement-a-new-model]]
+
+새로운 모델은 계속해서 출시됩니다. 만약 여러분이 새로운 모델을 구현하고 싶다면 다음 정보를 제공해 주세요.
+
+* 모델에 대한 간단한 설명과 논문 링크.
+* 구현이 공개되어 있다면 구현 링크.
+* 모델 가중치가 사용 가능하다면 가중치 링크.
+
+만약 모델을 직접 기여하고 싶으시다면, 알려주세요. 🤗 Transformers에 추가할 수 있도록 도와드리겠습니다!
+
+새로운 모델을 추가하는 방법에 대한 [상세 안내서와 템플릿](https://github.com/huggingface/transformers/tree/main/templates)을 제공하고 있으며, [🤗 Transformers에 새로운 모델을 추가하는 방법](https://huggingface.co/docs/transformers/add_new_model)에 대한 기술적인 안내서도 있습니다.
+
+## 문서를 추가하고 싶으신가요? [[do-you-want-to-add-documentation]]
+
+우리는 언제나 더 명확하고 정확한 문서를 제공하기 위하여 개선점을 찾고 있습니다. 오탈자나 부족한 내용, 분명하지 않거나 부정확한 내용 등을 알려주시면 개선하는 데 도움이 됩니다. 관심이 있으시다면 변경하거나 기여하실 수 있도록 도와드리겠습니다!
+
+문서를 생성, 빌드 및 작성하는 방법에 대한 자세한 내용은 [README](https://github.com/huggingface/transformers/tree/main/docs) 문서를 확인해 주세요.
+
+## 풀 리퀘스트(Pull Request) 생성하기 [[create-a-pull-request]]
+
+코드를 작성하기 전에 기존의 Pull Request나 이슈를 검색하여 누군가 이미 동일한 작업을 하고 있는지 확인하는 것이 좋습니다. 확실하지 않다면 피드백을 받기 위해 이슈를 열어보는 것이 좋습니다.
+
+🤗 Transformers에 기여하기 위해서는 기본적인 `git` 사용 능력이 필요합니다. `git`은 사용하기 쉬운 도구는 아니지만, 매우 훌륭한 매뉴얼을 제공합니다. 쉘(shell)에서 `git --help`을 입력하여 확인해보세요! 만약 책을 선호한다면, [Pro Git](https://git-scm.com/book/en/v2)은 매우 좋은 참고 자료가 될 것입니다.
+
+🤗 Transformers에 기여하려면 **[Python 3.8]((https://github.com/huggingface/transformers/blob/main/setup.py#L426))** 이상의 버전이 필요합니다. 기여를 시작하려면 다음 단계를 따르세요:
+
+1. 저장소 페이지에서 **[Fork](https://github.com/huggingface/transformers/fork)** 버튼을 클릭하여 저장소를 포크하세요. 이렇게 하면 코드의 복사본이 여러분의 GitHub 사용자 계정 아래에 생성됩니다.
+
+2. 포크한 저장소를 로컬 디스크로 클론하고, 기본 저장소를 원격(remote)으로 추가하세요:
+
+   ```bash
+   git clone git@github.com:<your Github handle>/transformers.git
+   cd transformers
+   git remote add upstream https://github.com/huggingface/transformers.git
+   ```
+
+3. 개발 변경 사항을 저장할 새 브랜치를 생성하세요:
+
+   ```bash
+   git checkout -b a-descriptive-name-for-my-changes
+   ```
+
+   🚨 절대 `main` 브랜치에서 작업하지 **마세요!**
+
+4. 가상 환경에서 다음 명령을 실행하여 개발 환경을 설정하세요:
+
+   ```bash
+   pip install -e ".[dev]"
+   ```
+
+   만약 이미 가상 환경에 🤗 Transformers가 설치되어 있다면, `-e` 플래그를 사용하여 설치하기 전에 `pip uninstall transformers`로 제거해주세요.
+   
+   여러분의 운영체제에 따라서, 그리고 🤗 Transformers의 선택적 의존성의 수가 증가하면서, 이 명령이 실패할 수도 있습니다. 그럴 경우 사용하려는 딥러닝 프레임워크(PyTorch, TensorFlow, 그리고/또는 Flax)를 설치한 후 아래 명령을 실행해주세요:
+
+   ```bash
+   pip install -e ".[quality]"
+   ```
+
+   대부분의 경우 이것으로 충분할 것입니다.
+
+5. 브랜치에서 기능을 개발하세요.
+
+   코드를 작업하는 동안 테스트 스위트(test suite)가 통과하는지 확인하세요. 다음과 같이 변경 사항에 영향을 받는 테스트를 실행하세요:
+
+   ```bash
+   pytest tests/<TEST_TO_RUN>.py
+   ```
+
+   테스트에 대한 더 많은 정보는 [테스트](https://huggingface.co/docs/transformers/testing) 가이드를 확인하세요.
+
+   🤗 Transformers는 `black`과 `ruff`를 사용하여 소스 코드의 형식을 일관되게 유지합니다. 변경 사항을 적용한 후에는 다음 명령으로 자동으로 스타일 교정 및 코드 검증을 수행하세요:
+
+   ```bash
+   make fixup
+   ```
+
+   이것은 또한 작업 중인 PR에서 수정한 파일에서만 작동하도록 최적화되어 있습니다.
+
+   검사를 하나씩 실행하려는 경우, 다음 명령으로 스타일 교정을 적용할 수 있습니다:
+
+   ```bash
+   make style
+   ```
+
+   🤗 Transformers는 또한 `ruff`와 몇 가지 사용자 정의 스크립트를 사용하여 코딩 실수를 확인합니다. CI를 통해 품질 관리가 수행되지만, 다음 명령으로 동일한 검사를 실행할 수 있습니다:
+
+   ```bash
+   make quality
+   ```
+
+   마지막으로, 새 모델을 추가할 때 일부 파일을 업데이트하는 것을 잊지 않도록 하기 위한 많은 스크립트가 있습니다. 다음 명령으로 이러한 스크립트를 실행할 수 있습니다:
+
+   ```bash
+   make repo-consistency
+   ```
+
+   이러한 검사에 대해 자세히 알아보고 관련 문제를 해결하는 방법은 [Pull Request에 대한 검사](https://huggingface.co/docs/transformers/pr_checks) 가이드를 확인하세요.
+
+   만약 `docs/source` 디렉터리 아래의 문서를 수정하는 경우, 문서가 빌드될 수 있는지 확인하세요. 이 검사는 Pull Request를 열 때도 CI에서 실행됩니다. 로컬 검사를 실행하려면 문서 빌더를 설치해야 합니다:
+   
+   ```bash
+   pip install ".[docs]"
+   ```
+
+   저장소의 루트 디렉터리에서 다음 명령을 실행하세요:
+
+   ```bash
+   doc-builder build transformers docs/source/en --build_dir ~/tmp/test-build
+   ```
+
+   이 명령은 `~/tmp/test-build` 폴더에 문서를 빌드하며, 생성된 Markdown 파일을 선호하는 편집기로 확인할 수 있습니다. Pull Request를 열 때 GitHub에서 문서를 미리 볼 수도 있습니다.
+
+   변경 사항에 만족하면 `git add`로 변경된 파일을 추가하고, `git commit`으로 변경 사항을 로컬에 기록하세요:
+
+   ```bash
+   git add modified_file.py
+   git commit
+   ```
+
+   [좋은 커밋 메시지](https://chris.beams.io/posts/git-commit/)를 작성하여 변경 사항을 명확하게 전달하세요!
+
+   변경 사항을 프로젝트 원본 저장소와 동기화하려면, PR을 *열기 전에* 브랜치를 `upstream/branch`로 리베이스(rebase)하세요. 또는 관리자의 요청에 이 작업이 필요할 수 있습니다:
+
+   ```bash
+   git fetch upstream
+   git rebase upstream/main
+   ```
+   
+   변경 사항을 브랜치에 푸시하세요:
+
+   ```bash
+   git push -u origin a-descriptive-name-for-my-changes
+   ```
+
+   이미 PR을 열었다면, `--force` 플래그와 함께 강제 푸시해야 합니다. 아직 PR이 열리지 않았다면 정상적으로 변경 사항을 푸시하면 됩니다.
+
+6. 이제 GitHub에서 포크한 저장소로 이동하고 **Pull request(풀 리퀘스트)**를 클릭하여 Pull Request를 열 수 있습니다. 아래의 [체크리스트](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md/#pull-request-checklist)에서 모든 항목에 체크 표시를 하세요. 준비가 완료되면 프로젝트 관리자에게 변경 사항을 보내 검토를 요청할 수 있습니다.
+
+7. 관리자가 변경 사항을 요청해도 괜찮습니다. 핵심 기여자들도 동일한 상황을 겪습니다! 모두가 변경 사항을 Pull Request에서 볼 수 있도록, 로컬 브랜치에서 작업하고 변경 사항을 포크한 저장소로 푸시하세요. 그러면 변경 사항이 자동으로 Pull Request에 나타납니다.
+
+### Pull Request 체크리스트 [[pull-request-checklist]]
+
+☐ Pull Request 제목은 기여 내용을 요약해야 합니다.<br>
+☐ Pull Request가 이슈를 해결하는 경우, Pull Request 설명에 이슈 번호를 언급하여 연관되어 있음을 알려주세요. (이슈를 확인하는 사람들이 해당 이슈에 대한 작업이 진행 중임을 알 수 있게 합니다).<br>
+☐ 작업이 진행중이라면 제목 앞에 `[WIP]`를 붙여주세요. 중복 작업을 피하고 병합할 준비가 된 PR과 구분하기에 유용합니다.<br>
+☐ 기존 테스트를 통과하는지 확인하세요.<br>
+☐ 새로운 기능을 추가하는 경우, 해당 기능에 대한 테스트도 추가하세요.<br>
+   - 새 모델을 추가하는 경우, `ModelTester.all_model_classes = (MyModel, MyModelWithLMHead,...)`을 사용하여 일반적인 테스트를 활성화하세요.
+   - 새 `@slow` 테스트를 추가하는 경우, 다음 명령으로 테스트를 통과하는지 확인하세요: `RUN_SLOW=1 python -m pytest tests/models/my_new_model/test_my_new_model.py`.
+   - 새 토크나이저를 추가하는 경우, 테스트를 작성하고 다음 명령으로 테스트를 통과하는지 확인하세요: `RUN_SLOW=1 python -m pytest tests/models/{your_model_name}/test_tokenization_{your_model_name}.py`. 
+   - CircleCI에서는 느린 테스트를 실행하지 않지만, GitHub Actions에서는 매일 밤 실행됩니다!<br>
+
+☐ 모든 공개 메소드는 유용한 기술문서를 가져야 합니다 (예를 들어 [`modeling_bert.py`](https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py) 참조).<br>
+☐ 저장소가 빠르게 성장하고 있으므로 저장소에 상당한 부담을 주는 이미지, 동영상 및 기타 텍스트가 아닌 파일은 추가하지 마세요. 대신 [`hf-internal-testing`](https://huggingface.co/hf-internal-testing)과 같은 Hub 저장소를 사용하여 이러한 파일을 호스팅하고 URL로 참조하세요. 문서와 관련된 이미지는 다음 저장소에 배치하는 것을 권장합니다: [huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images). 이 데이터셋 저장소에서 PR을 열어서 Hugging Face 멤버에게 병합을 요청할 수 있습니다.
+
+Pull Request에서 실행되는 검사에 대한 자세한 정보는 [Pull Request에 대한 검사](https://huggingface.co/docs/transformers/pr_checks) 가이드를 확인하세요.
+
+### 테스트 [[tests]]
+
+라이브러리 동작과 여러 예제를 테스트할 수 있는 광범위한 테스트 스위트가 포함되어 있습니다. 라이브러리 테스트는 [tests](https://github.com/huggingface/transformers/tree/main/tests) 폴더에, 예제 테스트는 [examples](https://github.com/huggingface/transformers/tree/main/examples) 폴더에 있습니다.
+
+속도가 빠른 `pytest`와 `pytest-xdist`를 선호합니다. 저장소의 루트 디렉터리에서 테스트를 실행할 *하위 폴더 경로 또는 테스트 파일 경로*를 지정하세요.
+
+```bash
+python -m pytest -n auto --dist=loadfile -s -v ./tests/models/my_new_model
+```
+
+마찬가지로 `examples` 디렉터리에서도 *하위 폴더 경로 또는 테스트 파일 경로*를 지정하세요. 예를 들어, 다음 명령은 PyTorch `examples` 디렉터리의 텍스트 분류 하위 폴더를 테스트합니다:
+
+```bash
+pip install -r examples/xxx/requirements.txt  # only needed the first time
+python -m pytest -n auto --dist=loadfile -s -v ./examples/pytorch/text-classification
+```
+
+이것이 실제로 `make test` 및 `make test-examples` 명령이 구현되는 방식입니다 (`pip install`은 제외합니다)!
+
+또한 특정 기능만 테스트하기 위한 더 작은 테스트를 지정할 수 있습니다.
+
+기본적으로 느린 테스트는 건너뛰지만 `RUN_SLOW` 환경 변수를 `yes`로 설정하여 실행할 수 있습니다. 이렇게 하면 많은 기가바이트 단위의 모델이 다운로드되므로 충분한 디스크 공간, 좋은 인터넷 연결과 많은 인내가 필요합니다!
+
+<Tip warning={true}>
+
+테스트를 실행하려면 *하위 폴더 경로 또는 테스트 파일 경로*를 지정하세요. 그렇지 않으면 `tests` 또는 `examples` 폴더의 모든 테스트를 실행하게 되어 매우 긴 시간이 걸립니다!
+
+</Tip>
+
+```bash
+RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./tests/models/my_new_model
+RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./examples/pytorch/text-classification
+```
+
+느린 테스트와 마찬가지로, 다음과 같이 테스트 중에 기본적으로 활성화되지 않는 다른 환경 변수도 있습니다:
+- `RUN_CUSTOM_TOKENIZERS`: 사용자 정의 토크나이저 테스트를 활성화합니다.
+- `RUN_PT_FLAX_CROSS_TESTS`: PyTorch + Flax 통합 테스트를 활성화합니다.
+- `RUN_PT_TF_CROSS_TESTS`: TensorFlow + PyTorch 통합 테스트를 활성화합니다.
+
+더 많은 환경 변수와 추가 정보는 [testing_utils.py](src/transformers/testing_utils.py)에서 찾을 수 있습니다.
+
+🤗 Transformers는 테스트 실행기로 `pytest`를 사용합니다. 그러나 테스트 스위트 자체에서는 `pytest` 관련 기능을 사용하지 않습니다.
+
+이것은 `unittest`가 완전히 지원된다는 것을 의미합니다. 다음은 `unittest`로 테스트를 실행하는 방법입니다:
+
+```bash
+python -m unittest discover -s tests -t . -v
+python -m unittest discover -s examples -t examples -v
+```
+
+### 스타일 가이드 [[style-guide]]
+
+문서는 [Google Python 스타일 가이드](https://google.github.io/styleguide/pyguide.html)를 따릅니다. 자세한 정보는 [문서 작성 가이드](https://github.com/huggingface/transformers/tree/main/docs#writing-documentation---specification)를 확인하세요.
+
+### Windows에서 개발 [[develop-on-windows]]
+
+Windows에서 개발할 경우([Windows Subsystem for Linux](https://learn.microsoft.com/en-us/windows/wsl/) 또는 WSL에서 작업하지 않는 한) Windows `CRLF` 줄 바꿈을 Linux `LF` 줄 바꿈으로 변환하도록 git을 구성해야 합니다:
+
+```bash
+git config core.autocrlf input
+```
+
+Windows에서 `make` 명령을 실행하는 한 가지 방법은 MSYS2를 사용하는 것입니다:
+
+1. [MSYS2](https://www.msys2.org/)를 다운로드합니다. `C:\msys64`에 설치되었다고 가정합니다.
+2. CLI에서 `C:\msys64\msys2.exe`를 엽니다 (시작 메뉴에서 사용 가능해야 함).
+3. 쉘에서 다음을 실행하여: `pacman -Syu` 및 `pacman -S make`로 `make`를 설치합니다.
+4. 환경 변수 PATH에 `C:\msys64\usr\bin`을 추가하세요.
+
+이제 모든 터미널 (Powershell, cmd.exe 등)에서 `make`를 사용할 수 있습니다! 🎉
+
+### 포크한 저장소를 상위 원본 브랜치(main)과 동기화하기 (Hugging Face 저장소) [[sync-a-forked-repository-with-upstream-main-the-hugging-face-repository]]
+
+포크한 저장소의 main 브랜치를 업데이트할 때, 다음 단계를 따라 수행해주세요. 이렇게 하면 각 upstream PR에 참조 노트가 추가되는 것을 피하고 이러한 PR에 관여하는 개발자들에게 불필요한 알림이 전송되는 것을 방지할 수 있습니다.
+
+1. 가능하면 포크된 저장소의 브랜치 및 PR을 사용하여 upstream과 동기화하지 마세요. 대신 포크된 main 저장소에 직접 병합하세요.
+2. PR이 반드시 필요한 경우, 브랜치를 확인한 후 다음 단계를 사용하세요:
+
+```bash
+git checkout -b your-branch-for-syncing
+git pull --squash --no-commit upstream main
+git commit -m '<your message without GitHub references>'
+git push --set-upstream origin your-branch-for-syncing
+```

docs/source/ko/custom_tools.md

@@ -20,7 +20,7 @@ Transformers와 관련하여 어떤 도구와 에이전트가 있는지 잘 모
 
 <Tip warning={true}>
 
-Transformers Agent는 실험 중인 API로 언제든지 변경될 수 있습니다. 
+Transformers Agents는 실험 중인 API로 언제든지 변경될 수 있습니다. 
 API 또는 기반 모델이 변경되기 쉽기 때문에 에이전트가 반환하는 결과도 달라질 수 있습니다.
 
 </Tip>

docs/source/ko/debugging.md

@@ -0,0 +1,306 @@
+<!--Copyright 2021 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# 디버깅 [[debugging]]
+
+## Multi-GPU 네트워크 문제 디버그 [[multigpu-network-issues-debug]]
+
+`DistributedDataParallel` 및 다중 GPU를 사용하여 훈련하거나 추론할 때, 프로세스 및/또는 노드 간의 상호 통신 문제가 발생하는 경우, 다음 스크립트를 사용하여 네트워크 문제를 진단할 수 있습니다.
+
+```bash
+wget https://raw.githubusercontent.com/huggingface/transformers/main/scripts/distributed/torch-distributed-gpu-test.py
+```
+
+예를 들어, 2개의 GPU가 상호 작용하는 방식을 테스트하려면 다음을 실행하세요:
+
+```bash
+python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py
+```
+두 프로세스가 서로 통신하고 GPU 메모리를 할당하는 경우, 각각 "OK" 상태를 출력합니다.
+
+더 많은 GPU 또는 노드의 경우 스크립트의 인수를 조정하면 됩니다.
+
+진단 스크립트 내에서 더 많은 세부 정보와 SLURM 환경에서 실행하는 방법에 대한 레시피를 찾을 수 있습니다.
+
+추가적인 디버그 수준은 다음과 같이 `NCCL_DEBUG=INFO` 환경 변수를 추가하는 것입니다:
+
+```bash
+NCCL_DEBUG=INFO python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py
+```
+
+이렇게 하면 NCCL 관련 디버그 정보가 많이 출력되며, 문제가 보고된 경우에는 인터넷에서 검색할 수 있습니다. 또는 출력을 해석하는 방법을 잘 모르는 경우 로그 파일을 이슈에 공유할 수 있습니다.
+
+
+
+## 언더플로 및 오버플로 감지 [[underflow-and-overflow-detection]]
+
+
+<Tip>
+
+이 기능은 현재 PyTorch에서만 사용할 수 있습니다.
+
+</Tip>
+
+<Tip>
+
+다중 GPU 훈련을 위해서는 DDP (`torch.distributed.launch`)가 필요합니다.
+
+</Tip>
+
+<Tip>
+
+이 기능은 `nn.Module`을 기반으로 하는 모델과 함께 사용할 수 있습니다.
+
+</Tip>
+
+`loss=NaN`이 나타나거나 모델이 `inf` 또는 `nan`으로 인해 다른 이상한 동작을 하는 경우, 언더플로 또는 오버플로의 첫 번째 발생 위치와 그 원인을 파악해야 합니다. 다행히도 이를 자동으로 감지하는 특수 모듈을 활성화하여 쉽게 알아낼 수 있습니다.
+
+[`Trainer`]를 사용하는 경우, 다음을 기존의 명령줄 인수에 추가하면 됩니다.
+
+```bash
+--debug underflow_overflow
+```
+또는 [`TrainingArguments`] 객체를 생성할 때 `debug="underflow_overflow"`를 전달합니다.
+
+자체 훈련 루프나 다른 Trainer를 사용하는 경우, 다음과 같이 수행할 수 있습니다. 
+
+```python
+from transformers.debug_utils import DebugUnderflowOverflow
+
+debug_overflow = DebugUnderflowOverflow(model)
+```
+
+[`~debug_utils.DebugUnderflowOverflow`]는 모델에 후크를 삽입하여 각 forward 호출 직후에 입력 및 출력 변수 및 해당 모듈의 가중치를 테스트합니다. 활성화나 가중치의 최소한 하나의 요소에서 `inf` 또는 `nan`이 감지되면 프로그램이 어설트되고 다음과 같은 보고서가 출력됩니다. (이 예제는 fp16 혼합 정밀도에서 `google/mt5-small`에서 캡처된 것입니다):
+
+```
+Detected inf/nan during batch_number=0
+Last 21 forward frames:
+abs min  abs max  metadata
+                  encoder.block.1.layer.1.DenseReluDense.dropout Dropout
+0.00e+00 2.57e+02 input[0]
+0.00e+00 2.85e+02 output
+[...]
+                  encoder.block.2.layer.0 T5LayerSelfAttention
+6.78e-04 3.15e+03 input[0]
+2.65e-04 3.42e+03 output[0]
+             None output[1]
+2.25e-01 1.00e+04 output[2]
+                  encoder.block.2.layer.1.layer_norm T5LayerNorm
+8.69e-02 4.18e-01 weight
+2.65e-04 3.42e+03 input[0]
+1.79e-06 4.65e+00 output
+                  encoder.block.2.layer.1.DenseReluDense.wi_0 Linear
+2.17e-07 4.50e+00 weight
+1.79e-06 4.65e+00 input[0]
+2.68e-06 3.70e+01 output
+                  encoder.block.2.layer.1.DenseReluDense.wi_1 Linear
+8.08e-07 2.66e+01 weight
+1.79e-06 4.65e+00 input[0]
+1.27e-04 2.37e+02 output
+                  encoder.block.2.layer.1.DenseReluDense.dropout Dropout
+0.00e+00 8.76e+03 input[0]
+0.00e+00 9.74e+03 output
+                  encoder.block.2.layer.1.DenseReluDense.wo Linear
+1.01e-06 6.44e+00 weight
+0.00e+00 9.74e+03 input[0]
+3.18e-04 6.27e+04 output
+                  encoder.block.2.layer.1.DenseReluDense T5DenseGatedGeluDense
+1.79e-06 4.65e+00 input[0]
+3.18e-04 6.27e+04 output
+                  encoder.block.2.layer.1.dropout Dropout
+3.18e-04 6.27e+04 input[0]
+0.00e+00      inf output
+```
+
+예제 출력은 간략성을 위해 중간 부분이 잘려 있습니다.
+
+두 번째 열은 절대적으로 가장 큰 요소의 값이며, 따라서 마지막 몇 개의 프레임을 자세히 살펴보면 입력과 출력이 `1e4` 범위에 있음을 알 수 있습니다. 따라서 이 훈련은 `fp16` 혼합 정밀도로 수행될 때 가장 마지막 단계에서 오버플로우가 발생했습니다 (`fp16`에서 `inf` 이전의 가장 큰 숫자는 `64e3`입니다). `fp16` 아래에서 오버플로우를 피하기 위해서는 활성화는 `1e4`보다 훨씬 작아야 합니다. 왜냐하면 `1e4 * 1e4 = 1e8`이기 때문에 큰 활성화와의 행렬 곱은 수치적인 오버플로우 조건으로 이어질 것입니다.
+
+추적의 맨 처음에서 어느 배치 번호에서 문제가 발생했는지 알 수 있습니다 (여기서 `Detected inf/nan during batch_number=0`은 문제가 첫 번째 배치에서 발생했음을 의미합니다).
+
+각 보고된 프레임은 해당 프레임이 보고하는 해당 모듈에 대한 완전한 항목을 선언하며, 이 프레임만 살펴보면 다음과 같습니다.
+
+```
+                  encoder.block.2.layer.1.layer_norm T5LayerNorm
+8.69e-02 4.18e-01 weight
+2.65e-04 3.42e+03 input[0]
+1.79e-06 4.65e+00 output
+```
+
+여기서 `encoder.block.2.layer.1.layer_norm`은 인코더의 두 번째 블록의 첫 번째 레이어에 대한 레이어 정규화를 의미하며, `forward`의 특정 호출은 `T5LayerNorm`입니다.
+
+이 보고서의 마지막 몇 개 프레임을 살펴보겠습니다:
+
+```
+Detected inf/nan during batch_number=0
+Last 21 forward frames:
+abs min  abs max  metadata
+[...]
+                  encoder.block.2.layer.1.DenseReluDense.wi_0 Linear
+2.17e-07 4.50e+00 weight
+1.79e-06 4.65e+00 input[0]
+2.68e-06 3.70e+01 output
+                  encoder.block.2.layer.1.DenseReluDense.wi_1 Linear
+8.08e-07 2.66e+01 weight
+1.79e-06 4.65e+00 input[0]
+1.27e-04 2.37e+02 output
+                  encoder.block.2.layer.1.DenseReluDense.wo Linear
+1.01e-06 6.44e+00 weight
+0.00e+00 9.74e+03 input[0]
+3.18e-04 6.27e+04 output
+                  encoder.block.2.layer.1.DenseReluDense T5DenseGatedGeluDense
+1.79e-06 4.65e+00 input[0]
+3.18e-04 6.27e+04 output
+                  encoder.block.2.layer.1.dropout Dropout
+3.18e-04 6.27e+04 input[0]
+0.00e+00      inf output
+```
+
+마지막 프레임은 `Dropout.forward` 함수에 대한 보고입니다. 첫 번째 항목은 유일한 입력을 나타내고 두 번째 항목은 유일한 출력을 나타냅니다. 이 함수가 `DenseReluDense` 클래스 내부의 `dropout` 속성에서 호출된 것을 볼 수 있습니다. 이는 첫 번째 레이어의 두 번째 블록에서 첫 번째 배치 중에 발생했다는 것을 알 수 있습니다. 마지막으로, 절대적으로 가장 큰 입력 요소는 `6.27e+04`이고 출력도 마찬가지로 `inf`입니다.
+
+여기에서는 `T5DenseGatedGeluDense.forward`가 출력 활성화를 생성하는데, 절대적으로 가장 큰 값이 약 62.7K인 것을 볼 수 있습니다. 이 값은 fp16의 최대 제한인 64K에 매우 근접합니다. 다음 프레임에서는 일부 요소를 0으로 만든 후 가중치를 재정규화하는 `Dropout`이 있습니다. 이로 인해 절대 최대값이 64K를 초과하고 오버플로우(`inf`)가 발생합니다.
+
+보시다시피, fp16 숫자의 경우 숫자가 매우 커질 때 이전 프레임을 살펴보아야 합니다.
+
+보고서를 `models/t5/modeling_t5.py`의 코드와 일치시켜 보겠습니다.
+
+```python
+class T5DenseGatedGeluDense(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)
+        self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)
+        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
+        self.dropout = nn.Dropout(config.dropout_rate)
+        self.gelu_act = ACT2FN["gelu_new"]
+
+    def forward(self, hidden_states):
+        hidden_gelu = self.gelu_act(self.wi_0(hidden_states))
+        hidden_linear = self.wi_1(hidden_states)
+        hidden_states = hidden_gelu * hidden_linear
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.wo(hidden_states)
+        return hidden_states
+```
+
+이제 `dropout` 호출과 이전의 모든 호출을 쉽게 확인할 수 있습니다.
+
+감지는 `forward` 후크에서 발생하므로, 이러한 보고서는 각 `forward`가 반환된 직후에 즉시 출력됩니다.
+
+전체 보고서로 돌아가서 문제에 대한 조치 및 수정을 하려면, 숫자가 증가하기 시작한 몇 개의 프레임 위로 이동해서 여기서 `fp32` 모드로 전환해야 합니다. 이렇게 해야 숫자가 곱해지거나 합쳐질 때 오버플로우되지 않을 가능성이 높습니다. 물론 다른 해결책도 있을 수 있습니다. 예를 들어, `amp`가 활성화된 경우 일시적으로 끄고 원래의 `forward`를 도우미 래퍼로 이동한 후 다음과 같이 할 수 있습니다:
+
+```python
+def _forward(self, hidden_states):
+    hidden_gelu = self.gelu_act(self.wi_0(hidden_states))
+    hidden_linear = self.wi_1(hidden_states)
+    hidden_states = hidden_gelu * hidden_linear
+    hidden_states = self.dropout(hidden_states)
+    hidden_states = self.wo(hidden_states)
+    return hidden_states
+
+
+import torch
+
+
+def forward(self, hidden_states):
+    if torch.is_autocast_enabled():
+        with torch.cuda.amp.autocast(enabled=False):
+            return self._forward(hidden_states)
+    else:
+        return self._forward(hidden_states)
+```
+
+자동 감지기는 전체 프레임의 입력과 출력에 대해서만 보고하므로, 어디를 살펴봐야 하는지 알면 특정 `forward` 함수의 중간 단계도 분석할 수 있습니다. 이 경우에는 `detect_overflow` 도우미 함수를 사용하여 원하는 위치에 감지기를 삽입할 수 있습니다. 예를 들어:
+
+```python
+from debug_utils import detect_overflow
+
+
+class T5LayerFF(nn.Module):
+    [...]
+
+    def forward(self, hidden_states):
+        forwarded_states = self.layer_norm(hidden_states)
+        detect_overflow(forwarded_states, "after layer_norm")
+        forwarded_states = self.DenseReluDense(forwarded_states)
+        detect_overflow(forwarded_states, "after DenseReluDense")
+        return hidden_states + self.dropout(forwarded_states)
+```
+
+여기서는 이를 추가하여 2개의 것을 추적하고 이제 `forwarded_states`의 `inf` 또는 `nan`이 중간에 감지되었는지를 추적합니다.
+
+실제로 위의 예제에서 각 호출이 `nn.Module`이기 때문에 탐지기가 이미 이를 보고합니다. 로컬에서 직접 계산하는 경우 이렇게 수행한다고 가정해 봅시다.
+
+또한, 자체 코드에서 디버거를 인스턴스화하는 경우 기본값에서 출력되는 프레임 수를 조정할 수 있습니다. 예를 들어:
+
+```python
+from transformers.debug_utils import DebugUnderflowOverflow
+
+debug_overflow = DebugUnderflowOverflow(model, max_frames_to_save=100)
+```
+
+### 특정 배치의 절댓값 최소 및 최대 값 추적 [[specific-batch-absolute-min-and-max-value-tracing]]
+
+동일한 디버깅 클래스는 언더플로우/오버플로우 감지 기능이 꺼진 상태에서 배치별 추적에도 사용할 수 있습니다.
+
+예를 들어, 특정 배치의 각 `forward` 호출의 모든 구성 성분에 대한 절대 최솟값과 최댓값을 확인하고, 이를 배치 1과 3에 대해서만 수행하려면 다음과 같이 이 클래스를 인스턴스화합니다:
+
+```python
+debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1, 3])
+```
+
+그러면 이제 배치 1과 3 전체가 언더플로우/오버플로우 감지기와 동일한 형식으로 추적됩니다.
+
+배치는 0부터 시작합니다.
+
+이는 프로그램이 특정 배치 번호 이후에 오작동하기 시작하는 것을 알고 있는 경우에 유용합니다. 그렇기 때문에 해당 영역으로 바로 이동할 수 있습니다. 이런 구성에 대한 샘플 축소된 출력은 다음과 같습니다.
+
+```
+                  *** Starting batch number=1 ***
+abs min  abs max  metadata
+                  shared Embedding
+1.01e-06 7.92e+02 weight
+0.00e+00 2.47e+04 input[0]
+5.36e-05 7.92e+02 output
+[...]
+                  decoder.dropout Dropout
+1.60e-07 2.27e+01 input[0]
+0.00e+00 2.52e+01 output
+                  decoder T5Stack
+     not a tensor output
+                  lm_head Linear
+1.01e-06 7.92e+02 weight
+0.00e+00 1.11e+00 input[0]
+6.06e-02 8.39e+01 output
+                   T5ForConditionalGeneration
+     not a tensor output
+
+                  *** Starting batch number=3 ***
+abs min  abs max  metadata
+                  shared Embedding
+1.01e-06 7.92e+02 weight
+0.00e+00 2.78e+04 input[0]
+5.36e-05 7.92e+02 output
+[...]
+```
+
+여기에서는 모델의 forward 호출 수와 동일한 수의 프레임이 덤프되므로 많은 수의 프레임이 생성됩니다. 따라서 원하는 것일 수도 있고 아닐 수도 있습니다. 그러나 때로는 일반 디버거보다 디버깅 목적으로 더 쉽게 사용할 수 있습니다. 예를 들어, 문제가 배치 번호 150에서 시작하는 경우 149와 150의 추적을 덤프하고 숫자가 어디서부터 다르게 되었는지 비교할 수 있습니다.
+
+또한, 훈련을 중지할 배치 번호를 지정할 수도 있습니다. 다음과 같이 지정할 수 있습니다.
+
+```python
+debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1, 3], abort_after_batch_num=3)
+```

docs/source/ko/llm_tutorial.md

@@ -0,0 +1,221 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+
+# 대규모 언어 모델로 생성하기 [[generation-with-llms]]
+
+[[open-in-colab]]
+
+LLM 또는 대규모 언어 모델은 텍스트 생성의 핵심 구성 요소입니다. 간단히 말하면, 주어진 입력 텍스트에 대한 다음 단어(정확하게는 토큰)를 예측하기 위해 훈련된 대규모 사전 훈련 변환기 모델로 구성됩니다. 토큰을 한 번에 하나씩 예측하기 때문에 새로운 문장을 생성하려면 모델을 호출하는 것 외에 더 복잡한 작업을 수행해야 합니다. 즉, 자기회귀 생성을 수행해야 합니다.
+
+자기회귀 생성은 몇 개의 초기 입력값을 제공한 후, 그 출력을 다시 모델에 입력으로 사용하여 반복적으로 호출하는 추론 과정입니다. 🤗 Transformers에서는 [`~generation.GenerationMixin.generate`] 메소드가 이 역할을 하며, 이는 생성 기능을 가진 모든 모델에서 사용 가능합니다.
+
+이 튜토리얼에서는 다음 내용을 다루게 됩니다:
+
+* LLM으로 텍스트 생성
+* 일반적으로 발생하는 문제 해결
+* LLM을 최대한 활용하기 위한 다음 단계
+
+시작하기 전에 필요한 모든 라이브러리가 설치되어 있는지 확인하세요:
+
+```bash
+pip install transformers bitsandbytes>=0.39.0 -q
+```
+
+
+## 텍스트 생성 [[generate-text]]
+
+[인과적 언어 모델링(causal language modeling)](tasks/language_modeling)을 목적으로 학습된 언어 모델은 일련의 텍스트 토큰을 입력으로 사용하고, 그 결과로 다음 토큰이 나올 확률 분포를 제공합니다.
+
+<!-- [GIF 1 -- FWD PASS] -->
+<figure class="image table text-center m-0 w-full">
+    <video
+        style="max-width: 90%; margin: auto;"
+        autoplay loop muted playsinline
+        src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/assisted-generation/gif_1_1080p.mov"
+    ></video>
+    <figcaption>"LLM의 전방 패스"</figcaption>
+</figure>
+
+LLM과 자기회귀 생성을 함께 사용할 때 핵심적인 부분은 이 확률 분포로부터 다음 토큰을 어떻게 고를 것인지입니다. 다음 반복 과정에 사용될 토큰을 결정하는 한, 어떠한 방법도 가능합니다. 확률 분포에서 가장 가능성이 높은 토큰을 선택하는 것처럼 간단할 수도 있고, 결과 분포에서 샘플링하기 전에 수십 가지 변환을 적용하는 것처럼 복잡할 수도 있습니다.
+
+<!-- [GIF 2 -- TEXT GENERATION] -->
+<figure class="image table text-center m-0 w-full">
+    <video
+        style="max-width: 90%; margin: auto;"
+        autoplay loop muted playsinline
+        src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/assisted-generation/gif_2_1080p.mov"
+    ></video>
+    <figcaption>"자기회귀 생성은 확률 분포에서 다음 토큰을 반복적으로 선택하여 텍스트를 생성합니다."</figcaption>
+</figure>
+
+위에서 설명한 과정은 어떤 종료 조건이 충족될 때까지 반복적으로 수행됩니다. 모델이 시퀀스의 끝(EOS 토큰)을 출력할 때까지를 종료 조건으로 하는 것이 이상적입니다. 그렇지 않은 경우에는 미리 정의된 최대 길이에 도달했을 때 생성이 중단됩니다.
+
+모델이 예상대로 동작하기 위해선 토큰 선택 단계와 정지 조건을 올바르게 설정하는 것이 중요합니다. 이러한 이유로, 각 모델에는 기본 생성 설정이 잘 정의된 [`~generation.GenerationConfig`] 파일이 함께 제공됩니다.
+
+코드를 확인해봅시다!
+
+<Tip>
+
+기본 LLM 사용에 관심이 있다면, 우리의 [`Pipeline`](pipeline_tutorial) 인터페이스로 시작하는 것을 추천합니다. 그러나 LLM은 양자화나 토큰 선택 단계에서의 미세한 제어와 같은 고급 기능들을 종종 필요로 합니다. 이러한 작업은 [`~generation.GenerationMixin.generate`]를 통해 가장 잘 수행될 수 있습니다. LLM을 이용한 자기회귀 생성은 자원을 많이 소모하므로, 적절한 처리량을 위해 GPU에서 실행되어야 합니다.
+
+</Tip>
+
+<!-- TODO: update example to llama 2 (or a newer popular baseline) when it becomes ungated -->
+먼저, 모델을 불러오세요.
+
+```py
+>>> from transformers import AutoModelForCausalLM
+
+>>> model = AutoModelForCausalLM.from_pretrained(
+...     "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
+... )
+```
+
+`from_pretrained` 함수를 호출할 때 2개의 플래그를 주목하세요:
+
+- `device_map`은 모델이 GPU로 이동되도록 합니다.
+- `load_in_4bit`는 리소스 요구 사항을 크게 줄이기 위해 [4비트 동적 양자화](main_classes/quantization)를 적용합니다.
+
+이 외에도 모델을 초기화하는 다양한 방법이 있지만, LLM을 처음 시작할 때 이 설정을 추천합니다.
+
+이어서 텍스트 입력을 [토크나이저](tokenizer_summary)으로 전처리하세요.
+
+```py
+>>> from transformers import AutoTokenizer
+
+>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b")
+>>> model_inputs = tokenizer(["A list of colors: red, blue"], return_tensors="pt").to("cuda")
+```
+
+`model_inputs` 변수에는 토큰화된 텍스트 입력과 함께 어텐션 마스크가 들어 있습니다. [`~generation.GenerationMixin.generate`]는 어텐션 마스크가 제공되지 않았을 경우에도 이를 추론하려고 노력하지만, 최상의 성능을 위해서는 가능하면 어텐션 마스크를 전달하는 것을 권장합니다. 
+
+마지막으로 [`~generation.GenerationMixin.generate`] 메소드를 호출해 생성된 토큰을 얻은 후, 이를 출력하기 전에 텍스트 형태로 변환하세요.
+
+```py
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'A list of colors: red, blue, green, yellow, black, white, and brown'
+```
+
+이게 전부입니다! 몇 줄의 코드만으로 LLM의 능력을 활용할 수 있게 되었습니다.
+
+
+## 일반적으로 발생하는 문제 [[common-pitfalls]]
+
+[생성 전략](generation_strategies)이 많고, 기본값이 항상 사용 사례에 적합하지 않을 수 있습니다. 출력이 예상과 다를 때 흔히 발생하는 문제와 이를 해결하는 방법에 대한 목록을 만들었습니다.
+
+```py
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
+
+>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b")
+>>> tokenizer.pad_token = tokenizer.eos_token  # Llama has no pad token by default
+>>> model = AutoModelForCausalLM.from_pretrained(
+...     "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
+... )
+```
+
+### 생성된 출력이 너무 짧거나 길다 [[generated-output-is-too-shortlong]]
+
+[`~generation.GenerationConfig`] 파일에서 별도로 지정하지 않으면, `generate`는 기본적으로 최대 20개의 토큰을 반환합니다. `generate` 호출에서 `max_new_tokens`을 수동으로 설정하여 반환할 수 있는 새 토큰의 최대 수를 설정하는 것이 좋습니다. LLM(정확하게는 [디코더 전용 모델](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt))은 입력 프롬프트도 출력의 일부로 반환합니다.
+
+
+```py
+>>> model_inputs = tokenizer(["A sequence of numbers: 1, 2"], return_tensors="pt").to("cuda")
+
+>>> # By default, the output will contain up to 20 tokens
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'A sequence of numbers: 1, 2, 3, 4, 5'
+
+>>> # Setting `max_new_tokens` allows you to control the maximum length
+>>> generated_ids = model.generate(**model_inputs, max_new_tokens=50)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'A sequence of numbers: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,'
+```
+
+### 잘못된 생성 모드 [[incorrect-generation-mode]]
+
+기본적으로 [`~generation.GenerationConfig`] 파일에서 별도로 지정하지 않으면, `generate`는 각 반복에서 가장 확률이 높은 토큰을 선택합니다(그리디 디코딩). 하려는 작업에 따라 이 방법은 바람직하지 않을 수 있습니다. 예를 들어, 챗봇이나 에세이 작성과 같은 창의적인 작업은 샘플링이 적합할 수 있습니다. 반면, 오디오를 텍스트로 변환하거나 번역과 같은 입력 기반 작업은 그리디 디코딩이 더 적합할 수 있습니다. `do_sample=True`로 샘플링을 활성화할 수 있으며, 이 주제에 대한 자세한 내용은 이 [블로그 포스트](https://huggingface.co/blog/how-to-generate)에서 볼 수 있습니다.
+
+```py
+>>> # Set seed or reproducibility -- you don't need this unless you want full reproducibility
+>>> from transformers import set_seed
+>>> set_seed(0)
+
+>>> model_inputs = tokenizer(["I am a cat."], return_tensors="pt").to("cuda")
+
+>>> # LLM + greedy decoding = repetitive, boring output
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'I am a cat. I am a cat. I am a cat. I am a cat'
+
+>>> # With sampling, the output becomes more creative!
+>>> generated_ids = model.generate(**model_inputs, do_sample=True)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'I am a cat.\nI just need to be. I am always.\nEvery time'
+```
+
+### 잘못된 패딩 [[wrong-padding-side]]
+
+LLM은 [디코더 전용](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt) 구조를 가지고 있어, 입력 프롬프트에 대해 지속적으로 반복 처리를 합니다. 입력 데이터의 길이가 다르면 패딩 작업이 필요합니다. LLM은 패딩 토큰에서 작동을 이어가도록 설계되지 않았기 때문에, 입력 왼쪽에 패딩이 추가 되어야 합니다. 그리고 어텐션 마스크도 꼭 `generate` 함수에 전달되어야 합니다!
+
+```py
+>>> # The tokenizer initialized above has right-padding active by default: the 1st sequence,
+>>> # which is shorter, has padding on the right side. Generation fails.
+>>> model_inputs = tokenizer(
+...     ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
+... ).to("cuda")
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids[0], skip_special_tokens=True)[0]
+''
+
+>>> # With left-padding, it works as expected!
+>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b", padding_side="left")
+>>> tokenizer.pad_token = tokenizer.eos_token  # Llama has no pad token by default
+>>> model_inputs = tokenizer(
+...     ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
+... ).to("cuda")
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'1, 2, 3, 4, 5, 6,'
+```
+
+<!-- TODO: when the prompting guide is ready, mention the importance of setting the right prompt in this section -->
+
+## 추가 자료 [[further-resources]]
+
+자기회귀 생성 프로세스는 상대적으로 단순한 편이지만, LLM을 최대한 활용하려면 여러 가지 요소를 고려해야 하므로 쉽지 않을 수 있습니다. LLM에 대한 더 깊은 이해와 활용을 위한 다음 단계는 아래와 같습니다:
+
+<!-- TODO: complete with new guides -->
+### 고급 생성 사용 [[advanced-generate-usage]]
+
+1. [가이드](generation_strategies)는 다양한 생성 방법을 제어하는 방법, 생성 설정 파일을 설정하는 방법, 출력을 스트리밍하는 방법에 대해 설명합니다.
+2. [`~generation.GenerationConfig`]와 [`~generation.GenerationMixin.generate`], [generate-related classes](internal/generation_utils)를 참조해보세요.
+
+### LLM 리더보드 [[llm-leaderboards]]
+
+1. [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)는 오픈 소스 모델의 품질에 중점을 둡니다.
+2. [Open LLM-Perf Leaderboard](https://huggingface.co/spaces/optimum/llm-perf-leaderboard)는 LLM 처리량에 중점을 둡니다.
+
+### 지연 시간 및 처리량 [[latency-and-throughput]]
+
+1. 메모리 요구 사항을 줄이려면, 동적 양자화에 대한 [가이드](main_classes/quantization)를 참조하세요.
+
+### 관련 라이브러리 [[related-libraries]]
+
+1. [`text-generation-inference`](https://github.com/huggingface/text-generation-inference)는 LLM을 위한 실제 운영 환경에 적합한 서버입니다.
+2. [`optimum`](https://github.com/huggingface/optimum)은 특정 하드웨어 장치에서 LLM을 최적화하기 위해 🤗 Transformers를 확장한 것입니다.

docs/source/ko/model_doc/llama.md

@@ -0,0 +1,117 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# LLaMA [[llama]]
+
+## 개요 [[overview]]
+
+LLaMA 모델은 Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, Guillaume Lample에 의해 제안된 [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971)에서 소개되었습니다. 이 모델은 7B에서 65B개의 파라미터까지 다양한 크기의 기초 언어 모델을 모아놓은 것입니다.
+
+논문의 초록은 다음과 같습니다:
+
+*"LLaMA는 7B에서 65B개의 파라미터 수를 가진 기초 언어 모델의 모음입니다. 우리는 수조 개의 토큰으로 모델을 훈련시켰고, 공개적으로 이용 가능한 데이터셋만을 사용하여 최고 수준의 모델을 훈련시킬 수 있음을 보여줍니다. 특히, LLaMA-13B 모델은 대부분의 벤치마크에서 GPT-3 (175B)를 능가하며, LLaMA-65B는 최고 수준의 모델인 Chinchilla-70B와 PaLM-540B에 버금가는 성능을 보입니다. 우리는 모든 모델을 연구 커뮤니티에 공개합니다."*
+
+팁:
+
+- LLaMA 모델의 가중치는 [이 양식](https://docs.google.com/forms/d/e/1FAIpQLSfqNECQnMkycAp2jP4Z9TFX0cGR4uf7b_fBxjY_OjhJILlKGA/viewform?usp=send_form)을 작성하여 얻을 수 있습니다.
+- 가중치를 다운로드한 후에는 이를 [변환 스크립트](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py)를 사용하여 Hugging Face Transformers 형식으로 변환해야합니다. 변환 스크립트를 실행하려면 아래의 예시 명령어를 참고하세요:
+
+```bash
+python src/transformers/models/llama/convert_llama_weights_to_hf.py \
+    --input_dir /path/to/downloaded/llama/weights --model_size 7B --output_dir /output/path
+```
+
+- 변환을 하였다면 모델과 토크나이저는 다음과 같이 로드할 수 있습니다:
+
+```python
+from transformers import LlamaForCausalLM, LlamaTokenizer
+
+tokenizer = LlamaTokenizer.from_pretrained("/output/path")
+model = LlamaForCausalLM.from_pretrained("/output/path")
+```
+
+스크립트를 실행하기 위해서는 모델을 float16 정밀도로 전부 로드할 수 있을 만큼의 충분한 CPU RAM이 필요합니다. (가장 큰 버전의 모델이 여러 체크포인트로 나뉘어 있더라도, 각 체크포인트는 모델의 각 가중치의 일부를 포함하고 있기 때문에 모든 체크포인트를 RAM에 로드해야 합니다) 65B 모델의 경우, 총 130GB의 RAM이 필요합니다.
+
+
+- LLaMA 토크나이저는 [sentencepiece](https://github.com/google/sentencepiece)를 기반으로 하는 BPE 모델입니다. sentencepiece의 특징 중 하나는 시퀀스를 디코딩할 때 첫 토큰이 단어의 시작이라면 (예를 들어 "Banana"), 토크나이저는 문자열 앞에 공백을 추가하지 않는다는 것입니다.
+
+이 모델은 [BlackSamorez](https://huggingface.co/BlackSamorez)의 기여와 함께, [zphang](https://huggingface.co/zphang)에 의해 제공되었습니다. Hugging Face에서의 구현 코드는 GPT-NeoX를 기반으로 하며 [여기](https://github.com/EleutherAI/gpt-neox)에서 찾을 수 있고, 저자의 코드 원본은 [여기](https://github.com/facebookresearch/llama)에서 확인할 수 있습니다.
+
+
+원래 LLaMA 모델을 기반으로 Meta AI에서 몇 가지 후속 작업을 발표했습니다:
+
+- **Llama2**: Llama2는 구조적인 몇 가지 수정(Grouped Query Attention)을 통해 개선된 버전이며, 2조 개의 토큰으로 사전 훈련이 되어 있습니다. Llama2에 대한 자세한 내용은 [이 문서](llama2)를 참고하세요.
+
+## 리소스 [[resources]]
+
+LLaMA를 시작하는 데 도움이 될 Hugging Face 및 커뮤니티(🌎로 표시)의 공식 자료 목록입니다. 여기에 자료를 제출하고 싶다면 Pull Request를 올려주세요! 추가할 자료는 기존의 자료와 중복되지 않고 새로운 내용을 보여주는 것이 좋습니다.
+
+<PipelineTag pipeline="text-classification"/>
+
+- LLaMA 모델을 텍스트 분류 작업에 적용하기 위한 프롬프트 튜닝 방법에 대한 [노트북](https://colab.research.google.com/github/bigscience-workshop/petals/blob/main/examples/prompt-tuning-sst2.ipynb#scrollTo=f04ba4d2) 🌎
+
+<PipelineTag pipeline="question-answering"/>
+
+- [Stack Exchange](https://stackexchange.com/)에서 질문에 답하는 LLaMA를 훈련하는 방법을 위한 [StackLLaMA: RLHF로 LLaMA를 훈련하는 실전 가이드](https://huggingface.co/blog/stackllama#stackllama-a-hands-on-guide-to-train-llama-with-rlhf) 🌎
+
+⚗️ 최적화
+- 제한된 메모리를 가진 GPU에서 xturing 라이브러리를 사용하여 LLaMA 모델을 미세 조정하는 방법에 대한 [노트북](https://colab.research.google.com/drive/1SQUXq1AMZPSLD4mk3A3swUIc6Y2dclme?usp=sharing) 🌎
+
+⚡️ 추론
+- 🤗 PEFT 라이브러리의 PeftModel을 사용하여 LLaMA 모델을 실행하는 방법에 대한 [노트북](https://colab.research.google.com/github/DominguesM/alpaca-lora-ptbr-7b/blob/main/notebooks/02%20-%20Evaluate.ipynb) 🌎
+- LangChain을 사용하여 PEFT 어댑터 LLaMA 모델을 로드하는 방법에 대한 [노트북](https://colab.research.google.com/drive/1l2GiSSPbajVyp2Nk3CFT4t3uH6-5TiBe?usp=sharing) 🌎
+
+🚀 배포
+- 🤗 PEFT 라이브러리와 사용자 친화적인 UI로 LLaMA 모델을 미세 조정하는 방법에 대한 [노트북](https://colab.research.google.com/github/lxe/simple-llama-finetuner/blob/master/Simple_LLaMA_FineTuner.ipynb#scrollTo=3PM_DilAZD8T) 🌎
+- Amazon SageMaker에서 텍스트 생성을 위해 Open-LLaMA 모델을 배포하는 방법에 대한 [노트북](https://github.com/aws/amazon-sagemaker-examples/blob/main/introduction_to_amazon_algorithms/jumpstart-foundation-models/text-generation-open-llama.ipynb) 🌎
+
+## LlamaConfig [[llamaconfig]]
+
+[[autodoc]] LlamaConfig
+
+
+## LlamaTokenizer [[llamatokenizer]]
+
+[[autodoc]] LlamaTokenizer
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - save_vocabulary
+
+## LlamaTokenizerFast [[llamatokenizerfast]]
+
+[[autodoc]] LlamaTokenizerFast
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - update_post_processor
+    - save_vocabulary
+
+## LlamaModel [[llamamodel]]
+
+[[autodoc]] LlamaModel
+    - forward
+
+
+## LlamaForCausalLM [[llamaforcausallm]]
+
+[[autodoc]] LlamaForCausalLM
+    - forward
+
+## LlamaForSequenceClassification [[llamaforsequenceclassification]]
+
+[[autodoc]] LlamaForSequenceClassification
+    - forward

docs/source/ko/model_doc/llama2.md

@@ -0,0 +1,129 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contains specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Llama2 [[llama2]]
+
+## 개요 [[overview]]
+
+Llama2 모델은 Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Ya1smine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing EllenTan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, Thomas Scialom의 논문 [LLaMA: Open Foundation and Fine-Tuned Chat Models](https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/)에서 제안되었습니다. 채팅 어플리케이션에 맞게 미세 조정된 체크포인트를 포함된 7B에서 70B 범위의 매개변수를 가진 기초 언어 모델 모음입니다!
+
+논문의 초록은 다음과 같습니다:
+
+*이 연구에서 우리는 70억에서 700억 파라미터의 범위에서 사전 훈련 및 미세 조정된 대규모 언어 모델(LLMs)의 모음인 Llama 2를 개발 및 공개합니다. Llama 2-Chat라고 불리는 미세 조정된 LLMs은 대화 사용 사례에 최적화되었습니다. 우리의 모델은 테스트한 대부분의 벤치마크에서 오픈 소스 채팅 모델보다 성능이 뛰어나며, 유용성과 안전성에 대한 인적 평가를 바탕으로 비공개 소스 모델을 대체할 수 있는 적절한 대안이 될 수 있습니다. 우리는 Llama 2-Chat의 미세 조정 및 안전성 향상의 접근 방식에 대한 자세한 설명을 제공하여 커뮤니티가 우리의 작업을 기반으로 LLMs의 책임있는 개발에 기여할 수 있도록 합니다.*
+
+[여기](https://huggingface.co/models?search=llama2)에서 모든 Llama2 모델을 확인할 수 있습니다.
+
+<Tip warning={true}>
+
+`Llama2` 모델은 `bfloat16`을 사용하여 훈련되었지만, 원래 추론은 `float16`을 사용합니다. 허브에 업로드된 체크포인트는 `torch_dtype = 'float16'`을 사용하며, 이는 `AutoModel` API에 의해 체크포인트를 `torch.float32`에서 `torch.float16`으로 캐스팅하는 데 사용됩니다. 
+
+온라인 가중치의 `dtype`은 `model = AutoModelForCausalLM.from_pretrained("path", torch_dtype = "auto")`를 사용하여 모델을 초기화할 때 `torch_dtype="auto"`를 사용하지 않는 한 대부분 관련이 없습니다. 그 이유는 모델이 먼저 다운로드될 것이고 (온라인 체크포인트의 `dtype`을 사용하여) 그다음에 기본 `dtype`인 `torch`로 캐스팅하고(`torch.float32`가 됨), 마지막으로 구성(configuration)에서 제공된 `torch_dtype`이 있는 경우 이를 사용하기 때문입니다.
+
+모델을 `float16`에서 훈련하는 것은 권장되지 않으며 `nan`을 생성하는 것으로 알려져 있습니다. 따라서 모델은 `bfloat16`에서 훈련되어야 합니다.
+
+</Tip>
+
+🍯 팁:
+
+- Llama2 모델의 가중치는 [이 양식](https://ai.meta.com/resources/models-and-libraries/llama-downloads/)을 작성하여 얻을 수 있습니다.
+- 아키텍처는 처음 버전의 Llama와 매우 유사하며, [이 논문](https://arxiv.org/pdf/2305.13245.pdf)의 내용에 따라 Grouped Query Attention (GQA)이 추가되었습니다.
+- `config.pretraining_tp`를 1과 다른 값으로 설정하면 더 정확하지만 느린 선형 레이어 계산이 활성화되어 원본 로짓과 더 잘 일치하게 됩니다.
+- 원래 모델은 `pad_id = -1`을 사용하는데, 이는 패딩 토큰이 없음을 의미합니다. 동일한 로직을 사용할 수 없으므로 `tokenizer.add_special_tokens({"pad_token":"<pad>"})`를 사용하여 패딩 토큰을 추가하고 이에 따라 토큰 임베딩 크기를 조정해야 합니다. 또한 `model.config.pad_token_id`를 설정해야 합니다. 모델의 `embed_tokens` 레이어는 `self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.config.padding_idx)`로 초기화되어, 패딩 토큰 인코딩이 0을 출력하도록 합니다. 따라서 초기화 시에 전달하는 것을 권장합니다.
+- 양식을 작성하고 모델 체크포인트 접근 권한을 얻은 후에는 이미 변환된 체크포인트를 사용할 수 있습니다. 그렇지 않고 자신의 모델을 직접 변환하려는 경우, [변환 스크립트](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py)를 자유롭게 사용하세요. 스크립트는 다음과 같은 예시의 명령어로 호출할 수 있습니다:
+
+```bash
+python src/transformers/models/llama/convert_llama_weights_to_hf.py \
+    --input_dir /path/to/downloaded/llama/weights --model_size 7B --output_dir /output/path
+```
+
+- 변환 후 모델과 토크나이저는 다음과 같이 로드할 수 있습니다:
+
+```python
+from transformers import LlamaForCausalLM, LlamaTokenizer
+
+tokenizer = LlamaTokenizer.from_pretrained("/output/path")
+model = LlamaForCausalLM.from_pretrained("/output/path")
+```
+
+스크립트를 실행하려면 모델을 float16 정밀도로 전부 호스트할 수 있을 만큼 충분한 CPU RAM이 필요합니다 (가장 큰 버전이 여러 체크포인트로 제공되더라도 각 체크포인트는 모델 가중치의 일부만을 포함하므로 모두 RAM에 로드해야 합니다). 75B 모델의 경우, 총 145GB의 RAM이 필요합니다.
+
+- LLaMA 토크나이저는 [sentencepiece](https://github.com/google/sentencepiece)를 기반으로 한 BPE 모델입니다. sentencepiece의 특징 중 하나는 시퀀스를 디코딩할 때 첫 번째 토큰이 단어의 시작이면 (예: "Banana") 토크나이저는 문자열 앞에 접두사 공간을 추가하지 않는 것입니다.
+
+이 모델은 [Arthur Zucker](https://huggingface.co/ArthurZ)가 [Lysandre Debut](https://huggingface.co/lysandre)의 도움을 받아 제공하였습니다. Hugging Face에서의 구현 코드는 [여기](https://github.com/EleutherAI/gpt-neox)의 GPT-NeoX 를 기반으로 합니다. 저자의 원래 코드는 [여기](https://github.com/facebookresearch/llama)에서 찾을 수 있습니다.
+
+## 리소스 [[resources]]
+
+LLaMA2를 시작하는 데 도움이 될 Hugging Face의 공식 및 커뮤니티(🌎로 표시) 리소스 목록입니다. 여기에 새로운 리소스를 추가하기 위해서 Pull Request를 열어 주시면 검토하겠습니다! 리소스는 기존 리소스와 중복되지 않는 새로운 것을 보여주는 것이 이상적입니다.
+
+- [Llama 2 is here - get it on Hugging Face](https://huggingface.co/blog/llama2), Llama 2에 관한 블로그 포스트와 🤗 Transformers 및 🤗 PEFT와 함께 사용하는 방법에 대한 내용입니다.
+- [LLaMA 2 - Every Resource you need](https://www.philschmid.de/llama-2), LLaMA 2에 대해 알아보고 빠르게 시작하는 데 필요한 관련 리소스의 모음입니다.
+
+<PipelineTag pipeline="text-generation"/>
+
+- Google Colab에서 QLoRA와 4-bit 정밀도를 사용하여 Llama 2를 미세 조정하는 방법에 대한 [노트북](https://colab.research.google.com/drive/1PEQyJO1-f6j0S_XJ8DV50NkpzasXkrzd?usp=sharing)입니다. 🌎
+- "Llama-v2-7b-guanaco" 모델을 4-bit QLoRA로 미세 조정하고 PDF에서 Q&A 데이터셋을 생성하는 방법에 대한 [노트북](https://colab.research.google.com/drive/134o_cXcMe_lsvl15ZE_4Y75Kstepsntu?usp=sharing)입니다. 🌎
+
+⚗️ 최적화
+- [Llama 2를 DPO로 미세 조정하기](https://huggingface.co/blog/dpo-trl), TRL 라이브러리의 DPO 방법을 사용하여 특정 데이터셋에서 Llama 2를 미세 조정하는 방법을 안내하는 가이드입니다.
+- [확장 가이드: Llama 2 명령어 조정](https://www.philschmid.de/instruction-tune-llama-2), 입력에서 명령어를 생성하도록 Llama 2를 훈련시키는 방법을 안내하는 가이드로, 명령어를 따르는 모델에서 명령어를 주는 모델로 변환합니다.
+- 개인 컴퓨터에서 QLoRA와 TRL을 사용하여 Llama 2 모델을 미세 조정하는 방법에 대한 [노트북](https://colab.research.google.com/drive/1SYpgFpcmtIUzdE7pxqknrM4ArCASfkFQ?usp=sharing)입니다. 🌎
+
+⚡️ 추론
+- AutoGPTQ 라이브러리의 GPTQ를 사용하여 Llama 2 모델을 양자화하는 방법에 대한 [노트북](https://colab.research.google.com/drive/1TC56ArKerXUpbgRy5vM3woRsbTEVNq7h?usp=sharing)입니다. 🌎
+- 로컬 컴퓨터나 Google Colab에서 4-bit 양자화로 Llama 2 채팅 모델을 실행하는 방법에 대한 [노트북](https://colab.research.google.com/drive/1X1z9Q6domMKl2CnEM0QGHNwidLfR4dW2?usp=sharing)입니다. 🌎
+
+🚀 배포
+- [Amazon SageMaker에서 LLaMA 2 (7-70B) 미세 조정하기](https://www.philschmid.de/sagemaker-llama2-qlora), Amazon SageMaker에서 QLoRA 미세 조정 및 배포에 이르기까지의 완전한 가이드입니다.
+- [Amazon SageMaker에서 Llama 2 7B/13B/70B 배포하기](https://www.philschmid.de/sagemaker-llama-llm), 안전하고 확장 가능한 배포를 위해 Hugging Face의 LLM DLC 컨테이너를 사용하는 방법에 대한 가이드입니다.
+
+
+## LlamaConfig [[llamaconfig]]
+
+[[autodoc]] LlamaConfig
+
+
+## LlamaTokenizer [[llamatokenizer]]
+
+[[autodoc]] LlamaTokenizer
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - save_vocabulary
+
+## LlamaTokenizerFast [[llamatokenizerfast]]
+
+[[autodoc]] LlamaTokenizerFast
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - update_post_processor
+    - save_vocabulary
+
+## LlamaModel [[llamamodel]]
+
+[[autodoc]] LlamaModel
+    - forward
+
+
+## LlamaForCausalLM [[llamaforcausallm]]
+
+[[autodoc]] LlamaForCausalLM
+    - forward
+
+## LlamaForSequenceClassification [[llamaforsequenceclassification]]
+
+[[autodoc]] LlamaForSequenceClassification
+    - forward

docs/source/ko/model_doc/whisper.md

@@ -0,0 +1,120 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Whisper [[whisper]]
+
+## 개요 [[overview]]
+
+Whisper 모델은 Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, Ilya Sutskever에 의해 [Robust Speech Recognition via Large-Scale Weak Supervision](https://cdn.openai.com/papers/whisper.pdf)에서 제안되었습니다.
+
+논문의 초록은 다음과 같습니다:
+
+*우리는 인터넷에서 대량의 오디오를 글로 옮긴 것을 예측하도록 간단히 훈련된 음성 처리 시스템의 성능을 연구합니다. 68만 시간의 다국어 및 다중 작업 지도(multitask supervision)에 확장했을 때, 결과 모델은 표준 벤치마크에 잘 일반화되며, 미세 조정이 필요 없는 제로샷 전송 설정에서 이전의 완전히 지도된(fully-supervised) 결과와 경쟁할 수 있는 경우가 많습니다. 사람과 비교하면, 이 모델은 사람의 정확도와 견고성에 근접합니다. 우리는 강력한 음성 처리를 위한 추가 작업의 기반이 될 모델과 추론 코드를 공개합니다.*
+
+
+
+팁:
+
+- 이 모델은 일반적으로 별도의 미세 조정 없이도 잘 작동합니다.
+- 아키텍처는 고전적인 인코더-디코더 아키텍처를 따르기 때문에, 추론을 위해 [`~generation.GenerationMixin.generate`] 함수를 사용합니다.
+- 현재 추론은 짧은 형식에만 구현되어 있으며, 오디오는 30초 미만의 세그먼트로 미리 분할되어야 합니다. 타임스탬프를 포함한 긴 형식에 대한 추론은 향후 릴리스에서 구현될 예정입니다.
+- [`WhisperProcessor`]를 사용하여 모델에 사용할 오디오를 준비하고, 예측된 ID를 텍스트로 디코딩할 수 있습니다.
+
+이 모델은 [Arthur Zucker](https://huggingface.co/ArthurZ)에 의해 제공되었습니다. 이 모델의 Tensorflow 버전은 [amyeroberts](https://huggingface.co/amyeroberts)에 의해 제공되었습니다.
+원본 코드는 [여기](https://github.com/openai/whisper)에서 찾을 수 있습니다.
+
+
+
+## WhisperConfig [[whisperconfig]]
+
+[[autodoc]] WhisperConfig
+
+## WhisperTokenizer [[whispertokenizer]]
+
+[[autodoc]] WhisperTokenizer
+    - set_prefix_tokens
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - save_vocabulary
+
+## WhisperTokenizerFast [[whispertokenizerfast]]
+
+[[autodoc]] WhisperTokenizerFast
+    - set_prefix_tokens
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - save_vocabulary
+
+## WhisperFeatureExtractor [[whisperfeatureextractor]]
+
+[[autodoc]] WhisperFeatureExtractor
+    - __call__
+
+## WhisperProcessor [[whisperprocessor]]
+
+[[autodoc]] WhisperProcessor
+    - __call__
+    - from_pretrained
+    - save_pretrained
+    - batch_decode
+    - decode
+
+## WhisperModel [[whispermodel]]
+
+[[autodoc]] WhisperModel
+    - forward
+    - _mask_input_features
+
+## WhisperForConditionalGeneration [[whisperforconditionalgeneration]]
+
+[[autodoc]] WhisperForConditionalGeneration
+    - forward
+
+## WhisperForAudioClassification [[whisperforaudioclassification]]
+
+[[autodoc]] WhisperForAudioClassification
+    - forward
+
+
+
+## TFWhisperModel [[tfwhispermodel]]
+
+[[autodoc]] TFWhisperModel
+    - call
+
+## TFWhisperForConditionalGeneration [[tfwhisperforconditionalgeneration]]
+
+[[autodoc]] TFWhisperForConditionalGeneration
+    - call
+
+
+## FlaxWhisperModel [[flaxwhispermodel]]
+
+[[autodoc]] FlaxWhisperModel
+    - __call__
+
+## FlaxWhisperForConditionalGeneration [[flaxwhisperforconditionalgeneration]]
+
+[[autodoc]] FlaxWhisperForConditionalGeneration
+    - __call__
+
+## FlaxWhisperForAudioClassification [[flaxwhisperforaudioclassification]]
+
+[[autodoc]] FlaxWhisperForAudioClassification
+    - __call__
+

docs/source/ko/perf_train_gpu_many.md

@@ -0,0 +1,533 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+
+⚠️ 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.
+
+-->
+
+# 다중 GPU에서 효율적인 훈련 [[efficient-training-on-multiple-gpus]]
+
+단일 GPU에서의 훈련이 너무 느리거나 모델 가중치가 단일 GPU의 메모리에 맞지 않는 경우, 다중-GPU 설정을 사용합니다. 단일 GPU에서 다중 GPU로 전환하기 위해서는 작업을 분산해야 합니다. 데이터, 텐서 또는 파이프라인과 같은 병렬화 기법을 사용하여 작업을 병렬로 처리할 수 있습니다. 그러나 이러한 설정을 모두에게 적용할 수 있는 완벽한 해결책은 없으며, 어떤 설정이 가장 적합한지는 사용하는 하드웨어에 따라 달라집니다. 이 문서는 주로 PyTorch 기반의 구현을 중심으로 설명하며, 대부분의 개념은 다른 프레임워크에도 적용될 수 있을 것으로 예상됩니다.
+
+<Tip>
+
+ 참고: [단일 GPU 섹션](perf_train_gpu_one)에서 소개된 전략(혼합 정밀도 훈련 또는 그래디언트 누적 등)은 일반적으로 모델 훈련에 적용되며, 다중-GPU 또는 CPU 훈련과 같은 다음 섹션으로 진입하기 전에 해당 섹션을 참고하는 것이 좋습니다.
+
+</Tip>
+
+먼저 1D 병렬화 기술에 대해 자세히 논의한 후, 이러한 기술을 결합하여 2D 및 3D 병렬화를 구현하여 더 빠른 훈련과 더 큰 모델을 지원하는 방법을 살펴볼 것입니다. 또한 다른 효과적인 대안 방식도 소개될 예정입니다.
+
+## 개념 [[concepts]]
+
+다음은 이 문서에서 자세히 설명될 주요 개념에 대한 간단한 설명입니다.
+
+1. **DataParallel (DP)** - 동일한 설정이 여러 번 복제되고, 각 설정에 데이터 일부를 받습니다. 처리는 병렬로 수행되며 모든 설정은 각 훈련 단계의 끝날 때 동기화됩니다.
+2. **TensorParallel (TP)** - 각 텐서는 여러 개의 묶음으로 분할되기에, 전체 텐서가 단일 GPU에 상주하는 대신 텐서의 각 샤드가 지정된 GPU에 상주합니다. 처리하는 동안 각 샤드는 서로 다른 GPU에서 개별적으로 병렬 처리되며 결과는 단계가 끝날 때 동기화됩니다. 분할이 수평 수준에서 이루어지기 때문에 이를 수평 병렬 처리라고 부를 수 있습니다.
+3. **PipelineParallel (PP)** - 모델이 수직으로 (레이어 수준) 여러 GPU에 분할되어 모델의 단일 GPU에는 하나 또는 여러 레이어가 배치됩니다. 각 GPU는 파이프라인의 서로 다른 단계를 병렬로 처리하며 작은 배치 묶음에서 작동합니다.
+4. **Zero Redundancy Optimizer (ZeRO)** - TP와 유사하게 텐서를 샤딩하지만, 전체 텐서는 순방향 또는 역방향 계산을 위해 재구성되므로 모델을 수정할 필요가 없습니다. 또한 제한된 GPU 메모리를 보완하기 위해 다양한 오프로드 기술을 지원합니다.
+5. **Sharded DDP** - ZeRO의 기본 개념으로 다른 ZeRO 구현에서도 사용되는 용어입니다.
+
+각 개념의 구체적인 내용에 대해 자세히 들어가기 전에 대규모 인프라에서 대규모 모델을 훈련하는 경우의 대략적인 결정 과정을 살펴보겠습니다.
+
+## 확장성 전략 [[scalability-strategy]]
+
+**⇨ 단일 노드 / 다중-GPU**
+* 모델이 단일 GPU에 맞는 경우:
+
+    1. DDP - 분산 DP
+    2. ZeRO - 상황과 구성에 따라 더 빠를 수도 있고 그렇지 않을 수도 있음
+
+* 모델이 단일 GPU에 맞지 않는 경우:
+
+    1. PP
+    2. ZeRO
+    3. TP
+
+    노드 내 연결 속도가 매우 빠른 NVLINK 또는 NVSwitch의 경우 세 가지 방법은 대부분 비슷한 성능을 보여야 하며, PP가 없는 경우 TP 또는 ZeRO보다 빠를 것입니다. TP의 정도도 차이를 만들 수 있습니다. 특정 설정에서 승자를 찾기 위해 실험하는 것이 가장 좋습니다.
+
+    TP는 거의 항상 단일 노드 내에서 사용됩니다. 즉, TP 크기 <= 노드당 GPU 수입니다.
+
+* 가장 큰 레이어가 단일 GPU에 맞지 않는 경우:
+
+    1. ZeRO를 사용하지 않는 경우 - PP만으로는 맞지 않으므로 TP를 반드시 사용해야 함
+    2. ZeRO를 사용하는 경우에는 위의 "단일 GPU" 항목과 동일
+
+
+**⇨ 다중 노드 /  다중 GPU**
+
+* 노드 간 연결 속도가 빠른 경우:
+
+    1. ZeRO - 모델에 대부분의 수정을 필요로 하지 않음
+    2. PP+TP+DP - 통신이 적지만 모델에 대대적인 변경이 필요함
+
+* 노드 간 연결 속도가 느리며, GPU 메모리가 여전히 부족한 경우:
+
+    1. DP+PP+TP+ZeRO-1
+	
+
+
+## 데이터 병렬화 [[data-parallelism]]
+
+2개의 GPU만으로도 대부분의 사용자들은 `DataParallel` (DP)과 `DistributedDataParallel` (DDP)을 통해 향상된 훈련 속도를 누릴 수 있습니다. 이는 PyTorch의 내장 기능입니다. 일반적으로 DDP를 사용하는 것이 좋으며, DP는 일부 모델에서 작동하지 않을 수 있으므로 주의해야 합니다. [PyTorch 문서](https://pytorch.org/docs/master/generated/torch.nn.DataParallel.html)에서도 DDP의 사용을 권장합니다.
+
+### DP vs DDP [[dp-vs-ddp]]
+
+`DistributedDataParallel` (DDP)은 일반적으로 `DataParallel` (DP)보다 빠르지만, 항상 그렇지는 않습니다:
+* DP는 파이썬 스레드 기반인 반면, DDP는 다중 프로세스 기반이기 때문에 GIL과 같은 파이썬 스레드 제한이 없습니다.
+* 그러나 GPU 카드 간의 느린 상호 연결성은 DDP로 인해 실제로 느린 결과를 낼 수 있습니다.
+
+이 두 모드 간의 GPU 간 통신 오버헤드의 주요 차이점은 다음과 같습니다:
+
+[DDP](https://pytorch.org/docs/master/notes/ddp.html):
+
+- 시작할 때, 주 프로세스가 모델을 gpu 0에서 다른 모든 gpu로 복제합니다.
+- 그런 다음 각 배치에 대해:
+   1. 각 gpu는 자체 미니 배치 데이터를 직접 사용합니다.
+   2. `backward` 동안 로컬 그래디언트가 준비되면, 모든 프로세스에 평균화됩니다.
+
+[DP](https://pytorch.org/docs/master/generated/torch.nn.DataParallel.html):
+
+각 배치에 대해:
+   1. gpu 0은 데이터 배치를 읽고 각 gpu에 미니 배치를 보냅니다.
+   2. 업데이트된 모델을 gpu 0에서 각 gpu로 복제합니다.
+   3. `forward`를 실행하고 각 gpu의 출력을 gpu 0으로 보내고 손실을 계산합니다.
+   4. gpu 0에서 모든 gpu로 손실을 분산하고 `backward`를 실행합니다.
+   5. 각 gpu에서 그래디언트를 gpu 0으로 보내고 이를 평균화합니다.
+
+DDP는 각 배치마다 그래디언트를 보내는 통신만을 수행하며, DP는 배치마다 5개의 다른 데이터 교환을 수행합니다.
+
+DP는 파이썬 스레드를 통해 프로세스 내에서 데이터를 복제하며, DDP는 [torch.distributed](https://pytorch.org/docs/master/distributed.html)를 통해 데이터를 복제합니다.
+
+DP에서는 gpu 0이 다른 gpu보다 훨씬 더 많은 작업을 수행하므로, gpu의 활용도가 낮아집니다.
+
+DDP는 여러 대의 컴퓨터에서 사용할 수 있지만, DP의 경우는 그렇지 않습니다.
+
+DP와 DDP 사이에는 다른 차이점이 있지만, 이 토론과는 관련이 없습니다.
+
+이 2가지 모드를 깊게 이해하고 싶다면, [이 문서](https://www.telesens.co/2019/04/04/distributed-data-parallel-training-using-pytorch-on-aws/)를 강력히 추천합니다. 이 문서는 멋진 다이어그램을 포함하고 있으며, 다양한 하드웨어에서 여러 벤치마크와 프로파일러 출력을 설명하여 필요한 세부 사항을 모두 설명합니다.
+
+실제 벤치마크를 살펴보겠습니다:
+
+| Type   | NVlink | Time |
+| :----- | -----  | ---: |
+| 2:DP   | Y      | 110s |
+| 2:DDP  | Y      | 101s |
+| 2:DDP  | N      | 131s |
+
+
+분석:
+
+여기서 DP는 NVlink가 있는 DDP보다 약 10% 느립니다. 그러나 NVlink가 없는 DDP보다 약 15% 빠릅니다.
+
+실제 차이는 각 GPU가 다른 GPU와 동기화해야 하는 데이터 양에 따라 달라질 것입니다. 동기화할 데이터가 많을수록 느린 링크가 총 실행 시간을 늦출 수 있습니다.
+
+다음은 전체 벤치마크 코드와 출력입니다:
+
+해당 벤치마크에서 `NCCL_P2P_DISABLE=1`을 사용하여 NVLink 기능을 비활성화했습니다.
+
+```
+
+# DP
+rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 \
+python examples/pytorch/language-modeling/run_clm.py \
+--model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
+--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
+
+{'train_runtime': 110.5948, 'train_samples_per_second': 1.808, 'epoch': 0.69}
+
+# DDP w/ NVlink
+rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 \
+python -m torch.distributed.launch --nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py \
+--model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
+--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
+
+{'train_runtime': 101.9003, 'train_samples_per_second': 1.963, 'epoch': 0.69}
+
+# DDP w/o NVlink
+rm -r /tmp/test-clm; NCCL_P2P_DISABLE=1 CUDA_VISIBLE_DEVICES=0,1 \
+python -m torch.distributed.launch --nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py \
+--model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
+--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
+
+{'train_runtime': 131.4367, 'train_samples_per_second': 1.522, 'epoch': 0.69}
+```
+
+하드웨어: 각각 24GB의 TITAN RTX 2개 + NVlink과 2개의 NVLink (`nvidia-smi topo -m`에서 `NV2`입니다.)
+소프트웨어: `pytorch-1.8-to-be` + `cuda-11.0` / `transformers==4.3.0.dev0`
+
+## ZeRO 데이터 병렬화 [[zero-data-parallelism]]
+
+ZeRO를 기반으로 한 데이터 병렬화 (ZeRO-DP)는 다음 [블로그 글](https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/)의 다음 다이어그램에서 설명되고 있습니다.
+![DeepSpeed-Image-1](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-zero.png)
+
+이 개념은 이해하기 어려울 수 있지만, 실제로는 매우 간단한 개념입니다. 이는 일반적인 `DataParallel` (DP)과 동일하지만, 전체 모델 매개변수, 그래디언트 및 옵티마이저 상태를 복제하는 대신 각 GPU는 그 중 일부만 저장합니다. 그리고 실행 시간에는 주어진 레이어에 대해 전체 레이어 매개변수가 필요할 때 각 GPU가 서로에게 필요한 부분을 제공하기 위해 동기화됩니다 - 그게 전부입니다.
+
+각각 3개의 레이어와 3개의 매개변수가 있는 간단한 모델을 생각해 봅시다:
+```
+La | Lb | Lc
+---|----|---
+a0 | b0 | c0
+a1 | b1 | c1
+a2 | b2 | c2
+```
+레이어 La에는 가중치 a0, a1 및 a2가 있습니다.
+
+3개의 GPU가 있는 경우, Sharded DDP (= Zero-DP)는 다음과 같이 모델을 3개의 GPU에 분할합니다:
+
+```
+GPU0:
+La | Lb | Lc
+---|----|---
+a0 | b0 | c0
+
+GPU1:
+La | Lb | Lc
+---|----|---
+a1 | b1 | c1
+
+GPU2:
+La | Lb | Lc
+---|----|---
+a2 | b2 | c2
+```
+
+일반적인 DNN 다이어그램을 상상해보면 이는 텐서 병렬 처리와 같은 수평 슬라이싱입니다. 수직 슬라이싱은 전체 레이어 그룹을 다른 GPU에 배치하는 것입니다. 이는 시작에 불과합니다.
+
+이제 이러한 각각의 GPU는 DP에서 작동하는 것과 마찬가지로 일반적인 미니 배치를 받습니다:
+```
+x0 => GPU0
+x1 => GPU1
+x2 => GPU2
+```
+
+입력은 수정되지 않은 상태로 일반 모델에 의해 처리될 것으로 간주합니다.
+
+먼저, 입력은 레이어 La에 도달합니다.
+
+GPU0에만 집중해 보겠습니다. x0은 순방향 경로를 수행하기 위해 a0, a1, a2 파라미터가 필요하지만 GPU0에는 a0만 있습니다. GPU1에서 a1을, GPU2에서 a2를 전송받아 모델의 모든 조각을 하나로 모읍니다.
+
+병렬적으로, GPU1은 미니 배치 x1을 받고 a1만 가지고 있지만, a0 및 a2 매개변수가 필요합니다. 따라서 GPU0 및 GPU2에서 이를 가져옵니다.
+
+GPU2도 동일한 작업을 수행합니다. 입력 x2를 받고 GPU0 및 GPU1에서 각각 a0과 a1을, 그리고 자신의 a2와 함께 전체 텐서를 복원합니다.
+
+3개의 GPU는 복원된 전체 텐서를 받고 forward가 수행됩니다.
+
+계산이 완료되면 더 이상 필요하지 않은 데이터는 삭제되고, 해당 데이터는 계산 중에만 사용됩니다. 복원은 사전 패치를 통해 효율적으로 수행됩니다.
+
+그리고 전체 프로세스는 레이어 Lb에 대해 반복되고, 그 다음 Lc로 순방향으로, 그다음은 역방향으로 Lc -> Lb -> La로 반복됩니다.
+
+개인적으로 이것은 효율적인 그룹 배낭 여행자의 중량 분배 전략처럼 들립니다:
+
+1. 사람 A가 텐트를 운반합니다.
+2. 사람 B가 난로를 운반합니다.
+3. 사람 C가 도끼를 운반합니다.
+
+이제 매일 밤 각자 가진 것을 다른 사람들과 공유하고, 가지지 않은 것은 다른 사람들로부터 받고, 아침에는 할당된 유형의 장비를 싸고 계속해서 여행을 진행합니다. 이것이 Sharded DDP / Zero DP입니다.
+
+이 전략을 각각 자신의 텐트, 난로 및 도끼를 개별적으로 운반해야 하는 단순한 전략과 비교해보면 훨씬 비효율적일 것입니다. 이것이 Pytorch의 DataParallel (DP 및 DDP)입니다.
+
+이 주제에 대해 논문을 읽을 때 다음 동의어를 만날 수 있습니다: Sharded, Partitioned.
+
+ZeRO가 모델 가중치를 분할하는 방식을 자세히 살펴보면, 텐서 병렬화와 매우 유사한 것을 알 수 있습니다. 이는 이후에 설명될 수직 모델 병렬화와는 달리 각 레이어의 가중치를 분할/분할하기 때문입니다.
+
+구현:
+
+- [DeepSpeed](https://www.deepspeed.ai/features/#the-zero-redundancy-optimizer)는 1단계 + 2단계 + 3단계의 ZeRO-DP를 제공합니다.
+- [Fairscale](https://github.com/facebookresearch/fairscale/#optimizer-state-sharding-zero)은 1단계 + 2단계 + 3단계의 ZeRO-DP를 제공합니다.
+- [`transformers` 통합](main_classes/trainer#trainer-integrations)
+
+## 네이티브 모델 병렬 처리(수직적) 및 파이프라인 병렬 처리[[naive-model-parallelism-vertical-and-pipeline-parallelism]]
+
+Naive Model Parallelism (MP)은 모델 레이어 그룹을 다중 GPU에 분산하는 방식입니다. 메커니즘은 상대적으로 간단합니다. 원하는 레이어를 `.to()`를 사용하여 원하는 장치로 전환하면 데이터가 해당 레이어로 들어오고 나갈 때 데이터도 레이어와 동일한 장치로 전환되고 나머지는 수정되지 않습니다.
+
+대부분의 모델이 그려지는 방식이 레이어를 세로로 슬라이스하기 때문에 이를 수직 모델 병렬화라고 부릅니다. 예를 들어 다음 다이어그램은 8레이어 모델을 보여줍니다:
+
+```
+===================  ===================
+|  0 | 1 | 2 | 3  |  |  4 | 5 | 6 | 7  |
+===================  ===================
+        gpu0                 gpu1
+```
+우리는 모델을 수직으로 2개로 분할하여 레이어 0-3을 GPU0에 배치하고 레이어 4-7을 GPU1에 배치했습니다.
+
+이제 데이터가 레이어 0에서 1로, 1에서 2로, 2에서 3으로 이동하는 동안에는 일반적인 모델입니다. 그러나 데이터가 레이어 3에서 레이어 4로 전달되어야 할 때는 GPU0에서 GPU1로 이동해야 하므로 통신 오버헤드가 발생합니다. 참여하는 GPU가 동일한 컴퓨팅 노드(예: 동일한 물리적인 기계)에 있는 경우 이 복사는 매우 빠릅니다. 그러나 GPU가 서로 다른 컴퓨팅 노드(예: 여러 기계)에 위치한 경우 통신 오버헤드는 상당히 크게 될 수 있습니다.
+
+그런 다음 레이어 4부터 5로, 6으로, 7로 진행되는 것은 일반적인 모델과 동일하게 진행되고, 7번째 레이어가 완료되면 데이터를 다시 레이어 0으로 보내거나 또는 레이블을 마지막 레이어로 보내야 할 필요가 있습니다. 이제 손실을 계산하고 옵티마이저가 작동할 수 있습니다.
+
+문제점:
+- 이 방식을 "naive" MP라고 부르는 이유는 주어진 상황에 하나의 GPU를 제외한 모든 GPU가 유휴 상태라는 점입니다. 따라서 4개의 GPU를 사용하는 경우 단일 GPU의 메모리 양을 4배로 늘리고 나머지 하드웨어는 무시하는 것과 거의 동일합니다. 또한 장치 간 데이터 복사의 오버헤드도 있습니다. 따라서 4개의 6GB 카드는 naive MP를 사용하여 1개의 24GB 카드와 동일한 크기를 수용할 수 있지만, 후자는 데이터 복사의 오버헤드가 없으므로 훈련을 더 빨리 완료합니다. 그러나 예를 들어 40GB 카드가 있고 45GB 모델을 맞추어야 할 경우 4개의 40GB 카드로 맞출 수 있습니다 (하지만 그래디언트와 옵티마이저 상태 때문에 가까스로 가능합니다).
+- 공유 임베딩은 GPU 간에 복사해야 할 수도 있습니다.
+
+파이프라인 병렬화 (PP)은 거의 naive MP와 동일하지만 GPU 유휴 상태 문제를 해결하기 위해 들어오는 배치를 마이크로 배치로 나누고 인공적으로 파이프라인을 생성하여 서로 다른 GPU가 동시에 계산에 참여할 수 있게 합니다.
+
+[GPipe 논문](https://ai.googleblog.com/2019/03/introducing-gpipe-open-source-library.html)에서 가져온 그림은 상단에 naive MP를, 하단에는 PP를 보여줍니다:
+
+![mp-pp](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-gpipe-bubble.png)
+
+하단 다이어그램에서 PP가 유휴 영역이 적은 것을 쉽게 볼 수 있습니다. 유휴 부분을 "bubble"이라고 합니다.
+
+다이어그램의 양쪽 부분은 참여하는 GPU가 4개인 병렬성을 보여줍니다. 즉, 4개의 GPU가 파이프라인에 참여합니다. 따라서 4개의 파이프 단계 F0, F1, F2 및 F3의 순방향 경로와 B3, B2, B1 및 B0의 역방향 경로가 있습니다.
+
+PP는 조정해야 할 새로운 하이퍼파라미터인 `chunks`를 도입합니다. 이는 동일한 파이프 단계를 통해 일련의 데이터를 묶어서 보내는 방식을 정의합니다. 예를 들어, 아래 다이어그램에서 `chunks=4`를 볼 수 있습니다. GPU0은 0, 1, 2 및 3 (F0,0, F0,1, F0,2, F0,3) 묶음에서 동일한 순방향 경로를 수행하고, 다른 GPU가 작업을 수행하기 시작하고 완료가 시작될 때만 GPU0이 묶음의 역순으로 3, 2, 1 및 0 (B0,3, B0,2, B0,1, B0,0) 경로를 수행합니다.
+
+개념적으로 이는 그래디언트 누적 단계 (GAS)와 동일한 개념입니다. 파이토치에서는 `chunks`를 사용하고 DeepSpeed에서는 동일한 하이퍼파라미터를 GAS로 참조합니다.
+
+묶음으로 인해 PP는 마이크로 배치 (MBS)의 개념을 도입합니다. DP는 전역 데이터 배치 크기를 미니 배치로 나눕니다. 따라서 DP 차수가 4이고 전역 배치 크기가 1024이면 256씩 4개의 미니 배치로 분할됩니다 (1024/4). 그리고 `chunks` (또는 GAS)의 수가 32이면 마이크로 배치 크기는 8이 됩니다 (256/32). 각 파이프라인 단계는 한 번에 하나의 마이크로 배치와 함께 작동합니다.
+
+DP + PP 설정의 전역 배치 크기를 계산하려면 `mbs*chunks*dp_degree` (`8*32*4=1024`)를 수행합니다.
+
+다이어그램으로 돌아가 보겠습니다.
+
+`chunks=1`로 설정하면 매우 비효율적인 naive MP가 생성되며, 매우 큰 `chunks` 값으로 설정하면 아주 작은 마이크로 배치 크기가 생성되어 효율적이지 않을 수 있습니다. 따라서 가장 효율적인 GPU 활용을 위해 어떤 값이 가장 적절한지 실험을 해야 합니다.
+
+다이어그램에서 보이는 것처럼 "dead" 시간의 버블이 존재하여 마지막 `forward` 단계가 `backward` 단계가 파이프라인을 완료하기를 기다려야 하는 상황이 발생하지만, `chunks`의 가장 적절한 값을 찾는 것의 목적은 모든 참여하는 GPU에서 동시에 고도로 활용되는 GPU 활용을 가능하게 하여 버블의 크기를 최소화하는 것입니다.
+
+해결책은 전통적인 파이프라인 API와 더 현대적인 솔루션으로 나뉩니다. 전통적인 파이프라인 API 솔루션과 현대적인 솔루션에 대해 알아보겠습니다.
+
+전통적인 파이프라인 API 솔루션:
+- 파이토치
+- FairScale
+- DeepSpeed
+- Megatron-LM
+
+현대적인 솔루션:
+- Varuna
+- Sagemaker
+
+전통적인 파이프라인 API 솔루션의 문제점:
+- 모델을 상당히 수정해야 한다는 점이 문제입니다. 파이프라인은 모듈의 정상적인 흐름을 `nn.Sequential` 시퀀스로 다시 작성해야 하므로 모델의 설계를 변경해야 할 수 있습니다.
+- 현재 파이프라인 API는 매우 제한적입니다. 파이프라인의 매우 첫 번째 단계에서 전달되는 많은 파이썬 변수가 있는 경우 이를 해결해야 합니다. 현재 파이프라인 인터페이스는 하나의 텐서 또는 텐서의 튜플을 유일한 입력 및 출력으로 요구합니다. 이러한 텐서는 마이크로 배치로 미니 배치로 묶을 것이므로 첫 번째 차원으로 배치 크기가 있어야 합니다. 가능한 개선 사항은 여기에서 논의되고 있습니다. https://github.com/pytorch/pytorch/pull/50693
+- 파이프 단계 수준에서 조건부 제어 흐름은 불가능합니다. 예를 들어, T5와 같은 인코더-디코더 모델은 조건부 인코더 단계를 처리하기 위해 특별한 해결책이 필요합니다.
+- 각 레이어를 정렬하여 하나의 모델의 출력이 다른 모델의 입력이 되도록해야 합니다.
+
+우리는 아직 Varuna와 SageMaker로 실험하지 않았지만, 해당 논문들은 위에서 언급한 문제들의 목록을 극복했고 사용자의 모델에 대한 변경 사항이 훨씬 적게 필요하다고 보고하고 있습니다.
+
+구현:
+- [파이토치](https://pytorch.org/docs/stable/pipeline.html) (파이토치-1.8에서 초기 지원, 1.9에서 점진적으로 개선되고 1.10에서 더 개선됨). [예제](https://github.com/pytorch/pytorch/blob/master/benchmarks/distributed/pipeline/pipe.py)도 참고하세요.
+- [FairScale](https://fairscale.readthedocs.io/en/latest/tutorials/pipe.html)
+- [DeepSpeed](https://www.deepspeed.ai/tutorials/pipeline/)
+- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)은 내부 구현을 가지고 있습니다 - API 없음.
+- [Varuna](https://github.com/microsoft/varuna)
+- [SageMaker](https://arxiv.org/abs/2111.05972) - 이는 AWS에서만 사용할 수 있는 소유 솔루션입니다.
+- [OSLO](https://github.com/tunib-ai/oslo) - 이는 Hugging Face Transformers를 기반으로 구현된 파이프라인 병렬화입니다.
+
+🤗 Transformers 상태: 이 작성 시점에서 모델 중 어느 것도 완전한 PP를 지원하지 않습니다. GPT2와 T5 모델은 naive MP를 지원합니다. 주요 장애물은 모델을 `nn.Sequential`로 변환하고 모든 입력을 텐서로 가져와야 하는 것을 처리할 수 없기 때문입니다. 현재 모델에는 이러한 변환을 매우 복잡하게 만드는 많은 기능이 포함되어 있어 제거해야 합니다.
+
+기타 접근 방법:
+
+DeepSpeed, Varuna 및 SageMaker는 [교차 파이프라인(Interleaved Pipeline)](https://docs.aws.amazon.com/sagemaker/latest/dg/model-parallel-core-features.html) 개념을 사용합니다.
+![interleaved-pipeline-execution](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-sagemaker-interleaved-pipeline.png)
+
+여기서는 버블(유휴 시간)을 역방향 패스에 우선순위를 부여하여 최소화합니다.
+
+Varuna는 가장 효율적인 스케줄링을 찾기 위해 시뮬레이션을 사용하여 스케줄을 개선하려고 합니다.
+
+OSLO는 `nn.Sequential`로 변환하지 않고 Transformers를 기반으로 한 파이프라인 병렬화를 구현했습니다.
+
+## 텐서 병렬 처리 [[tensor-parallelism]]
+
+텐서 병렬 처리에서는 각 GPU가 텐서의 일부분만 처리하고 전체 텐서가 필요한 연산에 대해서만 전체 텐서를 집계합니다.
+
+이 섹션에서는 [Megatron-LM](https://github.com/NVIDIA/Megatron-LM) 논문인 [Efficient Large-Scale Language Model Training on GPU Clusters](https://arxiv.org/abs/2104.04473)에서의 개념과 다이어그램을 사용합니다.
+
+Transformer의 주요 구성 요소는 fully connected `nn.Linear`와 비선형 활성화 함수인 `GeLU`입니다.
+
+Megatron 논문의 표기법을 따라 행렬의 점곱 부분을 `Y = GeLU(XA)`로 표현할 수 있습니다. 여기서 `X`와 `Y`는 입력 및 출력 벡터이고 `A`는 가중치 행렬입니다.
+
+행렬 형태로 계산을 살펴보면, 행렬 곱셈을 다중 GPU로 분할할 수 있는 방법을 쉽게 알 수 있습니다:
+![Parallel GEMM](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-tp-parallel_gemm.png)
+
+가중치 행렬 `A`를 `N`개의 GPU에 대해 열별로 분할하고 병렬로 행렬 곱셈 `XA_1`에서 `XA_n`까지 수행하면 `N`개의 출력 벡터 `Y_1, Y_2, ..., Y_n`가 생성되며 독립적으로 `GeLU`에 전달될 수 있습니다:
+![independent GeLU](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-tp-independent-gelu.png)
+
+이 원리를 사용하여 동기화가 필요하지 않은 GPU 간의 임의 깊이의 MLP를 업데이트할 수 있습니다. 그러나 결과 벡터를 샤드로부터 재구성해야 하는 마지막 단계까지는 GPU 간의 동기화가 필요합니다. Megatron-LM 논문의 저자들은 이에 대한 유용한 그림을 제공합니다:
+![parallel shard processing](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-tp-parallel_shard_processing.png)
+
+다중 헤드 어텐션 레이어의 병렬화는 더욱 간단합니다. 이미 독립적인 다중 헤드를 가지고 있기 때문에 이미 병렬화되어 있습니다!
+![parallel self-attention](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-tp-parallel_self_attention.png)
+
+특별 고려사항: TP는 매우 빠른 네트워크가 필요하므로 한 개 이상의 노드에서 TP를 수행하는 것은 권장되지 않습니다. 실제로 노드에 4개의 GPU가 있는 경우 TP의 최대 차수는 4입니다. TP 차수가 8인 경우 최소한 8개의 GPU가 있는 노드를 사용해야 합니다.
+
+이 섹션은 원래의 [더 자세한 TP 개요](https://github.com/huggingface/transformers/issues/10321#issuecomment-783543530)를 기반으로 합니다. 
+작성자는 [@anton-l](https://github.com/anton-l)입니다.
+
+SageMaker는 더 효율적인 처리를 위해 TP와 DP를 결합합니다.
+
+대체 이름:
+- DeepSpeed는 이를 [텐서 슬라이싱](https://www.deepspeed.ai/features/#model-parallelism)이라고 부릅니다.
+
+구현:
+- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)은 내부 구현을 가지고 있으므로 모델에 매우 특화되어 있습니다.
+- [parallelformers](https://github.com/tunib-ai/parallelformers) (현재는 추론에만 해당)
+- [SageMaker](https://arxiv.org/abs/2111.05972) - 이는 AWS에서만 사용할 수 있는 소유 솔루션입니다.
+- [OSLO](https://github.com/tunib-ai/oslo)은 Transformers를 기반으로 한 텐서 병렬 처리 구현을 가지고 있습니다.
+
+🤗 Transformers 현황:
+- core: 아직 핵심 부분에 구현되지 않음
+- 그러나 추론을 하려면 [parallelformers](https://github.com/tunib-ai/parallelformers)가 대부분의 모델을 지원합니다. 따라서 핵심 부분에 구현되기 전까지 그들의 것을 사용할 수 있습니다. 그리고 훈련 모드도 지원될 예정입니다.
+- Deepspeed-Inference는 CUDA 커널을 기반으로 하는 매우 빠른 추론 모드에서 BERT, GPT-2 및 GPT-Neo 모델을 지원합니다. 자세한 내용은 [여기](https://www.deepspeed.ai/tutorials/inference-tutorial/)를 참조하세요.
+
+## DP+PP [[dppp]]
+
+DeepSpeed [pipeline tutorial](https://www.deepspeed.ai/tutorials/pipeline/)에서 다음 다이어그램은 DP와 PP를 결합하는 방법을 보여줍니다.
+
+![dp-pp-2d](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-zero-dp-pp.png)
+
+여기서 DP 랭크 0은 GPU2를 보지 못하고, DP 랭크 1은 GPU3을 보지 못하는 것이 중요합니다. DP에게는 딱 2개의 GPU인 것처럼 데이터를 공급합니다. GPU0은 PP를 사용하여 GPU2에게 일부 작업을 "비밀리에" 할당합니다. 그리고 GPU1도 GPU3을 도움으로 삼아 같은 방식으로 작업합니다.
+
+각 차원마다 적어도 2개의 GPU가 필요하므로 최소한 4개의 GPU가 필요합니다.
+
+구현:
+- [DeepSpeed](https://github.com/microsoft/DeepSpeed)
+- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
+- [Varuna](https://github.com/microsoft/varuna)
+- [SageMaker](https://arxiv.org/abs/2111.05972)
+- [OSLO](https://github.com/tunib-ai/oslo)
+
+🤗 Transformers 현황: 아직 구현되지 않음
+
+## DP+PP+TP [[dppptp]]
+
+더 효율적인 훈련을 위해 PP와 TP 및 DP를 결합하여 3D 병렬 처리를 사용합니다. 다음 다이어그램에서 이를 확인할 수 있습니다.
+
+![dp-pp-tp-3d](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-deepspeed-3d.png)
+
+이 다이어그램은 [3D parallelism: Scaling to trillion-parameter models](https://www.microsoft.com/en-us/research/blog/deepspeed-extreme-scale-model-training-for-everyone/)이라는 블로그 글에서 확인할 수 있습니다.
+
+각 차원마다 적어도 2개의 GPU가 필요하므로 최소한 8개의 GPU가 필요합니다.
+
+구현:
+- [DeepSpeed](https://github.com/microsoft/DeepSpeed) - DeepSpeed는 더욱 효율적인 DP인 ZeRO-DP라고도 부릅니다.
+- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
+- [Varuna](https://github.com/microsoft/varuna)
+- [SageMaker](https://arxiv.org/abs/2111.05972)
+- [OSLO](https://github.com/tunib-ai/oslo)
+
+🤗 Transformers 현황: 아직 구현되지 않음. PP와 TP가 없기 때문입니다.
+
+## ZeRO DP+PP+TP [[zero-dppptp]]
+
+DeepSpeed의 주요 기능 중 하나는 DP의 확장인 ZeRO입니다. ZeRO-DP에 대해 이미 [ZeRO Data Parallelism](#zero-data-parallelism)에서 논의되었습니다. 일반적으로 이는 PP나 TP를 필요로하지 않는 독립적인 기능입니다. 그러나 PP와 TP와 결합할 수도 있습니다.
+
+ZeRO-DP가 PP와 (선택적으로 TP와) 결합되면 일반적으로 ZeRO 단계 1(옵티마이저 분할)만 활성화됩니다.
+
+이론적으로는 ZeRO 단계 2(그라디언트 분할)를 파이프라인 병렬 처리와 함께 사용할 수도 있지만, 이는 성능에 나쁜 영향을 미칠 것입니다. 각 마이크로 배치마다 그라디언트를 샤딩하기 전에 추가적인 리듀스-스캐터 컬렉티브가 필요하며, 이는 잠재적으로 상당한 통신 오버헤드를 추가합니다. 파이프라인 병렬 처리의 특성상 작은 마이크로 배치가 사용되며, 산술 연산 강도(마이크로 배치 크기)를 균형 있게 유지하면서 파이프라인 버블(마이크로 배치 수)을 최소화하는 것에 중점을 둡니다. 따라서 해당 통신 비용은 문제가 될 것입니다.
+
+또한, PP로 인해 정상보다 적은 수의 레이어가 있으므로 메모리 절약은 크지 않을 것입니다. PP는 이미 그래디언트 크기를 ``1/PP``로 줄이기 때문에 그래디언트 샤딩의 절약 효과는 순수 DP보다는 미미합니다.
+
+ZeRO 단계 3도 같은 이유로 좋은 선택이 아닙니다 - 더 많은 노드 간 통신이 필요합니다.
+
+그리고 ZeRO가 있기 때문에 다른 이점은 ZeRO-Offload입니다. 이는 단계 1이므로 옵티마이저 상태를 CPU로 오프로드할 수 있습니다.
+
+구현:
+- [Megatron-DeepSpeed](https://github.com/microsoft/Megatron-DeepSpeed) 및 [BigScience의 Megatron-Deepspeed](https://github.com/bigscience-workshop/Megatron-DeepSpeed), 이전 저장소의 포크입니다.
+- [OSLO](https://github.com/tunib-ai/oslo)
+
+중요한 논문:
+
+- [Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model](
+https://arxiv.org/abs/2201.11990)
+
+🤗 Transformers 현황: 아직 구현되지 않음, PP와 TP가 없기 때문입니다.
+
+## FlexFlow [[flexflow]]
+
+[FlexFlow](https://github.com/flexflow/FlexFlow)는 약간 다른 방식으로 병렬화 문제를 해결합니다.
+
+논문: ["Beyond Data and Model Parallelism for Deep Neural Networks" by Zhihao Jia, Matei Zaharia, Alex Aiken](https://arxiv.org/abs/1807.05358)
+
+이는 Sample-Operator-Attribute-Parameter를 기반으로 하는 일종의 4D 병렬화를 수행합니다.
+
+1. Sample = 데이터 병렬화 (샘플별 병렬)
+2. Operator = 단일 연산을 여러 하위 연산으로 병렬화
+3. Attribute = 데이터 병렬화 (길이별 병렬)
+4. Parameter = 모델 병렬화 (수평 또는 수직과 관계없이)
+
+예시:
+* Sample
+
+512 길이의 10개의 배치를 가정해 봅시다. 이를 sample 차원으로 2개의 장치에 병렬화하면, 10 x 512는 5 x 2 x 512가 됩니다.
+
+* Operator
+
+레이어 정규화를 수행한다면, 우선 std를 계산하고 두 번째로 mean을 계산한 다음 데이터를 정규화할 수 있습니다. Operator 병렬화는 std와 mean을 병렬로 계산할 수 있도록 합니다. 따라서 operator 차원으로 2개의 장치 (cuda:0, cuda:1)에 병렬화하면, 먼저 입력 데이터를 두 장치로 복사한 다음 cuda:0에서 std를 계산하고 cuda:1에서 동시에 mean을 계산합니다.
+
+* Attribute
+
+512 길이의 10개의 배치가 있습니다. 이를 attribute 차원으로 2개의 장치에 병렬화하면, 10 x 512는 10 x 2 x 256이 됩니다.
+
+* Parameter
+
+이는 tensor 모델 병렬화 또는 naive layer-wise 모델 병렬화와 유사합니다.
+
+![flex-flow-soap](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/parallelism-flexflow.jpeg)
+
+이 프레임워크의 중요한 점은 (1) GPU/TPU/CPU 대 (2) RAM/DRAM 대 (3) 빠른 인트라-커넥트 대 느린 인터-커넥트와 같은 리소스를 고려하여 어디에서 어떤 병렬화를 사용할지를 알고리즘적으로 자동으로 최적화한다는 것입니다.
+
+하나 매우 중요한 측면은 FlexFlow가 정적이고 고정된 워크로드를 가진 모델에 대한 DNN 병렬화를 최적화하기 위해 설계되었다는 것입니다. 동적인 동작을 가진 모델은 반복마다 다른 병렬화 전략을 선호할 수 있습니다.
+
+따라서 이 프레임워크의 장점은 선택한 클러스터에서 30분 동안 시뮬레이션을 실행하고 이 특정 환경을 최적으로 활용하기 위한 최상의 전략을 제안한다는 것입니다. 부품을 추가/제거/교체하면 실행하고 그에 대한 계획을 다시 최적화한 후 훈련할 수 있습니다. 다른 설정은 자체적인 사용자 정의 최적화를 가질 수 있습니다.
+
+🤗 Transformers 현황: 아직 통합되지 않음. 이미 [transformers.utils.fx](https://github.com/huggingface/transformers/blob/master/src/transformers/utils/fx.py)를 통해 모델을 FX-추적할 수 있으며, 이는 FlexFlow의 선행 조건입니다. 따라서 어떤 작업을 수행해야 FlexFlow가 우리의 모델과 함께 작동할 수 있는지 파악해야 합니다.
+
+
+## 어떤 전략을 사용해야 할까요? [[which-strategy-to-use-when]]
+
+다음은 어떤 병렬화 전략을 언제 사용해야 하는지에 대한 매우 대략적인 개요입니다. 각 목록의 첫 번째 전략이 일반적으로 더 빠릅니다.
+
+**⇨ 단일 GPU**
+
+* 모델이 단일 GPU에 맞는 경우:
+
+    1. 일반적인 사용
+
+* 모델이 단일 GPU에 맞지 않는 경우:
+
+    1. ZeRO + CPU 및 옵션으로 NVMe 언로드
+    2. 위와 동일하게 사용하되, 가장 큰 레이어가 단일 GPU에 맞지 않는 경우 Memory Centric Tiling(자세한 내용은 아래 참조)을 추가적으로 사용
+
+* 가장 큰 레이어가 단일 GPU에 맞지 않는 경우:
+
+1. ZeRO - [Memory Centric Tiling](https://deepspeed.readthedocs.io/en/latest/zero3.html#memory-centric-tiling) (MCT) 활성화. 이를 통해 크기가 매우 큰 레이어를 임의로 분할하여 순차적으로 실행할 수 있습니다. MCT는 GPU에 활성화된 매개변수의 수를 줄이지만 활성화 메모리에는 영향을 주지 않습니다. 현재 작성 기준으로 이 요구사항은 매우 드물기 때문에 사용자가 `torch.nn.Linear`를 수동으로 수정해야 합니다.
+
+**⇨ 단일 노드 / 다중 GPU**
+
+* 모델이 단일 GPU에 맞는 경우:
+
+    1. DDP - 분산 DP
+    2. ZeRO - 상황과 구성에 따라 빠를 수도 있고 그렇지 않을 수도 있습니다.
+
+* 모델이 단일 GPU에 맞지 않는 경우:
+
+    1. PP
+    2. ZeRO
+    3. TP
+
+    NVLINK 또는 NVSwitch를 통한 매우 빠른 인트라-노드 연결이 있는 경우 이 세 가지 방법은 거의 동등할 것이며, 이러한 연결이 없는 경우 PP가 TP나 ZeRO보다 빠를 것입니다. 또한 TP의 차수도 영향을 줄 수 있습니다. 특정 설정에서 우승자를 찾기 위해 실험하는 것이 가장 좋습니다.
+
+    TP는 거의 항상 단일 노드 내에서 사용됩니다. 즉, TP 크기 <= 노드당 GPU 수입니다.
+
+* 가장 큰 레이어가 단일 GPU에 맞지 않는 경우:
+
+    1. ZeRO를 사용하지 않을 경우 - PP만 사용할 수 없으므로 TP를 사용해야 합니다.
+    2. ZeRO를 사용할 경우, "단일 GPU"의 항목과 동일한 항목 참조
+
+
+**⇨ 다중 노드 / 다중 GPU**
+
+* 빠른 노드 간 연결이 있는 경우:
+
+    1. ZeRO - 모델에 대한 수정이 거의 필요하지 않습니다.
+    2. PP+TP+DP - 통신이 적지만 모델에 대한 대규모 변경이 필요합니다.
+
+* 느린 노드 간 연결 및 GPU 메모리 부족한 경우:
+
+    1. DP+PP+TP+ZeRO-1

docs/source/ko/tasks/audio_classification.md

@@ -0,0 +1,329 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# 오디오 분류[[audio_classification]]
+
+[[open-in-colab]]
+
+<Youtube id="KWwzcmG98Ds"/>
+
+오디오 분류는 텍스트와 마찬가지로 입력 데이터에 클래스 레이블 출력을 할당합니다. 유일한 차이점은 텍스트 입력 대신 원시 오디오 파형이 있다는 것입니다. 오디오 분류의 실제 적용 분야에는 화자의 의도 파악, 언어 분류, 소리로 동물 종을 식별하는 것 등이 있습니다.
+
+이 문서에서 방법을 알아보겠습니다:
+
+1. [MInDS-14](https://huggingface.co/datasets/PolyAI/minds14) 데이터 세트를 [Wav2Vec2](https://huggingface.co/facebook/wav2vec2-base)로 미세 조정하여 화자의 의도를 분류합니다.
+2. 추론에 미세 조정된 모델을 사용하세요.
+
+<Tip>
+이 튜토리얼에서 설명하는 작업은 아래의 모델 아키텍처에서 지원됩니다:
+
+<!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
+
+[Audio Spectrogram Transformer](../model_doc/audio-spectrogram-transformer), [Data2VecAudio](../model_doc/data2vec-audio), [Hubert](../model_doc/hubert), [SEW](../model_doc/sew), [SEW-D](../model_doc/sew-d), [UniSpeech](../model_doc/unispeech), [UniSpeechSat](../model_doc/unispeech-sat), [Wav2Vec2](../model_doc/wav2vec2), [Wav2Vec2-Conformer](../model_doc/wav2vec2-conformer), [WavLM](../model_doc/wavlm), [Whisper](../model_doc/whisper)
+
+<!--End of the generated tip-->
+
+</Tip>
+
+시작하기 전에 필요한 라이브러리가 모두 설치되어 있는지 확인하세요:
+
+```bash
+pip install transformers datasets evaluate
+```
+
+모델을 업로드하고 커뮤니티와 공유할 수 있도록 허깅페이스 계정에 로그인하는 것이 좋습니다. 메시지가 표시되면 토큰을 입력하여 로그인합니다:
+
+```py
+>>> from huggingface_hub import notebook_login
+
+>>> notebook_login()
+```
+
+## MInDS-14 데이터셋 불러오기[[load_minds_14_dataset]]
+
+먼저 🤗 Datasets 라이브러리에서 MinDS-14 데이터 세트를 가져옵니다:
+
+```py
+>>> from datasets import load_dataset, Audio
+
+>>> minds = load_dataset("PolyAI/minds14", name="en-US", split="train")
+```
+
+데이터 세트의 `train` 분할을 [`~datasets.Dataset.train_test_split`] 메소드를 사용하여 더 작은 훈련 및 테스트 집합으로 분할합니다. 이렇게 하면 전체 데이터 세트에 더 많은 시간을 소비하기 전에 모든 것이 작동하는지 실험하고 확인할 수 있습니다.
+
+```py
+>>> minds = minds.train_test_split(test_size=0.2)
+```
+
+이제 데이터 집합을 살펴볼게요:
+
+```py
+>>> minds
+DatasetDict({
+    train: Dataset({
+        features: ['path', 'audio', 'transcription', 'english_transcription', 'intent_class', 'lang_id'],
+        num_rows: 450
+    })
+    test: Dataset({
+        features: ['path', 'audio', 'transcription', 'english_transcription', 'intent_class', 'lang_id'],
+        num_rows: 113
+    })
+})
+```
+
+데이터 세트에는 `lang_id` 및 `english_transcription`과 같은 유용한 정보가 많이 포함되어 있지만 이 가이드에서는 `audio` 및 `intent_class`에 중점을 둘 것입니다. 다른 열은 [`~datasets.Dataset.remove_columns`] 메소드를 사용하여 제거합니다:
+
+```py
+>>> minds = minds.remove_columns(["path", "transcription", "english_transcription", "lang_id"])
+```
+
+예시를 살펴보겠습니다:
+
+```py
+>>> minds["train"][0]
+{'audio': {'array': array([ 0.        ,  0.        ,  0.        , ..., -0.00048828,
+         -0.00024414, -0.00024414], dtype=float32),
+  'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~APP_ERROR/602b9a5fbb1e6d0fbce91f52.wav',
+  'sampling_rate': 8000},
+ 'intent_class': 2}
+```
+
+두 개의 필드가 있습니다:
+
+- `audio`: 오디오 파일을 가져오고 리샘플링하기 위해 호출해야 하는 음성 신호의 1차원 `배열`입니다.
+- `intent_class`: 화자의 의도에 대한 클래스 ID를 나타냅니다.
+
+모델이 레이블 ID에서 레이블 이름을 쉽게 가져올 수 있도록 레이블 이름을 정수로 매핑하는 사전을 만들거나 그 반대로 매핑하는 사전을 만듭니다:
+
+```py
+>>> labels = minds["train"].features["intent_class"].names
+>>> label2id, id2label = dict(), dict()
+>>> for i, label in enumerate(labels):
+...     label2id[label] = str(i)
+...     id2label[str(i)] = label
+```
+
+이제 레이블 ID를 레이블 이름으로 변환할 수 있습니다:
+
+```py
+>>> id2label[str(2)]
+'app_error'
+```
+
+## 전처리[[preprocess]]
+
+다음 단계는 오디오 신호를 처리하기 위해 Wav2Vec2 특징 추출기를 가져오는 것입니다:
+
+```py
+>>> from transformers import AutoFeatureExtractor
+
+>>> feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base")
+```
+
+MinDS-14 데이터 세트의 샘플링 속도는 8000khz이므로(이 정보는 [데이터세트 카드](https://huggingface.co/datasets/PolyAI/minds14)에서 확인할 수 있습니다), 사전 훈련된 Wav2Vec2 모델을 사용하려면 데이터 세트를 16000kHz로 리샘플링해야 합니다:
+
+```py
+>>> minds = minds.cast_column("audio", Audio(sampling_rate=16_000))
+>>> minds["train"][0]
+{'audio': {'array': array([ 2.2098757e-05,  4.6582241e-05, -2.2803260e-05, ...,
+         -2.8419291e-04, -2.3305941e-04, -1.1425107e-04], dtype=float32),
+  'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~APP_ERROR/602b9a5fbb1e6d0fbce91f52.wav',
+  'sampling_rate': 16000},
+ 'intent_class': 2}
+```
+
+이제 전처리 함수를 만듭니다:
+
+1. 가져올 `오디오` 열을 호출하고 필요한 경우 오디오 파일을 리샘플링합니다.
+2. 오디오 파일의 샘플링 속도가 모델에 사전 훈련된 오디오 데이터의 샘플링 속도와 일치하는지 확인합니다. 이 정보는 Wav2Vec2 [모델 카드](https://huggingface.co/facebook/wav2vec2-base)에서 확인할 수 있습니다.
+3. 긴 입력이 잘리지 않고 일괄 처리되도록 최대 입력 길이를 설정합니다.
+
+```py
+>>> def preprocess_function(examples):
+...     audio_arrays = [x["array"] for x in examples["audio"]]
+...     inputs = feature_extractor(
+...         audio_arrays, sampling_rate=feature_extractor.sampling_rate, max_length=16000, truncation=True
+...     )
+...     return inputs
+```
+
+전체 데이터 세트에 전처리 기능을 적용하려면 🤗 Datasets [`~datasets.Dataset.map`] 함수를 사용합니다. `batched=True`를 설정하여 데이터 집합의 여러 요소를 한 번에 처리하면 `map`의 속도를 높일 수 있습니다. 필요하지 않은 열을 제거하고 `intent_class`의 이름을 모델이 예상하는 이름인 `label`로 변경합니다:
+
+```py
+>>> encoded_minds = minds.map(preprocess_function, remove_columns="audio", batched=True)
+>>> encoded_minds = encoded_minds.rename_column("intent_class", "label")
+```
+
+## 평가하기[[evaluate]]
+
+훈련 중에 메트릭을 포함하면 모델의 성능을 평가하는 데 도움이 되는 경우가 많습니다. 🤗 [Evaluate](https://huggingface.co/docs/evaluate/index) 라이브러리를 사용하여 평가 방법을 빠르게 가져올 수 있습니다. 이 작업에서는 [accuracy(정확도)](https://huggingface.co/spaces/evaluate-metric/accuracy) 메트릭을 가져옵니다(메트릭을 가져오고 계산하는 방법에 대한 자세한 내용은 🤗 Evalutate [빠른 둘러보기](https://huggingface.co/docs/evaluate/a_quick_tour) 참조하세요):
+
+```py
+>>> import evaluate
+
+>>> accuracy = evaluate.load("accuracy")
+```
+
+그런 다음 예측과 레이블을 [`~evaluate.EvaluationModule.compute`]에 전달하여 정확도를 계산하는 함수를 만듭니다:
+
+```py
+>>> import numpy as np
+
+
+>>> def compute_metrics(eval_pred):
+...     predictions = np.argmax(eval_pred.predictions, axis=1)
+...     return accuracy.compute(predictions=predictions, references=eval_pred.label_ids)
+```
+
+이제 `compute_metrics` 함수를 사용할 준비가 되었으며, 트레이닝을 설정할 때 이 함수를 사용합니다.
+
+## 훈련[[train]]
+
+<frameworkcontent>
+<pt>
+<Tip>
+
+[`Trainer`]로 모델을 미세 조정하는 데 익숙하지 않다면 기본 튜토리얼 [여기](../training#train-with-pytorch-trainer)을 살펴보세요!
+
+</Tip>
+
+이제 모델 훈련을 시작할 준비가 되었습니다! [`AutoModelForAudioClassification`]을 이용해서 Wav2Vec2를 불러옵니다. 예상되는 레이블 수와 레이블 매핑을 지정합니다:
+
+```py
+>>> from transformers import AutoModelForAudioClassification, TrainingArguments, Trainer
+
+>>> num_labels = len(id2label)
+>>> model = AutoModelForAudioClassification.from_pretrained(
+...     "facebook/wav2vec2-base", num_labels=num_labels, label2id=label2id, id2label=id2label
+... )
+```
+
+이제 세 단계만 남았습니다:
+
+1. 훈련 하이퍼파라미터를 [`TrainingArguments`]에 정의합니다. 유일한 필수 매개변수는 모델을 저장할 위치를 지정하는 `output_dir`입니다. `push_to_hub = True`를 설정하여 이 모델을 허브로 푸시합니다(모델을 업로드하려면 허깅 페이스에 로그인해야 합니다). 각 에폭이 끝날 때마다 [`Trainer`]가 정확도를 평가하고 훈련 체크포인트를 저장합니다.
+2. 모델, 데이터 세트, 토크나이저, 데이터 콜레이터, `compute_metrics` 함수와 함께 훈련 인자를 [`Trainer`]에 전달합니다.
+3. [`~Trainer.train`]을 호출하여 모델을 미세 조정합니다.
+
+
+```py
+>>> training_args = TrainingArguments(
+...     output_dir="my_awesome_mind_model",
+...     evaluation_strategy="epoch",
+...     save_strategy="epoch",
+...     learning_rate=3e-5,
+...     per_device_train_batch_size=32,
+...     gradient_accumulation_steps=4,
+...     per_device_eval_batch_size=32,
+...     num_train_epochs=10,
+...     warmup_ratio=0.1,
+...     logging_steps=10,
+...     load_best_model_at_end=True,
+...     metric_for_best_model="accuracy",
+...     push_to_hub=True,
+... )
+
+>>> trainer = Trainer(
+...     model=model,
+...     args=training_args,
+...     train_dataset=encoded_minds["train"],
+...     eval_dataset=encoded_minds["test"],
+...     tokenizer=feature_extractor,
+...     compute_metrics=compute_metrics,
+... )
+
+>>> trainer.train()
+```
+
+훈련이 완료되면 모든 사람이 모델을 사용할 수 있도록 [`~transformers.Trainer.push_to_hub`] 메소드를 사용하여 모델을 허브에 공유하세요:
+
+```py
+>>> trainer.push_to_hub()
+```
+</pt>
+</frameworkcontent>
+
+<Tip>
+
+For a more in-depth example of how to finetune a model for audio classification, take a look at the corresponding [PyTorch notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/audio_classification.ipynb).
+
+</Tip>
+
+## 추론[[inference]]
+
+이제 모델을 미세 조정했으니 추론에 사용할 수 있습니다!
+
+추론을 실행할 오디오 파일을 가져옵니다. 필요한 경우 오디오 파일의 샘플링 속도를 모델의 샘플링 속도와 일치하도록 리샘플링하는 것을 잊지 마세요!
+
+```py
+>>> from datasets import load_dataset, Audio
+
+>>> dataset = load_dataset("PolyAI/minds14", name="en-US", split="train")
+>>> dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
+>>> sampling_rate = dataset.features["audio"].sampling_rate
+>>> audio_file = dataset[0]["audio"]["path"]
+```
+
+추론을 위해 미세 조정한 모델을 시험해 보는 가장 간단한 방법은 [`pipeline`]에서 사용하는 것입니다. 모델을 사용하여 오디오 분류를 위한 `pipeline`을 인스턴스화하고 오디오 파일을 전달합니다:
+
+```py
+>>> from transformers import pipeline
+
+>>> classifier = pipeline("audio-classification", model="stevhliu/my_awesome_minds_model")
+>>> classifier(audio_file)
+[
+    {'score': 0.09766869246959686, 'label': 'cash_deposit'},
+    {'score': 0.07998877018690109, 'label': 'app_error'},
+    {'score': 0.0781070664525032, 'label': 'joint_account'},
+    {'score': 0.07667109370231628, 'label': 'pay_bill'},
+    {'score': 0.0755252093076706, 'label': 'balance'}
+]
+```
+
+원하는 경우 `pipeline`의 결과를 수동으로 복제할 수도 있습니다:
+
+<frameworkcontent>
+<pt>
+특징 추출기를 가져와서 오디오 파일을 전처리하고 `입력`을 PyTorch 텐서로 반환합니다:
+
+```py
+>>> from transformers import AutoFeatureExtractor
+
+>>> feature_extractor = AutoFeatureExtractor.from_pretrained("stevhliu/my_awesome_minds_model")
+>>> inputs = feature_extractor(dataset[0]["audio"]["array"], sampling_rate=sampling_rate, return_tensors="pt")
+```
+
+모델에 입력을 전달하고 로짓을 반환합니다:
+
+```py
+>>> from transformers import AutoModelForAudioClassification
+
+>>> model = AutoModelForAudioClassification.from_pretrained("stevhliu/my_awesome_minds_model")
+>>> with torch.no_grad():
+...     logits = model(**inputs).logits
+```
+
+확률이 가장 높은 클래스를 가져온 다음 모델의 `id2label` 매핑을 사용하여 이를 레이블로 변환합니다:
+
+```py
+>>> import torch
+
+>>> predicted_class_ids = torch.argmax(logits).item()
+>>> predicted_label = model.config.id2label[predicted_class_ids]
+>>> predicted_label
+'cash_deposit'
+```
+</pt>
+</frameworkcontent>

docs/source/ko/tokenizer_summary.md

@@ -0,0 +1,253 @@
+<!--Copyright 2020 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.
+
+-->
+
+# 토크나이저 요약[[summary-of-the-tokenizers]]
+
+[[open-in-colab]]
+
+이 페이지에서는 토큰화에 대해 자세히 살펴보겠습니다.
+
+<Youtube id="VFp38yj8h3A"/>
+
+[데이터 전처리하기 튜토리얼](preprocessing)에서 살펴본 것처럼, 텍스트를 토큰화하는 것은 텍스트를 단어 또는 서브워드로 분할하고 룩업 테이블을 통해 id로 변환하는 과정입니다.
+단어 또는 서브워드를 id로 변환하는 것은 간단하기 때문에 이번 문서에서는 텍스트를 단어 또는 서브워드로 쪼개는 것(즉, 텍스트를 토큰화하는 것)에 중점을 두겠습니다.
+구체적으로, 🤗 Transformers에서 사용되는 세 가지 주요 토큰화 유형인 [Byte-Pair Encoding (BPE)](#byte-pair-encoding), [WordPiece](#wordpiece), [SentencePiece](#sentencepiece)를 살펴보고 어떤 모델에서 어떤 토큰화 유형을 사용하는지 예시를 보여드리겠습니다.
+
+각 모델 페이지에 연결된 토크나이저의 문서를 보면 사전 훈련 모델에서 어떤 토크나이저를 사용했는지 알 수 있습니다.
+예를 들어, [`BertTokenizer`]를 보면 이 모델이 [WordPiece](#wordpiece)를 사용하는 것을 알 수 있습니다.
+
+## 개요[[introduction]]
+
+텍스트를 작은 묶음(chunk)으로 쪼개는 것은 보기보다 어려운 작업이며, 여러 가지 방법이 있습니다.
+예를 들어, `"Don't you love 🤗 Transformers? We sure do."` 라는 문장을 살펴보도록 하겠습니다.
+
+<Youtube id="nhJxYji1aho"/>
+
+위 문장을 토큰화하는 간단한 방법은 공백을 기준으로 쪼개는 것입니다.
+토큰화된 결과는 다음과 같습니다:
+
+```
+["Don't", "you", "love", "🤗", "Transformers?", "We", "sure", "do."]
+```
+이는 첫 번째 결과로는 합리적이지만, `"Transformers?"`와 `"do."`토큰을 보면 각각 `"Transformer"`와 `"do"`에 구두점이 붙어있는 것을 확인할 수 있습니다.
+구두점을 고려해야 모델이 단어의 다른 표현과 그 뒤에 올 수 있는 모든 가능한 구두점을 학습할 필요가 없습니다. 그렇지 않으면 모델이 학습해야 하는 표현의 수가 폭발적으로 증가하게 됩니다.
+
+구두점을 고려한 토큰화 결과는 다음과 같습니다:
+
+```
+["Don", "'", "t", "you", "love", "🤗", "Transformers", "?", "We", "sure", "do", "."]
+```
+
+이전보다 나아졌습니다. 하지만, `"Don't"`의 토큰화 결과도 수정이 필요합니다.
+`"Don't"`는 `"do not"`의 줄임말이기 때문에 `["Do", "n't"]`로 토큰화되는 것이 좋습니다.
+여기서부터 복잡해지기 시작합니다. 그리고 이 점이 각 모델마다 고유한 토큰화 유형이 존재하는 이유 중 하나입니다.
+텍스트를 토큰화하는 데 적용하는 규칙에 따라 동일한 텍스트에 대해 토큰화된 결과가 달라집니다.
+사전 훈련된 모델은 훈련 데이터를 토큰화하는 데 사용된 것과 동일한 규칙으로 토큰화된 입력을 제공해야만 제대로 작동합니다.
+
+[spaCy](https://spacy.io/)와 [Moses](http://www.statmt.org/moses/?n=Development.GetStarted)는 유명한 규칙 기반 토크나이저입니다. 예제에 *spaCy*와 *Moses* 를 적용한 결과는 다음과 같습니다:
+
+```
+["Do", "n't", "you", "love", "🤗", "Transformers", "?", "We", "sure", "do", "."]
+```
+
+보시다시피 공백 및 구두점 토큰화와 규칙 기반 토큰화가 사용됩니다.
+공백 및 구두점, 규칙 기반 토큰화은 모두 단어 문장을 단어로 쪼개는 단어 토큰화에 해당합니다.
+이 토큰화 방법은 텍스트를 더 작은 묶음(chunk)로 분할하는 가장 직관적인 방법이지만, 대규모 텍스트 말뭉치에 대해서는 문제가 발생할 수 있습니다.
+이 경우 공백 및 구두점 토큰화는 일반적으로 매우 큰 어휘(사용된 모든 고유 단어와 토큰 집합)을 생성합니다.
+*예를 들어*, [Transformer XL](model_doc/transformerxl)은 공백 및 구두점 토큰화를 사용해 어휘(vocabulary) 크기가 267,735입니다!
+
+어휘 크기가 크면 모델에 입력 및 출력 레이어로 엄청난 임베딩 행렬이 필요하므로 메모리와 시간 복잡성이 모두 증가합니다.
+일반적으로 트랜스포머 모델은 어휘 크기가 50,000개를 넘는 경우가 드물며, 특히 단일 언어에 대해서만 사전 훈련된 경우에는 더욱 그렇습니다.
+단순한 공백과 구두점 토큰화가 만족스럽지 않다면 단순히 문자를 토큰화하면 어떨까요?
+
+<Youtube id="ssLq_EK2jLE"/>
+
+문자 토큰화는 아주 간단하고 메모리와 시간 복잡도를 크게 줄일 수 있지만, 모델이 의미 있는 입력 표현을 학습하기에는 훨씬 더 어렵습니다.
+
+*예를 들어*, 문자 `"t"`에 대한 의미 있는 문맥 독립적 표현을 배우는 것 보다 단어 `"today"`에 대한 의미 있는 문맥 독립적 표현을 배우는 것이 훨씬 더 어렵습니다.
+문자 토큰화는 종종 성능 저하를 동반하기 때문에 두 가지 장점을 모두 얻기 위해 트랜스포머 모델은 **서브워드** 토큰화라고 하는 단어 수준과 문자 수준 토큰화의 하이브리드를 사용합니다.
+
+## 서브워드 토큰화[[subword-tokenization]]
+
+<Youtube id="zHvTiHr506c"/>
+
+서브워드 토큰화 알고리즘은 자주 사용되는 단어는 더 작은 하위 단어로 쪼개고, 드문 단어는 의미 있는 하위 단어로 분해되어야 한다는 원칙에 따라 작동합니다.
+예를 들어 `"annoyingly"`는 드문 단어로 간주되어 `"annoying"`과 `"ly"`로 분해될 수 있습니다.
+`"annoyingly"`가 `"annoying"`과 `"ly"`의 합성어인 반면, `"annoying"`과 `"ly"` 둘 다 독립적인 서브워드로 자주 등장합니다.
+이는 터키어와 같은 응집성 언어에서 특히 유용하며, 서브워드를 묶어 임의로 긴 복합 단어를 만들 수 있습니다.
+
+서브워드 토큰화를 사용하면 모델이 의미 있는 문맥 독립적 표현을 학습하면서 합리적인 어휘 크기를 가질 수 있습니다.
+또한, 서브워드 토큰화를 통해 모델은 이전에 본 적이 없는 단어를 알려진 서브워드로 분해하여 처리할 수 있습니다.
+
+예를 들어, [`~transformers.BertTokenizer`]는 `"I have a new GPU!"` 라는 문장을 아래와 같이 토큰화합니다:
+
+```py
+>>> from transformers import BertTokenizer
+
+>>> tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+>>> tokenizer.tokenize("I have a new GPU!")
+["i", "have", "a", "new", "gp", "##u", "!"]
+```
+
+대소문자가 없는 모델을 사용해 문장의 시작이 소문자로 표기되었습니다.
+단어 `["i", "have", "a", "new"]`는 토크나이저의 어휘에 속하지만, `"gpu"`는 속하지 않는 것을 확인할 수 있습니다.
+결과적으로 토크나이저는 `"gpu"`를 알려진 두 개의 서브워드로 쪼갭니다: `["gp" and "##u"]`.
+`"##"`은 토큰의 나머지 부분이 공백 없이 이전 토큰에 연결되어야(attach) 함을 의미합니다(토큰화 디코딩 또는 역전을 위해).
+
+또 다른 예로, [`~transformers.XLNetTokenizer`]는 이전에 예시 문장을 다음과 같이 토큰화합니다:
+```py
+>>> from transformers import XLNetTokenizer
+
+>>> tokenizer = XLNetTokenizer.from_pretrained("xlnet-base-cased")
+>>> tokenizer.tokenize("Don't you love 🤗 Transformers? We sure do.")
+["▁Don", "'", "t", "▁you", "▁love", "▁", "🤗", "▁", "Transform", "ers", "?", "▁We", "▁sure", "▁do", "."]
+```
+
+`"▁"`가 가지는 의미는 [SentencePiece](#sentencepiece)에서 다시 살펴보도록 하겠습니다.
+보다시피 `"Transformers"` 라는 드문 단어는 서브워드 `"Transform"`와 `"ers"`로 쪼개집니다.
+
+이제 다양한 하위 단어 토큰화 알고리즘이 어떻게 작동하는지 살펴보겠습니다.
+이러한 토큰화 알고리즘은 일반적으로 해당 모델이 학습되는 말뭉치에 대해 수행되는 어떤 형태의 학습에 의존한다는 점에 유의하세요.
+
+<a id='byte-pair-encoding'></a>
+
+### 바이트 페어 인코딩 (Byte-Pair Encoding, BPE)[[bytepair-encoding-bpe]]
+
+바이트 페어 인코딩(BPE)은 [Neural Machine Translation of Rare Words with Subword Units (Sennrich et
+al., 2015)](https://arxiv.org/abs/1508.07909) 에서 소개되었습니다.
+BPE는 훈련 데이터를 단어로 분할하는 사전 토크나이저(pre-tokenizer)에 의존합니다.
+사전 토큰화(Pretokenization)에는 [GPT-2](model_doc/gpt2), [Roberta](model_doc/roberta)와 같은 간단한 공백 토큰화가 있습니다.
+복잡한 사전 토큰화에는 규칙 기반 토큰화가 해당하는데, 훈련 말뭉치에서 각 단어의 빈도를 계산하기 위해 사용합니다.
+[XLM](model_doc/xlm), 대부분의 언어에서 Moses를 사용하는 [FlauBERT](model_doc/flaubert), Spacy와 ftfy를 사용하는 [GPT](model_doc/gpt)가 해당합니다.
+
+
+사전 토큰화 이후에, 고유 단어 집합가 생성되고 훈련 데이터에서 각 단어가 등장하는 빈도가 결정됩니다.
+다음으로, BPE는 고유 단어 집합에 나타나는 모든 기호로 구성된 기본 어휘를 생성하고 기본 어휘의 두 기호에서 새로운 기호를 형성하는 병합 규칙을 학습합니다.
+어휘가 원하는 어휘 크기에 도달할 때까지 위의 과정을 반복합니다.
+어휘 크기는 토크나이저를 훈련시키기 전에 정의해야 하는 하이퍼파라미터라는 점을 유의하세요.
+
+예를 들어, 사전 토큰화 후 빈도를 포함한 다음과 같은 어휘 집합이 결정되었다고 가정해 보겠습니다:
+
+```
+("hug", 10), ("pug", 5), ("pun", 12), ("bun", 4), ("hugs", 5)
+```
+
+결과적으로 기본 어휘는 `["b", "g", "h", "n", "p", "s", "u"]` 이고, 각 단어를 기본 어휘에 속하는 기호로 쪼개면 아래와 같습니다:
+
+```
+("h" "u" "g", 10), ("p" "u" "g", 5), ("p" "u" "n", 12), ("b" "u" "n", 4), ("h" "u" "g" "s", 5)
+```
+
+그런 다음 BPE는 가능한 각 기호 쌍의 빈도를 계산하여 가장 자주 발생하는 기호 쌍을 선택합니다.
+위의 예시에서 `"h"` 뒤에 오는 `"u"`는 _10 + 5 = 15_ 번 등장합니다. (`"hug"`에서 10번, `"hugs"`에서 5번 등장)
+
+하지만, 가장 등장 빈도가 높은 기호 쌍은 `"u"` 뒤에 오는 `"g"`입니다. _10 + 5 + 5 = 20_ 으로 총 20번 등장합니다.
+따라서 토크나이저가 병합하는 가장 첫 번째 쌍은 `"u"` 뒤에 오는 `"g"`입니다. `"ug"`가 어휘에 추가되어 어휘는 다음과 같습니다:
+
+```
+("h" "ug", 10), ("p" "ug", 5), ("p" "u" "n", 12), ("b" "u" "n", 4), ("h" "ug" "s", 5)
+```
+
+BPE는 다음으로 가장 많이 등장하는 기호 쌍을 식별합니다.
+`"u"` 뒤에 오는 `"n"`은 16번 등장해 `"un"` 으로 병합되어 어휘에 추가됩니다.
+그 다음으로 빈도수가 놓은 기호 쌍은 `"h"` 뒤에 오는 `"ug"`로 15번 등장합니다.
+다시 한 번 `"hug"`로 병합되어 어휘에 추가됩니다.
+
+현재 단계에서 어휘는 `["b", "g", "h", "n", "p", "s", "u", "ug", "un", "hug"]` 이고, 고유 단어 집합은 다음과 같습니다:
+
+```
+("hug", 10), ("p" "ug", 5), ("p" "un", 12), ("b" "un", 4), ("hug" "s", 5)
+```
+
+이 시점에서 바이트 페어 인코딩 훈련이 중단된다고 가정하면, 훈련된 병합 규칙은 새로운 단어에 적용됩니다(기본 어휘에 포함된 기호가 새로운 단어에 포함되지 않는 한).
+예를 들어, 단어 `"bug"`는 `["b", "ug"]`로 토큰화되지만, `"m"`이 기본 어휘에 없기 때문에 `"mug"`는 `["<unk>", "ug"]`로 토큰화될 것입니다.
+훈련 데이터에는 단일 문자가 최소한 한 번 등장하기 때문에 일반적으로 `"m"`과 같은 단일 문자는 `"<unk>"` 기호로 대체되지 않지만, 이모티콘과 같은 특별한 문자인 경우에는 대체될 수 있습니다.
+
+이전에 언급했듯이 어휘 크기(즉 기본 어휘 크기 + 병합 횟수)는 선택해야하는 하이퍼파라미터입니다.
+예를 들어 [GPT](model_doc/gpt)의 기본 어휘 크기는 478, 40,000번의 병합 이후에 훈련을 종료하기 때문에 어휘 크기가 40,478입니다.
+
+#### 바이트 수준 BPE (Byte-level BPE)[[bytelevel-bpe]]
+
+가능한 모든 기본 문자를 포함하는 기본 어휘의 크기는 굉장히 커질 수 있습니다. (예: 모든 유니코드 문자를 기본 문자로 간주하는 경우)
+더 나은 기본 어휘를 갖도록 [GPT-2](https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf)는 기본 어휘로 바이트(bytes)를 사용합니다.
+이 방식은 모든 기본 문자가 어휘에 포함되도록 하면서 기본 어휘의 크기를 256으로 제한합니다.
+구두점을 다루는 추가적인 규칙을 사용해 GPT2 토크나이저는 모든 텍스트를 <unk> 기호 없이 토큰화할 수 있습니다.
+[GPT-2](model_doc/gpt)의 어휘 크기는 50,257로 256 바이트 크기의 기본 토큰, 특별한 end-of-text 토큰과 50,000번의 병합으로 학습한 기호로 구성됩니다.
+
+<a id='wordpiece'></a>
+
+### 워드피스 (WordPiece)[[wordpiece]]
+
+워드피스는 [BERT](model_doc/bert), [DistilBERT](model_doc/distilbert), [Electra](model_doc/electra)에 사용된 서브워드 토큰화 알고리즘입니다.
+이 알고리즘은 [Japanese and Korean Voice Search (Schuster et al., 2012)](https://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/37842.pdf)에서 소개되었고, BPE와 굉장히 유사합니다.
+워드피스는 훈련 데이터에 등장하는 모든 문자로 기본 어휘를 초기화한 후, 주어진 병합 규칙에 따라 점진적으로 학습합니다.
+BPE와는 대조적으로 워드피스는 가장 빈도수가 높은 기호 쌍을 선택하지 않고, 어휘에 추가되었을 때 훈련 데이터의 우도가 최대화되는 쌍을 선택합니다.
+
+정확히 무슨 의미일까요?
+이전 예시를 참조하면, 훈련 데이터의 우도 값을 최대화하는 것은 모든 기호 쌍 중에서 첫 번째 기호와 두 번째 기호의 확률로 나눈 확률이 가장 큰 기호 쌍을 찾는 것과 동일합니다.
+예를 들어 `"ug"`의 확률이 `"u"`와 `"g"` 각각으로 쪼개졌을 때 보다 높아야 `"u"` 뒤에 오는 `"g"`는 병합될 것입니다.
+직관적으로 워드피스는 두 기호를 병합하여 _잃는_ 것을 평가하여 그만한 _가치_가 있는지 확인한다는 점에서 BPE와 약간 다릅니다.
+
+<a id='unigram'></a>
+
+### 유니그램 (Unigram)[[unigram]]
+
+유니그램은 [Subword Regularization: Improving Neural Network Translation Models with Multiple Subword Candidates (Kudo, 2018)](https://arxiv.org/pdf/1804.10959.pdf)에서 제안된 서브워드 토큰화 알고리즘입니다.
+BPE나 워드피스와 달리 유니그램은 기본 어휘를 많은 수의 기호로 초기화한 후 각 기호를 점진적으로 줄여 더 작은 어휘를 얻습니다.
+예를 들어 기본 어휘는 모든 사전 토큰화된 단어와 가장 일반적인 하위 문자열에 해당할 수 있습니다.
+유니그램은 transformers 모델에서 직접적으로 사용되지는 않지만, [SentencePiece](#sentencepiece)와 함께 사용됩니다.
+
+각 훈련 단계에서 유니그램 알고리즘은 현재 어휘와 유니그램 언어 모델이 주어졌을 때 훈련 데이터에 대한 손실(흔히 로그 우도로 정의됨)을 정의합니다.
+그런 다음 어휘의 각 기호에 대해 알고리즘은 해당 기호를 어휘에서 제거할 경우 전체 손실이 얼마나 증가할지 계산합니다.
+이후에 유니그램은 손실 증가율이 가장 낮은 기호의 p(보통 10% 또는 20%) 퍼센트를 제거합니다. (제거되는 기호는 훈련 데이터에 대한 전체 손실에 가장 작은 영향을 미칩니다.)
+어휘가 원하는 크기에 도달할 때까지 이 과정을 반복합니다.
+유니그램 알고리즘은 항상 기본 문자를 포함해 어떤 단어라도 토큰화할 수 있습니다.
+유니그램이 병합 규칙에 기반하지 않기 떄문에 (BPE나 워드피스와는 대조적으로), 해당 알고리즘은 훈련 이후에 새로운 텍스트를 토큰화하는데 여러 가지 방법이 있습니다.
+
+예를 들어, 훈련된 유니그램 토큰화가 다음과 같은 어휘를 가진다면:
+
+```
+["b", "g", "h", "n", "p", "s", "u", "ug", "un", "hug"],
+```
+
+`"hugs"`는 두 가지로 토큰화할 수 있습니다. `["hug", "s"]`와 `["h", "ug", "s"]` 또는 `["h", "u", "g", "s"]`.
+
+그렇다면 어떤 토큰화 방법을 선택해야 할까요?
+유니그램은 어휘를 저장하는 것 외에도 훈련 말뭉치에 각 토큰의 확률을 저장하여 훈련 후 가능한 각 토큰화의 확률을 계산할 수 있도록 합니다.
+이 알고리즘은 단순히 실제로 가장 가능성이 높은 토큰화를 선택하지만, 확률에 따라 가능한 토큰화를 샘플링할 수 있는 가능성도 제공합니다.
+이러한 확률은 토크나이저가 학습한 손실에 의해 정의됩니다.
+
+단어로 구성된 훈련 데이터를 \\(x_{1}, \dots, x_{N}\\)라 하고, 단어 \\(x_{i}\\)에 대한 가능한 모든 토큰화 결과를 \\(S(x_{i})\\)라 한다면, 전체 손실은 다음과 같이 정의됩니다:
+
+$$\mathcal{L} = -\sum_{i=1}^{N} \log \left ( \sum_{x \in S(x_{i})} p(x) \right )$$
+
+
+
+<a id='sentencepiece'></a>
+
+### 센텐스피스 (SentencePiece)[[sentencepiece]]
+
+지금까지 다룬 토큰화 알고리즘은 동일한 문제를 가집니다: 입력 텍스트는 공백을 사용하여 단어를 구분한다고 가정합니다.
+하지만, 모든 언어에서 단어를 구분하기 위해 공백을 사용하지 않습니다.
+한가지 가능한 해결방안은 특정 언어에 특화된 사전 토크나이저를 사용하는 것입니다. 예를 들어 [XLM](model_doc/xlm)은 특정 중국어, 일본어, 태국어 사전 토크나이저를 사용합니다.
+이 문제를 일반적인 방법으로 해결하기 위해, [SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing (Kudo et al., 2018)](https://arxiv.org/pdf/1808.06226.pdf)는 입력을 스트림으로 처리해 공백를 하나의 문자로 사용합니다.
+이후에 BPE 또는 유니그램 알고리즘을 사용해 적절한 어휘를 구성합니다.
+
+[`XLNetTokenizer`]는 센텐스피스를 사용하기 때문에, 위에서 다룬 예시에서 어휘에  `"▁"`가 포함되어있습니다.
+모든 토큰을 합친 후 `"▁"`을 공백으로 대체하면 되기 때문에 센텐스피스로 토큰화된 결과는 디코딩하기 수월합니다.
+
+transformers에서 제공하는 센텐스피스 토크나이저를 사용하는 모든 모델은 유니그램과 함께 사용됩니다. 
+[ALBERT](model_doc/albert), [XLNet](model_doc/xlnet), [Marian](model_doc/marian), [T5](model_doc/t5) 모델이 센텐스피스 토크나이저를 사용합니다.

docs/source/ms/index.md

@@ -76,6 +76,7 @@ Dokumentasi disusun kepada lima bahagian:
 1. **[BLOOM](model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[Bros](model_doc/bros)** (from NAVER) released with the paper [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
 1. **[ByT5](model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
 1. **[CamemBERT](model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
@@ -288,6 +289,7 @@ Flax), PyTorch, dan/atau TensorFlow.
 |            BLIP-2             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             BLOOM             |       ❌       |       ✅       |       ✅        |         ❌         |      ❌      |
 |          BridgeTower          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|             Bros              |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           CamemBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |            CANINE             |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         Chinese-CLIP          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/pt/tasks/sequence_classification.md

@@ -70,7 +70,7 @@ Crie uma função de pré-processamento para tokenizar o campo `text` e truncar
 ...     return tokenizer(examples["text"], truncation=True)
 ```
 
-Use a função [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) do 🤗 Datasets para aplicar a função de pré-processamento em todo o conjunto de dados. Você pode acelerar a função `map` definindo `batched=True` para processar vários elementos do conjunto de dados de uma só vez:
+Use a função [`map`](https://huggingface.co/docs/datasets/process#map) do 🤗 Datasets para aplicar a função de pré-processamento em todo o conjunto de dados. Você pode acelerar a função `map` definindo `batched=True` para processar vários elementos do conjunto de dados de uma só vez:
 
 ```py
 tokenized_imdb = imdb.map(preprocess_function, batched=True)

docs/source/pt/tasks/token_classification.md

@@ -128,7 +128,7 @@ Aqui está como você pode criar uma função para realinhar os tokens e rótulo
 ...     return tokenized_inputs
 ```
 
-Use a função [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) do 🤗 Datasets para tokenizar e alinhar os rótulos em todo o conjunto de dados. Você pode acelerar a função `map` configurando `batched=True` para processar vários elementos do conjunto de dados de uma só vez:
+Use a função [`map`](https://huggingface.co/docs/datasets/process#map) do 🤗 Datasets para tokenizar e alinhar os rótulos em todo o conjunto de dados. Você pode acelerar a função `map` configurando `batched=True` para processar vários elementos do conjunto de dados de uma só vez:
 
 ```py
 >>> tokenized_wnut = wnut.map(tokenize_and_align_labels, batched=True)

docs/source/zh/quicktour.md

@@ -407,7 +407,7 @@ tensor([[0.0021, 0.0018, 0.0115, 0.2121, 0.7725],
 
 ## Trainer - PyTorch优化训练循环
 
-所有的模型都是标准的[`torch.nn.Module`](https://pytorch.org/docs/stable/nn.html#torch.nn.Module), 所以你可以在任何典型的训练模型中使用它们. 当你编写自己的训练循环时W, 🤗 Transformers为PyTorch提供了一个[`Trainer`]类, 它包含了基础的训练循环并且为诸如分布式训练, 混合精度等特性增加了额外的功能.
+所有的模型都是标准的[`torch.nn.Module`](https://pytorch.org/docs/stable/nn.html#torch.nn.Module), 所以你可以在任何典型的训练模型中使用它们. 当你编写自己的训练循环时, 🤗 Transformers为PyTorch提供了一个[`Trainer`]类, 它包含了基础的训练循环并且为诸如分布式训练, 混合精度等特性增加了额外的功能.
 
 取决于你的任务, 你通常可以传递以下的参数给[`Trainer`]:
 

examples/README.md

@@ -19,7 +19,7 @@ We host a wide range of example scripts for multiple learning frameworks. Simply
 
 We also have some [research projects](https://github.com/huggingface/transformers/tree/main/examples/research_projects), as well as some [legacy examples](https://github.com/huggingface/transformers/tree/main/examples/legacy). Note that unlike the main examples these are not actively maintained, and may require specific older versions of dependencies in order to run. 
 
-While we strive to present as many use cases as possible, the example scripts are just that - examples. It is expected that they won't work out-of-the box on your specific problem and that you will be required to change a few lines of code to adapt them to your needs. To help you with that, most of the examples fully expose the preprocessing of the data, allowing you to tweak and edit them as required.
+While we strive to present as many use cases as possible, the example scripts are just that - examples. It is expected that they won't work out-of-the-box on your specific problem and that you will be required to change a few lines of code to adapt them to your needs. To help you with that, most of the examples fully expose the preprocessing of the data, allowing you to tweak and edit them as required.
 
 Please discuss on the [forum](https://discuss.huggingface.co/) or in an [issue](https://github.com/huggingface/transformers/issues) a feature you would like to implement in an example before submitting a PR; we welcome bug fixes, but since we want to keep the examples as simple as possible it's unlikely that we will merge a pull request adding more functionality at the cost of readability.
 

examples/flax/_tests_requirements.txt

@@ -5,4 +5,6 @@ nltk
 rouge-score
 seqeval
 tensorboard
-evaluate >= 0.2.0
+evaluate >= 0.2.0
+torch
+accelerate

examples/flax/language-modeling/run_bart_dlm_flax.py

@@ -139,7 +139,7 @@ class ModelArguments:
         default=None,
         metadata={
             "help": (
-                "The model checkpoint for weights initialization.Don't set if you want to train a model from scratch."
+                "The model checkpoint for weights initialization. Don't set if you want to train a model from scratch."
             )
         },
     )
@@ -531,6 +531,7 @@ def main():
             data_args.dataset_config_name,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            num_proc=data_args.preprocessing_num_workers,
         )
 
         if "validation" not in datasets.keys():
@@ -540,6 +541,7 @@ def main():
                 split=f"train[:{data_args.validation_split_percentage}%]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
             datasets["train"] = load_dataset(
                 data_args.dataset_name,
@@ -547,6 +549,7 @@ def main():
                 split=f"train[{data_args.validation_split_percentage}%:]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
     else:
         data_files = {}
@@ -562,6 +565,7 @@ def main():
             data_files=data_files,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            num_proc=data_args.preprocessing_num_workers,
         )
 
         if "validation" not in datasets.keys():
@@ -571,6 +575,7 @@ def main():
                 split=f"train[:{data_args.validation_split_percentage}%]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
             datasets["train"] = load_dataset(
                 extension,
@@ -578,6 +583,7 @@ def main():
                 split=f"train[{data_args.validation_split_percentage}%:]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
     # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
     # https://huggingface.co/docs/datasets/loading_datasets.html.
@@ -684,7 +690,7 @@ def main():
     # might be slower to preprocess.
     #
     # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
-    # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+    # https://huggingface.co/docs/datasets/process#map
     tokenized_datasets = tokenized_datasets.map(
         group_texts,
         batched=True,

examples/flax/language-modeling/run_clm_flax.py

@@ -140,7 +140,7 @@ class ModelArguments:
         default=None,
         metadata={
             "help": (
-                "The model checkpoint for weights initialization.Don't set if you want to train a model from scratch."
+                "The model checkpoint for weights initialization. Don't set if you want to train a model from scratch."
             )
         },
     )
@@ -421,6 +421,7 @@ def main():
             cache_dir=model_args.cache_dir,
             keep_in_memory=False,
             token=model_args.token,
+            num_proc=data_args.preprocessing_num_workers,
         )
 
         if "validation" not in dataset.keys():
@@ -430,6 +431,7 @@ def main():
                 split=f"train[:{data_args.validation_split_percentage}%]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
             dataset["train"] = load_dataset(
                 data_args.dataset_name,
@@ -437,6 +439,7 @@ def main():
                 split=f"train[{data_args.validation_split_percentage}%:]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
     else:
         data_files = {}
@@ -455,6 +458,7 @@ def main():
             cache_dir=model_args.cache_dir,
             **dataset_args,
             token=model_args.token,
+            num_proc=data_args.preprocessing_num_workers,
         )
 
         if "validation" not in dataset.keys():
@@ -465,6 +469,7 @@ def main():
                 cache_dir=model_args.cache_dir,
                 **dataset_args,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
             dataset["train"] = load_dataset(
                 extension,
@@ -473,6 +478,7 @@ def main():
                 cache_dir=model_args.cache_dir,
                 **dataset_args,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
     # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
     # https://huggingface.co/docs/datasets/loading_datasets.html.
@@ -574,9 +580,9 @@ def main():
         if block_size > config.max_position_embeddings:
             logger.warning(
                 f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
-                "Picking 1024 instead. You can change that default value by passing --block_size xxx."
+                f"Using block_size={min(1024, config.max_position_embeddings)} instead. You can change that default value by passing --block_size xxx."
             )
-            block_size = 1024
+            block_size = min(1024, config.max_position_embeddings)
     else:
         if data_args.block_size > tokenizer.model_max_length:
             logger.warning(
@@ -607,7 +613,7 @@ def main():
     # to preprocess.
     #
     # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
-    # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+    # https://huggingface.co/docs/datasets/process#map
 
     lm_datasets = tokenized_datasets.map(
         group_texts,

examples/flax/language-modeling/run_mlm_flax.py

@@ -145,7 +145,7 @@ class ModelArguments:
         default=None,
         metadata={
             "help": (
-                "The model checkpoint for weights initialization.Don't set if you want to train a model from scratch."
+                "The model checkpoint for weights initialization. Don't set if you want to train a model from scratch."
             )
         },
     )
@@ -458,6 +458,7 @@ def main():
             data_args.dataset_config_name,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            num_proc=data_args.preprocessing_num_workers,
         )
 
         if "validation" not in datasets.keys():
@@ -467,6 +468,7 @@ def main():
                 split=f"train[:{data_args.validation_split_percentage}%]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
             datasets["train"] = load_dataset(
                 data_args.dataset_name,
@@ -474,6 +476,7 @@ def main():
                 split=f"train[{data_args.validation_split_percentage}%:]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
     else:
         data_files = {}
@@ -489,6 +492,7 @@ def main():
             data_files=data_files,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            num_proc=data_args.preprocessing_num_workers,
         )
 
         if "validation" not in datasets.keys():
@@ -498,6 +502,7 @@ def main():
                 split=f"train[:{data_args.validation_split_percentage}%]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
             datasets["train"] = load_dataset(
                 extension,
@@ -505,6 +510,7 @@ def main():
                 split=f"train[{data_args.validation_split_percentage}%:]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
     # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
     # https://huggingface.co/docs/datasets/loading_datasets.html.
@@ -625,7 +631,7 @@ def main():
         # might be slower to preprocess.
         #
         # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
-        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+        # https://huggingface.co/docs/datasets/process#map
         tokenized_datasets = tokenized_datasets.map(
             group_texts,
             batched=True,

examples/flax/language-modeling/run_t5_mlm_flax.py

@@ -139,7 +139,7 @@ class ModelArguments:
         default=None,
         metadata={
             "help": (
-                "The model checkpoint for weights initialization.Don't set if you want to train a model from scratch."
+                "The model checkpoint for weights initialization. Don't set if you want to train a model from scratch."
             )
         },
     )
@@ -572,6 +572,7 @@ def main():
             data_args.dataset_config_name,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            num_proc=data_args.preprocessing_num_workers,
         )
 
         if "validation" not in datasets.keys():
@@ -581,6 +582,7 @@ def main():
                 split=f"train[:{data_args.validation_split_percentage}%]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
             datasets["train"] = load_dataset(
                 data_args.dataset_name,
@@ -588,6 +590,7 @@ def main():
                 split=f"train[{data_args.validation_split_percentage}%:]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
     else:
         data_files = {}
@@ -603,6 +606,7 @@ def main():
             data_files=data_files,
             cache_dir=model_args.cache_dir,
             token=model_args.token,
+            num_proc=data_args.preprocessing_num_workers,
         )
 
         if "validation" not in datasets.keys():
@@ -612,6 +616,7 @@ def main():
                 split=f"train[:{data_args.validation_split_percentage}%]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
             datasets["train"] = load_dataset(
                 extension,
@@ -619,6 +624,7 @@ def main():
                 split=f"train[{data_args.validation_split_percentage}%:]",
                 cache_dir=model_args.cache_dir,
                 token=model_args.token,
+                num_proc=data_args.preprocessing_num_workers,
             )
     # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
     # https://huggingface.co/docs/datasets/loading_datasets.html.
@@ -715,7 +721,7 @@ def main():
     # might be slower to preprocess.
     #
     # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
-    # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+    # https://huggingface.co/docs/datasets/process#map
     tokenized_datasets = tokenized_datasets.map(
         group_texts,
         batched=True,

examples/flax/question-answering/run_qa.py

@@ -62,7 +62,7 @@ from transformers.utils import check_min_version, send_example_telemetry
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.33.0.dev0")
+check_min_version("4.34.0")
 
 Array = Any
 Dataset = datasets.arrow_dataset.Dataset

examples/flax/speech-recognition/README.md

@@ -0,0 +1,68 @@
+<!---
+Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
+# Automatic Speech Recognition - Flax Examples
+
+## Sequence to Sequence
+
+The script [`run_flax_speech_recognition_seq2seq.py`](https://github.com/huggingface/transformers/blob/main/examples/flax/speech-recognition/run_flax_speech_recognition_seq2seq.py) 
+can be used to fine-tune any [Flax Speech Sequence-to-Sequence Model](https://huggingface.co/docs/transformers/main/en/model_doc/auto#transformers.FlaxAutoModelForSpeechSeq2Seq) 
+for automatic speech recognition on one of the [official speech recognition datasets](https://huggingface.co/datasets?task_ids=task_ids:automatic-speech-recognition) 
+or a custom dataset. This includes the Whisper model from OpenAI, or a warm-started Speech-Encoder-Decoder Model, 
+an example for which is included below.
+
+### Whisper Model
+
+We can load all components of the Whisper model directly from the pretrained checkpoint, including the pretrained model 
+weights, feature extractor and tokenizer. We simply have to specify the id of fine-tuning dataset and the necessary
+training hyperparameters.
+
+The following example shows how to fine-tune the [Whisper small](https://huggingface.co/openai/whisper-small) checkpoint 
+on the Hindi subset of the [Common Voice 13](https://huggingface.co/datasets/mozilla-foundation/common_voice_13_0) dataset.
+Note that before running this script you must accept the dataset's [terms of use](https://huggingface.co/datasets/mozilla-foundation/common_voice_13_0) 
+and register your Hugging Face Hub token on your device by running `huggingface-hub login`.
+
+```bash
+python run_flax_speech_recognition_seq2seq.py \
+	--model_name_or_path="openai/whisper-small" \
+	--dataset_name="mozilla-foundation/common_voice_13_0" \
+	--dataset_config_name="hi" \
+	--language="hindi" \
+	--train_split_name="train+validation" \
+	--eval_split_name="test" \
+	--output_dir="./whisper-small-hi-flax" \
+	--per_device_train_batch_size="16" \
+	--per_device_eval_batch_size="16" \
+	--num_train_epochs="10" \
+	--learning_rate="1e-4" \
+	--warmup_steps="500" \
+	--logging_steps="25" \
+	--generation_max_length="40" \
+	--preprocessing_num_workers="32" \
+	--dataloader_num_workers="32" \
+	--max_duration_in_seconds="30" \
+	--text_column_name="sentence" \
+	--overwrite_output_dir \
+	--do_train \
+	--do_eval \
+	--predict_with_generate \
+	--push_to_hub \
+	--use_auth_token
+```
+
+On a TPU v4-8, training should take approximately 25 minutes, with a final cross-entropy loss of 0.02 and word error 
+rate of **34%**. See the checkpoint [sanchit-gandhi/whisper-small-hi-flax](https://huggingface.co/sanchit-gandhi/whisper-small-hi-flax)
+for an example training run.

examples/flax/speech-recognition/requirements.txt

@@ -0,0 +1,8 @@
+datasets[audio]>=2.14.0
+jax>=0.3.6
+jaxlib>=0.3.6
+flax>=0.4.1
+optax>=0.0.8
+torch>=1.9.0
+jiwer
+evaluate

examples/flax/speech-recognition/run_flax_speech_recognition_seq2seq.py

@@ -0,0 +1,857 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the Flax library models for sequence to sequence speech recognition.
+"""
+# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
+
+import logging
+import os
+import sys
+import time
+from dataclasses import field
+from functools import partial
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import datasets
+import evaluate
+import flax
+import jax
+import jax.numpy as jnp
+import numpy as np
+import optax
+from datasets import DatasetDict, load_dataset
+from flax import jax_utils, traverse_util
+from flax.jax_utils import pad_shard_unpad, unreplicate
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from huggingface_hub import Repository, create_repo
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoProcessor,
+    AutoTokenizer,
+    FlaxAutoModelForSpeechSeq2Seq,
+    HfArgumentParser,
+    Seq2SeqTrainingArguments,
+    is_tensorboard_available,
+)
+from transformers.file_utils import get_full_repo_name
+from transformers.utils import check_min_version, send_example_telemetry
+from transformers.utils.versions import require_version
+
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risk.
+check_min_version("4.34.0")
+
+require_version("datasets>=2.14.0", "To fix: pip install -r examples/flax/speech-recogintion/requirements.txt")
+
+logger = logging.getLogger(__name__)
+
+
+@flax.struct.dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    feature_extractor_name: Optional[str] = field(
+        default=None, metadata={"help": "feature extractor name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    model_revision: str = field(
+        default="main",
+        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "Will use the token generated when running `transformers-cli login` (necessary to use this script "
+            "with private models)."
+        },
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": (
+                "Floating-point format in which the model weights should be initialized and trained. Choose one of"
+                " `[float32, float16, bfloat16]`."
+            )
+        },
+    )
+    num_beams: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": (
+                "Number of beams to use for evaluation. This argument will be passed to `model.generate`, "
+                "which is used during evaluation."
+            )
+        },
+    )
+
+
+@flax.struct.dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: str = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    text_column: Optional[str] = field(
+        default=None,
+        metadata={"help": "The name of the column in the datasets containing the full texts (for summarization)."},
+    )
+    dataset_cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Path to cache directory for saving and loading datasets"}
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+            "value if set."
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={"help": "Filter audio files that are longer than `max_duration_in_seconds` seconds"},
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0,
+        metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"},
+    )
+    max_label_length: float = field(
+        default=128,
+        metadata={"help": "Truncate transcriptions that are longer `max_eval_length` tokens."},
+    )
+    pad_input_to_multiple_of: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "If set will pad the input sequence to a multiple of the provided value. "
+            "This is important to avoid triggering recompilations on TPU. If unspecified, will default to padding the inputs to max length."
+        },
+    )
+    pad_target_to_multiple_of: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "If set will pad the target sequence to a multiple of the provided value. "
+            "This is important to avoid triggering recompilations on TPU. If unspecified, will default to padding the targets to max length."
+        },
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only do data preprocessing and skip training. "
+            "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
+            "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
+            "so that the cached datasets can consequently be loaded in distributed training"
+        },
+    )
+    train_split_name: str = field(
+        default="train",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="validation",
+        metadata={
+            "help": "The name of the evaluation data set split to use (via the datasets library). Defaults to 'validation'"
+        },
+    )
+    do_lower_case: bool = field(
+        default=True,
+        metadata={"help": "Whether the target text should be lower cased."},
+    )
+    language: str = field(
+        default=None,
+        metadata={
+            "help": (
+                "Language for multilingual fine-tuning. This argument should be set for multilingual fine-tuning "
+                "only. For English speech recognition, it should be set to `None`."
+            )
+        },
+    )
+    task: str = field(
+        default="transcribe",
+        metadata={"help": "Task, either `transcribe` for speech recognition or `translate` for speech translation."},
+    )
+
+
+def shift_tokens_right(label_ids: np.array, decoder_start_token_id: int) -> np.ndarray:
+    """
+    Shift label ids one token to the right.
+    """
+    shifted_label_ids = np.zeros_like(label_ids)
+    shifted_label_ids[:, 1:] = label_ids[:, :-1]
+    shifted_label_ids[:, 0] = decoder_start_token_id
+
+    return shifted_label_ids
+
+
+@flax.struct.dataclass
+class FlaxDataCollatorSpeechSeq2SeqWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor ([`Wav2Vec2Processor`])
+            The processor used for proccessing the data.
+        decoder_start_token_id (:obj: `int`)
+            The begin-of-sentence of the decoder.
+        input_padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned input sequences (according to the model's padding side and padding index)
+            among:
+            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+              sequence if provided).
+            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
+              maximum acceptable input length for the model if that argument is not provided.
+            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
+              different lengths).
+        target_padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned target sequences (according to the model's padding side and padding index).
+            See above for details.
+        max_input_length (:obj:`float`, `optional`):
+            Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
+        max_target_length (:obj:`int`, `optional`):
+            Maximum length of the ``labels`` of the returned list and optionally padding length (see above).
+        pad_input_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the input sequence to a multiple of the provided value.
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+        pad_target_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the target sequence to a multiple of the provided value.
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+    """
+
+    processor: Any
+    decoder_start_token_id: int
+    input_padding: Union[bool, str] = "longest"
+    target_padding: Union[bool, str] = "max_length"
+    max_input_length: Optional[float] = None
+    max_target_length: Optional[int] = None
+    pad_input_to_multiple_of: Optional[int] = None
+    pad_target_to_multiple_of: Optional[int] = None
+
+    def __call__(self, features: List[Dict[str, Union[List[int], np.ndarray]]]) -> Dict[str, np.ndarray]:
+        # split inputs and labels since they have to be of different lengths and need
+        # different padding methods
+        model_input_name = self.processor.model_input_names[0]
+
+        # dataloader returns a list of features which we convert to a dict
+        input_features = {model_input_name: [feature[model_input_name] for feature in features]}
+        label_features = {"input_ids": [feature["labels"] for feature in features]}
+
+        # reformat list to dict and set to pytorch format
+        batch = self.processor.feature_extractor.pad(
+            input_features,
+            max_length=self.max_input_length,
+            padding=self.input_padding,
+            pad_to_multiple_of=self.pad_input_to_multiple_of,
+            return_tensors="np",
+        )
+
+        labels_batch = self.processor.tokenizer.pad(
+            label_features,
+            max_length=self.max_target_length,
+            padding=self.target_padding,
+            pad_to_multiple_of=self.pad_target_to_multiple_of,
+            return_tensors="np",
+        )
+
+        # if bos token is appended in previous tokenization step,
+        # cut bos token here as it's append later anyways
+        labels = labels_batch["input_ids"]
+        if (labels[:, 0] == self.decoder_start_token_id).all().item():
+            labels = labels[:, 1:]
+            labels_batch.attention_mask = labels_batch.attention_mask[:, 1:]
+
+        decoder_input_ids = shift_tokens_right(labels, self.decoder_start_token_id)
+
+        # replace padding with -100 to ignore correctly when computing the loss
+        labels = np.ma.array(labels, mask=np.not_equal(labels_batch.attention_mask, 1))
+        labels = labels.filled(fill_value=-100)
+
+        batch["labels"] = labels
+        batch["decoder_input_ids"] = decoder_input_ids
+
+        return batch
+
+
+class TrainState(train_state.TrainState):
+    dropout_rng: jnp.ndarray
+
+    def replicate(self):
+        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
+
+
+def write_metric(summary_writer, train_metrics, eval_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+def create_learning_rate_fn(
+    num_train_steps: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, linear_decay learning rate function."""
+    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
+    decay_fn = optax.linear_schedule(
+        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
+    )
+    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
+    return schedule_fn
+
+
+def main():
+    # 1. Parse input arguments
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
+
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
+    # information sent is the one passed as arguments along with your JAX/Flax versions.
+    send_example_telemetry("run_speech_recognition_seq2seq", model_args, data_args, framework="flax")
+
+    # 2. Setup logging
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    # Set the verbosity to info of the Transformers logger.
+    # We only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Check the output dir is valid
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use `--overwrite_output_dir` to overcome."
+        )
+
+    # Handle the repository creation
+    if training_args.push_to_hub:
+        if training_args.hub_model_id is None:
+            repo_name = get_full_repo_name(
+                Path(training_args.output_dir).absolute().name, token=training_args.hub_token
+            )
+        else:
+            repo_name = training_args.hub_model_id
+        create_repo(repo_name, exist_ok=True, token=training_args.hub_token)
+        repo = Repository(training_args.output_dir, clone_from=repo_name, token=training_args.hub_token)
+
+    # 3. Load dataset
+    raw_datasets = DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.train_split_name,
+            cache_dir=data_args.dataset_cache_dir,
+            token=True if model_args.use_auth_token else None,
+        )
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.eval_split_name,
+            cache_dir=data_args.dataset_cache_dir,
+            token=True if model_args.use_auth_token else None,
+        )
+
+    if not training_args.do_train and not training_args.do_eval:
+        raise ValueError(
+            "Cannot not train and not do evaluation. At least one of training or evaluation has to be performed."
+        )
+
+    if data_args.audio_column_name not in next(iter(raw_datasets.values())).column_names:
+        raise ValueError(
+            f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+            "Make sure to set `--audio_column_name` to the correct audio column - one of "
+            f"{', '.join(next(iter(raw_datasets.values())).column_names)}."
+        )
+
+    if data_args.text_column_name not in next(iter(raw_datasets.values())).column_names:
+        raise ValueError(
+            f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
+            "Make sure to set `--text_column_name` to the correct text column - one of "
+            f"{', '.join(next(iter(raw_datasets.values())).column_names)}."
+        )
+
+    # 5. Load pretrained model, tokenizer, and feature extractor
+    config = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        token=True if model_args.use_auth_token else None,
+    )
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.feature_extractor_name if model_args.feature_extractor_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        token=True if model_args.use_auth_token else None,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_fast=model_args.use_fast_tokenizer,
+        revision=model_args.model_revision,
+        token=True if model_args.use_auth_token else None,
+    )
+
+    model = FlaxAutoModelForSpeechSeq2Seq.from_pretrained(
+        model_args.model_name_or_path,
+        config=config,
+        dtype=getattr(jnp, model_args.dtype),
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        token=True if model_args.use_auth_token else None,
+    )
+
+    if model.config.decoder_start_token_id is None:
+        raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
+
+    # 6. Resample speech dataset: `datasets` takes care of automatically loading and resampling the audio,
+    # so we just need to set the correct target sampling rate.
+    raw_datasets = raw_datasets.cast_column(
+        data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
+    )
+
+    # 7. Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    max_input_length = int(data_args.max_duration_in_seconds * feature_extractor.sampling_rate)
+    min_input_length = int(data_args.min_duration_in_seconds * feature_extractor.sampling_rate)
+    max_label_length = (
+        data_args.max_label_length if data_args.max_label_length is not None else model.config.max_length
+    )
+    pad_input_to_multiple_of = data_args.pad_input_to_multiple_of
+    pad_target_to_multiple_of = data_args.pad_target_to_multiple_of
+    audio_column_name = data_args.audio_column_name
+    num_workers = data_args.preprocessing_num_workers
+    text_column_name = data_args.text_column_name
+    model_input_name = feature_extractor.model_input_names[0]
+    do_lower_case = data_args.do_lower_case
+
+    if training_args.do_train and data_args.max_train_samples is not None:
+        raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval and data_args.max_eval_samples is not None:
+        raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
+
+    if data_args.language is not None:
+        # We only need to set the task id when the language is specified (i.e. in a multilingual setting)
+        tokenizer.set_prefix_tokens(language=data_args.language, task=data_args.task)
+
+    def prepare_dataset(batch):
+        # process audio
+        sample = batch[audio_column_name]
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        # process audio length
+        batch[model_input_name] = inputs.get(model_input_name)[0]
+        batch["input_length"] = len(sample["array"])
+
+        # process targets
+        input_str = batch[text_column_name].lower() if do_lower_case else batch[text_column_name]
+        batch["labels"] = tokenizer(input_str).input_ids
+        return batch
+
+    vectorized_datasets = raw_datasets.map(
+        prepare_dataset,
+        remove_columns=next(iter(raw_datasets.values())).column_names,
+        num_proc=num_workers,
+        desc="preprocess train dataset",
+    )
+
+    # filter training data with inputs longer than max_input_length
+    def is_audio_in_length_range(length):
+        return min_input_length < length < max_input_length
+
+    vectorized_datasets = vectorized_datasets.filter(
+        is_audio_in_length_range,
+        num_proc=num_workers,
+        input_columns=["input_length"],
+    )
+
+    # for large datasets it is advised to run the preprocessing on a
+    # single machine first with `args.preprocessing_only` since there will mostly likely
+    # be a timeout when running the script in distributed mode.
+    # In a second step `args.preprocessing_only` can then be set to `False` to load the
+    # cached dataset
+    if data_args.preprocessing_only:
+        cache = {k: v.cache_files for k, v in vectorized_datasets.items()}
+        logger.info(f"Data preprocessing finished. Files cached at {cache}.")
+        return
+
+    # 8. Load Metric
+    metric = evaluate.load("wer")
+
+    def compute_metrics(preds, labels):
+        # replace padded labels by the padding token
+        for idx in range(len(labels)):
+            labels[idx][labels[idx] == -100] = tokenizer.pad_token_id
+
+        pred_str = tokenizer.batch_decode(preds, skip_special_tokens=True)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(labels, skip_special_tokens=True)
+
+        wer = metric.compute(predictions=pred_str, references=label_str)
+        return {"wer": wer}
+
+    # 9. Save feature extractor, tokenizer and config
+    feature_extractor.save_pretrained(training_args.output_dir)
+    tokenizer.save_pretrained(training_args.output_dir)
+    config.save_pretrained(training_args.output_dir)
+
+    processor = AutoProcessor.from_pretrained(training_args.output_dir)
+
+    data_collator = FlaxDataCollatorSpeechSeq2SeqWithPadding(
+        processor=processor,
+        decoder_start_token_id=model.config.decoder_start_token_id,
+        input_padding="longest",
+        target_padding="longest",
+        max_target_length=max_label_length,
+        pad_input_to_multiple_of=pad_input_to_multiple_of,
+        pad_target_to_multiple_of=pad_target_to_multiple_of if pad_target_to_multiple_of else max_label_length,
+    )
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+
+            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+    else:
+        logger.warning(
+            "Unable to display metrics through TensorBoard because the package is not installed: "
+            "Please run pip install tensorboard to enable."
+        )
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    rng, dropout_rng = jax.random.split(rng)
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    per_device_eval_batch_size = int(training_args.per_device_eval_batch_size)
+    eval_batch_size = per_device_eval_batch_size * jax.device_count()
+    steps_per_epoch = len(vectorized_datasets["train"]) // train_batch_size
+    total_train_steps = steps_per_epoch * num_epochs
+
+    # Create learning rate schedule
+    linear_decay_lr_schedule_fn = create_learning_rate_fn(
+        len(vectorized_datasets["train"]),
+        training_args.warmup_steps,
+        training_args.learning_rate,
+    )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        # find out all LayerNorm parameters
+        layer_norm_candidates = ["layer_norm", "self_attn_layer_norm", "final_layer_norm", "encoder_attn_layer_norm"]
+        layer_norm_named_params = {
+            layer[-2:]
+            for layer_norm_name in layer_norm_candidates
+            for layer in flat_params.keys()
+            if layer_norm_name in "".join(layer).lower()
+        }
+        flat_mask = {path: (path[-1] != "bias" and path[-2:] not in layer_norm_named_params) for path in flat_params}
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    adamw = optax.adamw(
+        learning_rate=linear_decay_lr_schedule_fn,
+        b1=training_args.adam_beta1,
+        b2=training_args.adam_beta2,
+        eps=training_args.adam_epsilon,
+        weight_decay=training_args.weight_decay,
+        mask=decay_mask_fn,
+    )
+
+    # Setup train state
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw, dropout_rng=dropout_rng)
+
+    # label smoothed cross entropy
+    def loss_fn(logits, labels, label_smoothing_factor=0.0):
+        """
+        The label smoothing implementation is adapted from Flax's official example:
+        https://github.com/google/flax/blob/87a211135c6a377c8f29048a1cac3840e38b9da4/examples/wmt/train.py#L104
+        """
+        vocab_size = logits.shape[-1]
+        confidence = 1.0 - label_smoothing_factor
+        low_confidence = (1.0 - confidence) / (vocab_size - 1)
+        normalizing_constant = -(
+            confidence * jnp.log(confidence) + (vocab_size - 1) * low_confidence * jnp.log(low_confidence + 1e-20)
+        )
+        soft_labels = onehot(labels, vocab_size, on_value=confidence, off_value=low_confidence)
+
+        loss = optax.softmax_cross_entropy(logits, soft_labels)
+        loss = loss - normalizing_constant
+
+        # ignore padded tokens from loss, i.e. where labels are not set to -100
+        padding_mask = labels >= 0
+        loss = loss * padding_mask
+        loss = loss.sum()
+        num_labels = padding_mask.sum()
+        return loss, num_labels
+
+    # Define gradient update step fn
+    def train_step(state, batch, label_smoothing_factor=0.0):
+        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+
+        def compute_loss(params):
+            labels = batch.pop("labels")
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+            loss, num_labels = loss_fn(logits, labels, label_smoothing_factor)
+            return loss, num_labels
+
+        grad_fn = jax.value_and_grad(compute_loss, has_aux=True)
+        (loss, num_labels), grad = grad_fn(state.params)
+        num_labels = jax.lax.psum(num_labels, "batch")
+
+        # true loss = total loss / total samples
+        loss = jax.lax.psum(loss, "batch")
+        loss = jax.tree_util.tree_map(lambda x: x / num_labels, loss)
+
+        # true grad = total grad / total samples
+        grad = jax.lax.psum(grad, "batch")
+        grad = jax.tree_util.tree_map(lambda x: x / num_labels, grad)
+        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
+
+        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
+        return new_state, metrics
+
+    # Define eval fn
+    def eval_step(params, batch, label_smoothing_factor=0.0):
+        labels = batch.pop("labels")
+        logits = model(**batch, params=params, train=False)[0]
+
+        loss, num_labels = loss_fn(logits, labels, label_smoothing_factor)
+        num_labels = jax.lax.psum(num_labels, "batch")
+
+        # true loss = total loss / total samples
+        loss = jax.lax.psum(loss, "batch")
+        loss = jax.tree_util.tree_map(lambda x: x / num_labels, loss)
+
+        metrics = {"loss": loss}
+        return metrics
+
+    # Define generation function
+    num_beams = model_args.num_beams if model_args.num_beams is not None else model.config.num_beams
+    gen_kwargs = {"max_length": max_label_length, "num_beams": num_beams}
+
+    def generate_step(params, batch):
+        model.params = params
+        output_ids = model.generate(batch[model_input_name], attention_mask=batch.get("attention_mask"), **gen_kwargs)
+        return output_ids.sequences
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(
+        partial(train_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch", donate_argnums=(0,)
+    )
+    p_eval_step = jax.pmap(partial(eval_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch")
+    p_generate_step = jax.pmap(generate_step, "batch")
+
+    # Replicate the train state on each device
+    state = state.replicate()
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(vectorized_datasets['train'])}")
+    logger.info(f"  Num Epochs = {num_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    train_time = 0
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+
+        train_metrics = []
+
+        # Generate an epoch by shuffling sampling indices from the train dataset and create a data loader
+        vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed)
+        train_loader = DataLoader(
+            vectorized_datasets["train"],
+            batch_size=train_batch_size,
+            drop_last=True,
+            collate_fn=data_collator,
+            num_workers=training_args.dataloader_num_workers,
+        )
+        # train
+        for batch in tqdm(train_loader, desc="Training...", position=1, leave=False):
+            batch = shard(batch.data)
+            state, train_metric = p_train_step(state, batch)
+            train_metrics.append(train_metric)
+
+        train_time += time.time() - train_start
+
+        train_metric = unreplicate(train_metric)
+
+        epochs.write(
+            f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate:"
+            f" {train_metric['learning_rate']})"
+        )
+
+        # ======================== Evaluating ==============================
+        eval_metrics = []
+        eval_preds = []
+        eval_labels = []
+
+        eval_loader = DataLoader(
+            vectorized_datasets["eval"],
+            batch_size=eval_batch_size,
+            drop_last=False,
+            collate_fn=data_collator,
+            num_workers=training_args.dataloader_num_workers,
+        )
+        for batch in tqdm(eval_loader, desc="Evaluating...", position=2, leave=False):
+            # Model forward
+            labels = batch["labels"]
+
+            metrics = pad_shard_unpad(p_eval_step, static_return=True)(
+                state.params, batch.data, min_device_batch=per_device_eval_batch_size
+            )
+            eval_metrics.append(metrics)
+
+            # generation
+            if training_args.predict_with_generate:
+                generated_ids = pad_shard_unpad(p_generate_step)(state.params, batch.data)
+                eval_preds.extend(jax.device_get(generated_ids.reshape(-1, gen_kwargs["max_length"])))
+                eval_labels.extend(labels)
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_util.tree_map(jnp.mean, eval_metrics)
+
+        # compute WER metric
+        wer_desc = ""
+        if training_args.predict_with_generate:
+            wer_metric = compute_metrics(eval_preds, eval_labels)
+            eval_metrics.update(wer_metric)
+            wer_desc = " ".join([f"Eval {key}: {value} |" for key, value in wer_metric.items()])
+
+        # Print metrics and update progress bar
+        desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']} | {wer_desc})"
+        epochs.write(desc)
+        epochs.desc = desc
+
+        # Save metrics
+        if has_tensorboard and jax.process_index() == 0:
+            cur_step = epoch * (len(vectorized_datasets["train"]) // train_batch_size)
+            write_metric(summary_writer, train_metrics, eval_metrics, train_time, cur_step)
+
+        # save checkpoint after each epoch and push checkpoint to the hub
+        if jax.process_index() == 0:
+            params = jax.device_get(jax.tree_util.tree_map(lambda x: x[0], state.params))
+            model.save_pretrained(training_args.output_dir, params=params)
+            tokenizer.save_pretrained(training_args.output_dir)
+            if training_args.push_to_hub:
+                repo.push_to_hub(commit_message=f"Saving weights and logs of epoch {epoch}", blocking=False)
+
+
+if __name__ == "__main__":
+    main()

examples/flax/summarization/run_summarization_flax.py

@@ -159,7 +159,7 @@ class ModelArguments:
         default=None,
         metadata={
             "help": (
-                "The model checkpoint for weights initialization.Don't set if you want to train a model from scratch."
+                "The model checkpoint for weights initialization. Don't set if you want to train a model from scratch."
             )
         },
     )

examples/flax/test_flax_examples.py

@@ -32,6 +32,7 @@ SRC_DIRS = [
         "summarization",
         "token-classification",
         "question-answering",
+        "speech-recognition",
     ]
 ]
 sys.path.extend(SRC_DIRS)
@@ -41,6 +42,7 @@ if SRC_DIRS is not None:
     import run_clm_flax
     import run_flax_glue
     import run_flax_ner
+    import run_flax_speech_recognition_seq2seq
     import run_mlm_flax
     import run_qa
     import run_summarization_flax
@@ -252,3 +254,32 @@ class ExamplesTests(TestCasePlus):
             result = get_results(tmp_dir)
             self.assertGreaterEqual(result["eval_f1"], 30)
             self.assertGreaterEqual(result["eval_exact"], 30)
+
+    @slow
+    def test_run_flax_speech_recognition_seq2seq(self):
+        tmp_dir = self.get_auto_remove_tmp_dir()
+        testargs = f"""
+            run_flax_speech_recognition_seq2seq.py
+            --model_name_or_path openai/whisper-tiny.en
+            --dataset_name hf-internal-testing/librispeech_asr_dummy
+            --dataset_config clean
+            --train_split_name validation
+            --eval_split_name validation
+            --output_dir {tmp_dir}
+            --overwrite_output_dir
+            --num_train_epochs=2
+            --max_train_samples 10
+            --max_eval_samples 10
+            --warmup_steps=8
+            --do_train
+            --do_eval
+            --learning_rate=2e-4
+            --per_device_train_batch_size=2
+            --per_device_eval_batch_size=1
+            --predict_with_generate
+        """.split()
+
+        with patch.object(sys, "argv", testargs):
+            run_flax_speech_recognition_seq2seq.main()
+            result = get_results(tmp_dir, split="eval")
+            self.assertLessEqual(result["eval_wer"], 0.05)

examples/flax/text-classification/run_flax_glue.py

@@ -55,7 +55,7 @@ from transformers.utils import check_min_version, send_example_telemetry
 
 logger = logging.getLogger(__name__)
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.33.0.dev0")
+check_min_version("4.34.0")
 
 Array = Any
 Dataset = datasets.arrow_dataset.Dataset

examples/flax/token-classification/run_flax_ner.py

@@ -56,7 +56,7 @@ from transformers.utils.versions import require_version
 
 logger = logging.getLogger(__name__)
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.33.0.dev0")
+check_min_version("4.34.0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/token-classification/requirements.txt")