Patch release: v4.21.2

Accept trust_remote_code and ignore it in PreTrainedModel.from_pretrained (#18428 )
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
2022-08-24 14:53:27 +01:00 · 2022-08-24 14:40:44 +01:00 · 2022-08-04 09:26:24 -04:00 · 2022-08-04 09:24:45 -04:00 · 2022-07-27 15:08:28 +02:00 · 2022-07-27 15:08:28 +02:00
569 changed files with 43555 additions and 5317 deletions
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@@ -65,19 +65,20 @@ jobs:
    run_tests_torch_and_tf:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            RUN_PT_TF_CROSS_TESTS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch_and_tf-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch_and_tf-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng git-lfs
            - run: git lfs install
            - run: pip install --upgrade pip
@@ -87,7 +88,7 @@ jobs:
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - run: pip install git+https://github.com/huggingface/accelerate
            - save_cache:
-                key: v0.4-{{ checksum "setup.py" }}
+                key: v0.5-{{ checksum "setup.py" }}
                paths:
                    - '~/.cache/pip'
            - run: python utils/tests_fetcher.py | tee test_preparation.txt
@@ -105,19 +106,20 @@ jobs:
    run_tests_torch_and_tf_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            RUN_PT_TF_CROSS_TESTS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch_and_tf-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch_and_tf-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng git-lfs
            - run: git lfs install
            - run: pip install --upgrade pip
@@ -127,7 +129,7 @@ jobs:
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - run: pip install git+https://github.com/huggingface/accelerate
            - save_cache:
-                key: v0.4-{{ checksum "setup.py" }}
+                key: v0.5-{{ checksum "setup.py" }}
                paths:
                    - '~/.cache/pip'
            - run: |
@@ -140,19 +142,20 @@ jobs:
    run_tests_torch_and_flax:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            RUN_PT_FLAX_CROSS_TESTS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch_and_flax-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch_and_flax-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,flax,torch,testing,sentencepiece,torch-speech,vision]
@@ -160,7 +163,7 @@ jobs:
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - run: pip install git+https://github.com/huggingface/accelerate
            - save_cache:
-                key: v0.4-{{ checksum "setup.py" }}
+                key: v0.5-{{ checksum "setup.py" }}
                paths:
                    - '~/.cache/pip'
            - run: python utils/tests_fetcher.py | tee test_preparation.txt
@@ -178,19 +181,20 @@ jobs:
    run_tests_torch_and_flax_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            RUN_PT_FLAX_CROSS_TESTS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch_and_flax-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch_and_flax-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,flax,torch,testing,sentencepiece,torch-speech,vision]
@@ -198,7 +202,7 @@ jobs:
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - run: pip install git+https://github.com/huggingface/accelerate
            - save_cache:
-                key: v0.4-{{ checksum "setup.py" }}
+                key: v0.5-{{ checksum "setup.py" }}
                paths:
                    - '~/.cache/pip'
            - run: |
@@ -211,18 +215,19 @@ jobs:
    run_tests_torch:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng time
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,torch,testing,sentencepiece,torch-speech,vision,timm]
@@ -230,7 +235,7 @@ jobs:
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - run: pip install git+https://github.com/huggingface/accelerate
            - save_cache:
-                  key: v0.4-torch-{{ checksum "setup.py" }}
+                  key: v0.5-torch-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/tests_fetcher.py | tee test_preparation.txt
@@ -248,18 +253,19 @@ jobs:
    run_tests_torch_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,torch,testing,sentencepiece,torch-speech,vision,timm]
@@ -267,7 +273,7 @@ jobs:
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - run: pip install git+https://github.com/huggingface/accelerate
            - save_cache:
-                  key: v0.4-torch-{{ checksum "setup.py" }}
+                  key: v0.5-torch-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: |
@@ -280,25 +286,26 @@ jobs:
    run_tests_tf:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-tf-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-tf-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,tf-cpu,testing,sentencepiece,tf-speech,vision]
            - run: pip install tensorflow_probability
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - save_cache:
-                  key: v0.4-tf-{{ checksum "setup.py" }}
+                  key: v0.5-tf-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/tests_fetcher.py | tee test_preparation.txt
@@ -316,25 +323,26 @@ jobs:
    run_tests_tf_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-tf-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-tf-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,tf-cpu,testing,sentencepiece,tf-speech,vision]
            - run: pip install tensorflow_probability
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - save_cache:
-                  key: v0.4-tf-{{ checksum "setup.py" }}
+                  key: v0.5-tf-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: |
@@ -347,24 +355,25 @@ jobs:
    run_tests_flax:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                keys:
-                    - v0.4-flax-{{ checksum "setup.py" }}
-                    - v0.4-{{ checksum "setup.py" }}
+                    - v0.5-flax-{{ checksum "setup.py" }}
+                    - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[flax,testing,sentencepiece,flax-speech,vision]
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - save_cache:
-                  key: v0.4-flax-{{ checksum "setup.py" }}
+                  key: v0.5-flax-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/tests_fetcher.py | tee test_preparation.txt
@@ -382,24 +391,25 @@ jobs:
    run_tests_flax_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                keys:
-                    - v0.4-flax-{{ checksum "setup.py" }}
-                    - v0.4-{{ checksum "setup.py" }}
+                    - v0.5-flax-{{ checksum "setup.py" }}
+                    - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[flax,testing,sentencepiece,vision,flax-speech]
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - save_cache:
-                  key: v0.4-flax-{{ checksum "setup.py" }}
+                  key: v0.5-flax-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: |
@@ -412,26 +422,27 @@ jobs:
    run_tests_pipelines_torch:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            RUN_PIPELINE_TESTS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,torch,testing,sentencepiece,torch-speech,vision,timm]
            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - save_cache:
-                  key: v0.4-torch-{{ checksum "setup.py" }}
+                  key: v0.5-torch-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/tests_fetcher.py | tee test_preparation.txt
@@ -449,26 +460,27 @@ jobs:
    run_tests_pipelines_torch_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            RUN_PIPELINE_TESTS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,torch,testing,sentencepiece,torch-speech,vision,timm]
            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.11.0+cpu.html
            - run: pip install https://github.com/kpu/kenlm/archive/master.zip
            - save_cache:
-                  key: v0.4-torch-{{ checksum "setup.py" }}
+                  key: v0.5-torch-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: |
@@ -481,24 +493,25 @@ jobs:
    run_tests_pipelines_tf:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            RUN_PIPELINE_TESTS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-tf-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-tf-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,tf-cpu,testing,sentencepiece]
            - run: pip install tensorflow_probability
            - save_cache:
-                  key: v0.4-tf-{{ checksum "setup.py" }}
+                  key: v0.5-tf-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/tests_fetcher.py | tee test_preparation.txt
@@ -516,24 +529,25 @@ jobs:
    run_tests_pipelines_tf_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            RUN_PIPELINE_TESTS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-tf-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-tf-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,tf-cpu,testing,sentencepiece]
            - run: pip install tensorflow_probability
            - save_cache:
-                  key: v0.4-tf-{{ checksum "setup.py" }}
+                  key: v0.5-tf-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: |
@@ -546,21 +560,22 @@ jobs:
    run_tests_custom_tokenizers:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            RUN_CUSTOM_TOKENIZERS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-custom_tokenizers-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-custom_tokenizers-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[ja,testing,sentencepiece,jieba,spacy,ftfy,rjieba]
            - run: python -m unidic download
            - save_cache:
-                  key: v0.4-custom_tokenizers-{{ checksum "setup.py" }}
+                  key: v0.5-custom_tokenizers-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: |
@@ -579,24 +594,25 @@ jobs:
    run_examples_torch:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch_examples-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch_examples-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,torch,sentencepiece,testing,torch-speech]
            - run: pip install -r examples/pytorch/_tests_requirements.txt
            - save_cache:
-                  key: v0.4-torch_examples-{{ checksum "setup.py" }}
+                  key: v0.5-torch_examples-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/tests_fetcher.py --filters examples tests | tee test_preparation.txt
@@ -614,24 +630,25 @@ jobs:
    run_examples_torch_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch_examples-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch_examples-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,torch,sentencepiece,testing,torch-speech]
            - run: pip install -r examples/pytorch/_tests_requirements.txt
            - save_cache:
-                  key: v0.4-torch_examples-{{ checksum "setup.py" }}
+                  key: v0.5-torch_examples-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: |
@@ -644,23 +661,24 @@ jobs:
    run_examples_flax:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                keys:
-                    - v0.4-flax_examples-{{ checksum "setup.py" }}
-                    - v0.4-{{ checksum "setup.py" }}
+                    - v0.5-flax_examples-{{ checksum "setup.py" }}
+                    - v0.5-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
-            - run: sudo pip install .[flax,testing,sentencepiece]
+            - run: pip install .[flax,testing,sentencepiece]
            - run: pip install -r examples/flax/_tests_requirements.txt
            - save_cache:
-                  key: v0.4-flax_examples-{{ checksum "setup.py" }}
+                  key: v0.5-flax_examples-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/tests_fetcher.py --filters examples tests | tee test_preparation.txt
@@ -678,23 +696,24 @@ jobs:
    run_examples_flax_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                keys:
-                    - v0.4-flax_examples-{{ checksum "setup.py" }}
-                    - v0.4-{{ checksum "setup.py" }}
+                    - v0.5-flax_examples-{{ checksum "setup.py" }}
+                    - v0.5-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
-            - run: sudo pip install .[flax,testing,sentencepiece]
+            - run: pip install .[flax,testing,sentencepiece]
            - run: pip install -r examples/flax/_tests_requirements.txt
            - save_cache:
-                  key: v0.4-flax_examples-{{ checksum "setup.py" }}
+                  key: v0.5-flax_examples-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: |
@@ -707,27 +726,28 @@ jobs:
    run_tests_hub:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            HUGGINGFACE_CO_STAGING: yes
            RUN_GIT_LFS_TESTS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-hub-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: sudo apt-get install git-lfs
+                      - v0.5-hub-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
+            - run: sudo apt-get -y update && sudo apt-get install git-lfs
            - run: |
                git config --global user.email "ci@dummy.com"
                git config --global user.name "ci"
            - run: pip install --upgrade pip
            - run: pip install .[torch,sentencepiece,testing]
            - save_cache:
-                  key: v0.4-hub-{{ checksum "setup.py" }}
+                  key: v0.5-hub-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/tests_fetcher.py | tee test_preparation.txt
@@ -745,27 +765,28 @@ jobs:
    run_tests_hub_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            HUGGINGFACE_CO_STAGING: yes
            RUN_GIT_LFS_TESTS: yes
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-hub-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: sudo apt-get install git-lfs
+                      - v0.5-hub-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
+            - run: sudo apt-get -y update && sudo apt-get install git-lfs
            - run: |
                git config --global user.email "ci@dummy.com"
                git config --global user.name "ci"
            - run: pip install --upgrade pip
            - run: pip install .[torch,sentencepiece,testing]
            - save_cache:
-                  key: v0.4-hub-{{ checksum "setup.py" }}
+                  key: v0.5-hub-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: |
@@ -778,22 +799,23 @@ jobs:
    run_tests_onnxruntime:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[torch,testing,sentencepiece,onnxruntime,vision,rjieba]
            - save_cache:
-                  key: v0.4-onnx-{{ checksum "setup.py" }}
+                  key: v0.5-onnx-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/tests_fetcher.py | tee test_preparation.txt
@@ -811,22 +833,23 @@ jobs:
    run_tests_onnxruntime_all:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[torch,testing,sentencepiece,onnxruntime,vision]
            - save_cache:
-                  key: v0.4-onnx-{{ checksum "setup.py" }}
+                  key: v0.5-onnx-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: |
@@ -839,21 +862,22 @@ jobs:
    check_code_quality:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        resource_class: large
        environment:
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-code_quality-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-code_quality-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[all,quality]
            - save_cache:
-                  key: v0.4-code_quality-{{ checksum "setup.py" }}
+                  key: v0.5-code_quality-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: black --check --preview examples tests src utils
@@ -862,25 +886,27 @@ jobs:
            - run: python utils/sort_auto_mappings.py --check_only
            - run: flake8 examples tests src utils
            - run: doc-builder style src/transformers docs/source --max_len 119 --check_only --path_to_docs docs/source
+            - run: python utils/check_doc_toc.py

    check_repository_consistency:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        resource_class: large
        environment:
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-repository_consistency-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-repository_consistency-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[all,quality]
            - save_cache:
-                  key: v0.4-repository_consistency-{{ checksum "setup.py" }}
+                  key: v0.5-repository_consistency-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/check_copies.py
@@ -895,18 +921,19 @@ jobs:
    run_tests_layoutlmv2_and_v3:
        working_directory: ~/transformers
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
+            PYTEST_TIMEOUT: 120
        resource_class: xlarge
        parallelism: 1
        steps:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.4-torch-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
+                      - v0.5-torch-{{ checksum "setup.py" }}
+                      - v0.5-{{ checksum "setup.py" }}
            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
            - run: pip install --upgrade pip
            - run: pip install .[torch,testing,vision]
@@ -915,7 +942,7 @@ jobs:
            - run: sudo apt install tesseract-ocr
            - run: pip install pytesseract
            - save_cache:
-                  key: v0.4-torch-{{ checksum "setup.py" }}
+                  key: v0.5-torch-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: python utils/tests_fetcher.py | tee test_preparation.txt
@@ -933,7 +960,7 @@ jobs:
 # TPU JOBS
    run_examples_tpu:
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        environment:
            OMP_NUM_THREADS: 1
            TRANSFORMERS_IS_CI: yes
@@ -953,7 +980,7 @@ jobs:

    cleanup-gke-jobs:
        docker:
-            - image: circleci/python:3.7
+            - image: cimg/python:3.7.12
        steps:
            - gcp-gke/install
            - gcp-gke/update-kubeconfig-with-credentials:
--- a/.github/ISSUE_TEMPLATE/bug-report.yml
+++ b/.github/ISSUE_TEMPLATE/bug-report.yml
@@ -7,7 +7,6 @@ body:
    attributes:
      label: System Info
      description: Please share your system info with us. You can run the command `transformers-cli env` and copy-paste its output below.
-      render: shell
      placeholder: transformers version, platform, python version, ...
    validations:
      required: true
@@ -118,4 +117,3 @@ body:
    attributes:
      label: Expected behavior
      description: "A clear and concise description of what you would expect to happen."
-      render: shell
--- a/.github/conda/meta.yaml
+++ b/.github/conda/meta.yaml
@@ -25,7 +25,7 @@ requirements:
    - sacremoses
    - regex !=2019.12.17
    - protobuf
-    - tokenizers >=0.10.1,<0.11.0
+    - tokenizers >=0.11.1,!=0.11.3,<0.13
    - pyyaml >=5.1
  run:
    - python
@@ -40,7 +40,7 @@ requirements:
    - sacremoses
    - regex !=2019.12.17
    - protobuf
-    - tokenizers >=0.10.1,<0.11.0
+    - tokenizers >=0.11.1,!=0.11.3,<0.13
    - pyyaml >=5.1

 test:
--- a/.github/workflows/add-model-like.yml
+++ b/.github/workflows/add-model-like.yml
@@ -27,7 +27,7 @@ jobs:
        id: cache
        with:
          path: ~/venv/
-          key: v3-tests_model_like-${{ hashFiles('setup.py') }}
+          key: v4-tests_model_like-${{ hashFiles('setup.py') }}

      - name: Create virtual environment on cache miss
        if: steps.cache.outputs.cache-hit != 'true'
--- a/.github/workflows/build-docker-images.yml
+++ b/.github/workflows/build-docker-images.yml
@@ -5,6 +5,7 @@ on:
    branches:
      - docker-image*
  repository_dispatch:
+  workflow_call:
  schedule:
    - cron: "0 1 * * *"

@@ -39,6 +40,33 @@ jobs:
          push: true
          tags: huggingface/transformers-all-latest-gpu

+  latest-with-torch-nightly-docker:
+    name: "Nightly PyTorch + Stable TensorFlow"
+    runs-on: ubuntu-latest
+    steps:
+      -
+        name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      -
+        name: Check out code
+        uses: actions/checkout@v2
+      -
+        name: Login to DockerHub
+        uses: docker/login-action@v1
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+      -
+        name: Build and push
+        uses: docker/build-push-action@v2
+        with:
+          context: ./docker/transformers-all-latest-gpu
+          build-args: |
+            REF=main
+            PYTORCH=pre
+          push: true
+          tags: huggingface/transformers-all-latest-torch-nightly-gpu
+
  latest-torch-deepspeed-docker:
    name: "Latest PyTorch + DeepSpeed"
    runs-on: ubuntu-latest
@@ -65,6 +93,32 @@ jobs:
          push: true
          tags: huggingface/transformers-pytorch-deepspeed-latest-gpu

+  nightly-torch-deepspeed-docker:
+    name: "Nightly PyTorch + DeepSpeed"
+    runs-on: ubuntu-latest
+    steps:
+      -
+        name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      -
+        name: Check out code
+        uses: actions/checkout@v2
+      -
+        name: Login to DockerHub
+        uses: docker/login-action@v1
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+      -
+        name: Build and push
+        uses: docker/build-push-action@v2
+        with:
+          context: ./docker/transformers-pytorch-deepspeed-nightly-gpu
+          build-args: |
+            REF=main
+          push: true
+          tags: huggingface/transformers-pytorch-deepspeed-nightly-gpu
+
  doc-builder:
    name: "Doc builder"
    runs-on: ubuntu-latest
--- a/.github/workflows/build-past-ci-docker-images.yml
+++ b/.github/workflows/build-past-ci-docker-images.yml
@@ -0,0 +1,108 @@
+name: Build docker images (Past CI)
+
+on:
+  push:
+    branches:
+      - past-ci-docker-image*
+
+concurrency:
+  group: docker-images-builds
+  cancel-in-progress: false
+
+jobs:
+  past-pytorch-docker:
+    name: "Past PyTorch Docker"
+    strategy:
+      fail-fast: false
+      matrix:
+        version: ["1.11", "1.10", "1.9", "1.8", "1.7", "1.6", "1.5", "1.4"]
+    runs-on: ubuntu-latest
+    steps:
+      -
+        name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      -
+        name: Check out code
+        uses: actions/checkout@v2
+      -
+        name: Login to DockerHub
+        uses: docker/login-action@v1
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+      -
+        name: Build and push
+        uses: docker/build-push-action@v2
+        with:
+          context: ./docker/transformers-past-gpu
+          build-args: |
+            REF=main
+            FRAMEWORK=pytorch
+            VERSION=${{ matrix.version }}
+          push: true
+          tags: huggingface/transformers-pytorch-past-${{ matrix.version }}-gpu
+
+  past-tensorflow-docker:
+    name: "Past TensorFlow Docker"
+    strategy:
+      fail-fast: false
+      matrix:
+        version: ["2.8", "2.7", "2.6", "2.5"]
+    runs-on: ubuntu-latest
+    steps:
+      -
+        name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      -
+        name: Check out code
+        uses: actions/checkout@v2
+      -
+        name: Login to DockerHub
+        uses: docker/login-action@v1
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+      -
+        name: Build and push
+        uses: docker/build-push-action@v2
+        with:
+          context: ./docker/transformers-past-gpu
+          build-args: |
+            REF=main
+            FRAMEWORK=tensorflow
+            VERSION=${{ matrix.version }}
+          push: true
+          tags: huggingface/transformers-tensorflow-past-${{ matrix.version }}-gpu
+
+  past-tensorflow-docker-2-4:
+    name: "Past TensorFlow Docker"
+    strategy:
+      fail-fast: false
+      matrix:
+        version: ["2.4"]
+    runs-on: ubuntu-latest
+    steps:
+      -
+        name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v1
+      -
+        name: Check out code
+        uses: actions/checkout@v2
+      -
+        name: Login to DockerHub
+        uses: docker/login-action@v1
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+      -
+        name: Build and push
+        uses: docker/build-push-action@v2
+        with:
+          context: ./docker/transformers-past-gpu
+          build-args: |
+            REF=main
+            BASE_DOCKER_IMAGE=nvidia/cuda:11.0.3-cudnn8-devel-ubuntu20.04
+            FRAMEWORK=tensorflow
+            VERSION=${{ matrix.version }}
+          push: true
+          tags: huggingface/transformers-tensorflow-past-${{ matrix.version }}-gpu
--- a/.github/workflows/model-templates.yml
+++ b/.github/workflows/model-templates.yml
@@ -21,7 +21,7 @@ jobs:
        id: cache
        with:
          path: ~/venv/
-          key: v3-tests_templates-${{ hashFiles('setup.py') }}
+          key: v4-tests_templates-${{ hashFiles('setup.py') }}

      - name: Create virtual environment on cache miss
        if: steps.cache.outputs.cache-hit != 'true'
--- a/.github/workflows/self-nightly-scheduled.yml
+++ b/.github/workflows/self-nightly-scheduled.yml
@@ -1,250 +1,236 @@
-name: Self-hosted runner; Nightly (scheduled)
+name: Self-hosted runner (nightly)
+
+# Note that each job's dependencies go into a corresponding docker file.
+#
+# For example for `run_all_tests_torch_cuda_extensions_gpu` the docker image is
+# `huggingface/transformers-pytorch-deepspeed-latest-gpu`, which can be found at
+# `docker/transformers-pytorch-deepspeed-latest-gpu/Dockerfile`

 on:
-    push:
-        branches:
-            - nightly_ci*
-    repository_dispatch:
-    schedule:
-        - cron: "0 0 */3 * *"
+  repository_dispatch:
+  schedule:
+    - cron: "0 16 * * *"

 env:
-    HF_HOME: /mnt/cache
-    TRANSFORMERS_IS_CI: yes
-    RUN_SLOW: yes
-    OMP_NUM_THREADS: 16
-    MKL_NUM_THREADS: 16
-    PYTEST_TIMEOUT: 600
-    SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
+  HF_HOME: /mnt/cache
+  TRANSFORMERS_IS_CI: yes
+  OMP_NUM_THREADS: 8
+  MKL_NUM_THREADS: 8
+  RUN_SLOW: yes
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
+  TF_FORCE_GPU_ALLOW_GROWTH: true
+  RUN_PT_TF_CROSS_TESTS: 1

 jobs:
-    run_all_tests_torch_gpu:
-        runs-on: [self-hosted, docker-gpu, single-gpu]
-        container:
-            image: pytorch/pytorch:1.10.0-cuda11.3-cudnn8-runtime
-            options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
-        steps:
-            - name: Launcher docker
-              uses: actions/checkout@v2
+  setup:
+    name: Setup
+    strategy:
+      matrix:
+        machine_type: [single-gpu, multi-gpu]
+    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    container:
+      image: huggingface/transformers-all-latest-torch-nightly-gpu
+      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    outputs:
+      matrix: ${{ steps.set-matrix.outputs.matrix }}
+    steps:
+      - name: Update clone
+        working-directory: /transformers
+        run: |
+          git fetch && git checkout ${{ github.sha }}

-            - name: NVIDIA-SMI
-              run: |
-                  nvidia-smi
+      - name: Cleanup
+        working-directory: /transformers
+        run: |
+          rm -rf tests/__pycache__
+          rm -rf tests/models/__pycache__
+          rm -rf reports

-            - name: Install dependencies
-              run: |
-                  apt -y update && apt install -y libsndfile1-dev git espeak-ng
-                  pip install --upgrade pip
-                  pip install .[integrations,sklearn,testing,onnxruntime,sentencepiece,torch-speech,vision,timm]
-                  pip install https://github.com/kpu/kenlm/archive/master.zip
-                  pip install --pre torch torchvision torchaudio -f https://download.pytorch.org/whl/nightly/cu113/torch_nightly.html -U
+      - id: set-matrix
+        name: Identify models to test
+        working-directory: /transformers/tests
+        run: |
+          echo "::set-output name=matrix::$(python3 -c 'import os; tests = os.getcwd(); model_tests = os.listdir(os.path.join(tests, "models")); d1 = sorted(list(filter(os.path.isdir, os.listdir(tests)))); d2 = sorted(list(filter(os.path.isdir, [f"models/{x}" for x in model_tests]))); d1.remove("models"); d = d2 + d1; print(d)')"

-            - name: Are GPUs recognized by our DL frameworks
-              run: |
-                utils/print_env.py
+      - name: NVIDIA-SMI
+        run: |
+          nvidia-smi

-            - name: Run all tests on GPU
-              run: |
-                  python -m pytest -n 1 -v --dist=loadfile --make-reports=tests_torch_gpu tests
+  run_tests_single_gpu:
+    name: Model tests
+    strategy:
+      fail-fast: false
+      matrix:
+        folders: ${{ fromJson(needs.setup.outputs.matrix) }}
+        machine_type: [single-gpu]
+    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    container:
+      image: huggingface/transformers-all-latest-torch-nightly-gpu
+      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    needs: setup
+    steps:
+      - 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 }}"
+          matrix_folders=${{ matrix.folders }}
+          matrix_folders=${matrix_folders/'models/'/'models_'}
+          echo "$matrix_folders"
+          echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV

-            - name: Failure short reports
-              if: ${{ always() }}
-              run: cat reports/tests_torch_gpu/failures_short.txt
+      - name: Update clone
+        working-directory: /transformers
+        run: git fetch && git checkout ${{ github.sha }}

-            - name: Run examples tests on GPU
-              if: ${{ always() }}
-              env:
-                  OMP_NUM_THREADS: 16
-                  MKL_NUM_THREADS: 16
-                  RUN_SLOW: yes
-                  HF_HOME: /mnt/cache
-                  TRANSFORMERS_IS_CI: yes
-              run: |
-                  pip install -r examples/pytorch/_tests_requirements.txt
-                  python -m pytest -n 1 -v --dist=loadfile --make-reports=examples_torch_gpu examples
+      - name: NVIDIA-SMI
+        run: |
+          nvidia-smi

-            - name: Failure short reports
-              if: ${{ always() }}
-              run: cat reports/examples_torch_gpu/failures_short.txt
+      - name: Environment
+        working-directory: /transformers
+        run: |
+          python3 utils/print_env.py

-            - name: Run all pipeline tests on GPU
-              if: ${{ always() }}
-              env:
-                  RUN_PIPELINE_TESTS: yes
-              run: |
-                  python -m pytest -n 1 -v --dist=loadfile -m is_pipeline_test --make-reports=tests_torch_pipeline_gpu tests
+      - name: Run all tests on GPU
+        working-directory: /transformers
+        run: python3 -m pytest -v --make-reports=${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }} tests/${{ matrix.folders }}

-            - name: Failure short reports
-              if: ${{ always() }}
-              run: cat reports/tests_torch_pipeline_gpu/failures_short.txt
+      - 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@v2
-              with:
-                  name: run_all_tests_torch_gpu_test_reports
-                  path: reports
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: ${{ matrix.machine_type }}_run_all_tests_gpu_${{ env.matrix_folders }}_test_reports
+          path: /transformers/reports/${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }}

-    run_all_tests_torch_multi_gpu:
-        runs-on: [self-hosted, docker-gpu, multi-gpu]
-        container:
-            image: pytorch/pytorch:1.10.0-cuda11.3-cudnn8-runtime
-            options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
-        steps:
-            - name: Launcher docker
-              uses: actions/checkout@v2
+  run_tests_multi_gpu:
+    name: Model tests
+    strategy:
+      fail-fast: false
+      matrix:
+        folders: ${{ fromJson(needs.setup.outputs.matrix) }}
+        machine_type: [multi-gpu]
+    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    container:
+      image: huggingface/transformers-all-latest-torch-nightly-gpu
+      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    needs: setup
+    steps:
+      - 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 }}"
+          matrix_folders=${{ matrix.folders }}
+          matrix_folders=${matrix_folders/'models/'/'models_'}
+          echo "$matrix_folders"
+          echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV

-            - name: NVIDIA-SMI
-              continue-on-error: true
-              run: |
-                  nvidia-smi
+      - name: Update clone
+        working-directory: /transformers
+        run: git fetch && git checkout ${{ github.sha }}

-            - name: Install dependencies
-              run: |
-                  apt -y update && apt install -y libsndfile1-dev git espeak-ng
-                  pip install --upgrade pip
-                  pip install .[integrations,sklearn,testing,onnxruntime,sentencepiece,torch-speech,vision,timm]
-                  pip install https://github.com/kpu/kenlm/archive/master.zip
-                  pip install --pre torch torchvision torchaudio -f https://download.pytorch.org/whl/nightly/cu113/torch_nightly.html -U
+      - name: NVIDIA-SMI
+        run: |
+          nvidia-smi

-            - name: Are GPUs recognized by our DL frameworks
-              run: |
-                utils/print_env.py
+      - name: Environment
+        working-directory: /transformers
+        run: |
+          python3 utils/print_env.py

-            - name: Run all tests on GPU
-              env:
-                  MKL_SERVICE_FORCE_INTEL: 1
-              run: |
-                  python -m pytest -n 1 -v --dist=loadfile --make-reports=tests_torch_multi_gpu tests
+      - name: Run all tests on GPU
+        working-directory: /transformers
+        run: python3 -m pytest -v --make-reports=${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }} tests/${{ matrix.folders }}

-            - name: Failure short reports
-              if: ${{ always() }}
-              run: cat reports/tests_torch_multi_gpu/failures_short.txt
+      - 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: Run all pipeline tests on GPU
-              if: ${{ always() }}
-              env:
-                  RUN_PIPELINE_TESTS: yes
-              run: |
-                  python -m pytest -n 1 -v --dist=loadfile -m is_pipeline_test --make-reports=tests_torch_pipeline_multi_gpu tests
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: ${{ matrix.machine_type }}_run_all_tests_gpu_${{ env.matrix_folders }}_test_reports
+          path: /transformers/reports/${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }}

-            - name: Failure short reports
-              if: ${{ always() }}
-              run: cat reports/tests_torch_pipeline_multi_gpu/failures_short.txt
+  run_all_tests_torch_cuda_extensions_gpu:
+    name: Torch CUDA extension tests
+    strategy:
+      fail-fast: false
+      matrix:
+        machine_type: [single-gpu, multi-gpu]
+    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker') }}
+    needs: setup
+    container:
+      image: huggingface/transformers-pytorch-deepspeed-nightly-gpu
+      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    steps:
+      - name: Update clone
+        working-directory: /workspace/transformers
+        run: git fetch && git checkout ${{ github.sha }}

-            - name: Test suite reports artifacts
-              if: ${{ always() }}
-              uses: actions/upload-artifact@v2
-              with:
-                  name: run_all_tests_torch_multi_gpu_test_reports
-                  path: reports
+      # To avoid unknown test failures
+      - name: Pre build DeepSpeed *again*
+        working-directory: /workspace
+        run: |
+          python3 -m pip uninstall -y deepspeed
+          rm -rf DeepSpeed
+          git clone https://github.com/microsoft/DeepSpeed && cd DeepSpeed && rm -rf build
+          DS_BUILD_CPU_ADAM=1 DS_BUILD_AIO=1 DS_BUILD_UTILS=1 python3 -m pip install . --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check

-    run_all_tests_torch_cuda_extensions_gpu:
-        runs-on: [self-hosted, docker-gpu, single-gpu]
-        container:
-            image: nvcr.io/nvidia/pytorch:21.03-py3
-            options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
-        steps:
-            - name: Launcher docker
-              uses: actions/checkout@v2
+      - name: NVIDIA-SMI
+        run: |
+          nvidia-smi

-            - name: NVIDIA-SMI
-              run: |
-                  nvidia-smi
+      - name: Environment
+        working-directory: /workspace/transformers
+        run: |
+          python utils/print_env.py

-            - name: Install dependencies
-              run: |
-                  apt -y update && apt install -y libaio-dev libsndfile1-dev git espeak-ng
-                  pip install --upgrade pip
-                  pip install --pre torch torchvision torchaudio -f https://download.pytorch.org/whl/nightly/cu113/torch_nightly.html -U
-                  pip install .[deepspeed-testing]
-                  pip install https://github.com/kpu/kenlm/archive/master.zip
-                  pip install git+https://github.com/microsoft/DeepSpeed
+      - name: Run all tests on GPU
+        working-directory: /workspace/transformers
+        run: |
+          python -m pytest -v --make-reports=${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu tests/deepspeed tests/extended

-            - name: Are GPUs recognized by our DL frameworks
-              run: |
-                utils/print_env.py
+      - name: Failure short reports
+        if: ${{ failure() }}
+        continue-on-error: true
+        run: cat /workspace/transformers/reports/${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu/failures_short.txt

-            - name: Run all tests on GPU
-              run: |
-                  python -m pytest -n 1 -v --dist=loadfile --make-reports=tests_torch_cuda_extensions_gpu tests/deepspeed tests/extended
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: ${{ matrix.machine_type }}_run_tests_torch_cuda_extensions_gpu_test_reports
+          path: /workspace/transformers/reports/${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu

-            - name: Failure short reports
-              if: ${{ always() }}
-              run: cat reports/tests_torch_cuda_extensions_gpu/failures_short.txt
-
-            - name: Test suite reports artifacts
-              if: ${{ always() }}
-              uses: actions/upload-artifact@v2
-              with:
-                  name: run_tests_torch_cuda_extensions_gpu_test_reports
-                  path: reports
-
-    run_all_tests_torch_cuda_extensions_multi_gpu:
-        runs-on: [self-hosted, docker-gpu, multi-gpu]
-        container:
-            image: nvcr.io/nvidia/pytorch:21.03-py3
-            options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
-        steps:
-            - name: Launcher docker
-              uses: actions/checkout@v2
-
-            - name: NVIDIA-SMI
-              continue-on-error: true
-              run: |
-                  nvidia-smi
-
-            - name: Install dependencies
-              run: |
-                  apt -y update && apt install -y libaio-dev libsndfile1-dev git espeak-ng
-                  pip install --upgrade pip
-                  pip install --pre torch torchvision torchaudio -f https://download.pytorch.org/whl/nightly/cu113/torch_nightly.html -U
-                  rm -rf ~/.cache/torch_extensions/ # shared between conflicting builds
-                  pip install .[testing,fairscale]
-                  pip install https://github.com/kpu/kenlm/archive/master.zip
-                  pip install git+https://github.com/microsoft/DeepSpeed # testing bleeding edge
-
-            - name: Are GPUs recognized by our DL frameworks
-              run: |
-                utils/print_env.py
-
-            - name: Run all tests on GPU
-              run: |
-                  python -m pytest -n 1 -v --dist=loadfile --make-reports=tests_torch_cuda_extensions_multi_gpu tests/deepspeed tests/extended
-
-            - name: Failure short reports
-              if: ${{ always() }}
-              run: cat reports/tests_torch_cuda_extensions_multi_gpu/failures_short.txt
-
-            - name: Test suite reports artifacts
-              if: ${{ always() }}
-              uses: actions/upload-artifact@v2
-              with:
-                  name: run_tests_torch_cuda_extensions_multi_gpu_test_reports
-                  path: reports
-
-    send_results:
-        name: Send results to webhook
-        runs-on: ubuntu-latest
-        if: always()
-        needs: [
-                run_all_tests_torch_gpu,
-                run_all_tests_torch_multi_gpu,
-                run_all_tests_torch_cuda_extensions_gpu,
-                run_all_tests_torch_cuda_extensions_multi_gpu
-        ]
-        steps:
-            - uses: actions/checkout@v2
-
-            - uses: actions/download-artifact@v2
-
-            - 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_PAST_FUTURE: ${{ secrets.CI_SLACK_CHANNEL_ID_PAST_FUTURE }}
-
-              run: |
-                  pip install slack_sdk
-                  python utils/notification_service.py scheduled nightly-torch
+  send_results:
+    name: Send results to webhook
+    runs-on: ubuntu-latest
+    if: always()
+    needs: [setup, run_tests_single_gpu, run_tests_multi_gpu, run_all_tests_torch_cuda_extensions_gpu]
+    steps:
+      - uses: actions/checkout@v2
+      - uses: actions/download-artifact@v2
+      - 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_DUMMY_TESTS: ${{ secrets.CI_SLACK_CHANNEL_DUMMY_TESTS }}
+          CI_SLACK_REPORT_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID_PAST_FUTURE }}
+          CI_EVENT: nightly-build
+        # We pass `needs.setup.outputs.matrix` as the argument. A processing in `notification_service.py` to change
+        # `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
+        run: |
+          pip install slack_sdk
+          python utils/notification_service.py "${{ needs.setup.outputs.matrix }}"
--- a/.github/workflows/self-past-caller.yml
+++ b/.github/workflows/self-past-caller.yml
@@ -0,0 +1,136 @@
+name: Self-hosted runner (past-ci-caller)
+
+on:
+  push:
+    branches:
+      - run-past-ci*
+
+jobs:
+  run_past_ci_pytorch_1-11:
+    name: PyTorch 1.11
+    if: always()
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: pytorch
+      version: "1.11"
+    secrets: inherit
+
+  run_past_ci_pytorch_1-10:
+    name: PyTorch 1.10
+    if: always()
+    needs: [run_past_ci_pytorch_1-11]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: pytorch
+      version: "1.10"
+    secrets: inherit
+
+  run_past_ci_pytorch_1-9:
+    name: PyTorch 1.9
+    if: always()
+    needs: [run_past_ci_pytorch_1-10]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: pytorch
+      version: "1.9"
+    secrets: inherit
+
+  run_past_ci_pytorch_1-8:
+    name: PyTorch 1.8
+    if: always()
+    needs: [run_past_ci_pytorch_1-9]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: pytorch
+      version: "1.8"
+    secrets: inherit
+
+  run_past_ci_pytorch_1-7:
+    name: PyTorch 1.7
+    if: always()
+    needs: [run_past_ci_pytorch_1-8]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: pytorch
+      version: "1.7"
+    secrets: inherit
+
+  run_past_ci_pytorch_1-6:
+    name: PyTorch 1.6
+    if: always()
+    needs: [run_past_ci_pytorch_1-7]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: pytorch
+      version: "1.6"
+    secrets: inherit
+
+  run_past_ci_pytorch_1-5:
+    name: PyTorch 1.5
+    if: always()
+    needs: [run_past_ci_pytorch_1-6]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: pytorch
+      version: "1.5"
+    secrets: inherit
+
+  run_past_ci_pytorch_1-4:
+    name: PyTorch 1.4
+    if: always()
+    needs: [run_past_ci_pytorch_1-5]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: pytorch
+      version: "1.4"
+    secrets: inherit
+
+  run_past_ci_tensorflow_2-8:
+    name: TensorFlow 2.8
+    if: always()
+    needs: [run_past_ci_pytorch_1-4]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: tensorflow
+      version: "2.8"
+    secrets: inherit
+
+  run_past_ci_tensorflow_2-7:
+    name: TensorFlow 2.7
+    if: always()
+    needs: [run_past_ci_tensorflow_2-8]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: tensorflow
+      version: "2.7"
+    secrets: inherit
+
+  run_past_ci_tensorflow_2-6:
+    name: TensorFlow 2.6
+    if: always()
+    needs: [run_past_ci_tensorflow_2-7]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: tensorflow
+      version: "2.6"
+    secrets: inherit
+
+  run_past_ci_tensorflow_2-5:
+    name: TensorFlow 2.5
+    if: always()
+    needs: [run_past_ci_tensorflow_2-6]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: tensorflow
+      version: "2.5"
+    secrets: inherit
+
+  run_past_ci_tensorflow_2-4:
+    name: TensorFlow 2.4
+    if: always()
+    needs: [run_past_ci_tensorflow_2-5]
+    uses: ./.github/workflows/self-past.yml
+    with:
+      framework: tensorflow
+      version: "2.4"
+    secrets: inherit
--- a/.github/workflows/self-past.yml
+++ b/.github/workflows/self-past.yml
@@ -0,0 +1,192 @@
+name: Self-hosted runner (past)
+
+# Note that each job's dependencies go into a corresponding docker file.
+#
+# For example for `run_all_tests_torch_cuda_extensions_gpu` the docker image is
+# `huggingface/transformers-pytorch-deepspeed-latest-gpu`, which can be found at
+# `docker/transformers-pytorch-deepspeed-latest-gpu/Dockerfile`
+
+on:
+  workflow_call:
+    inputs:
+      framework:
+        required: true
+        type: string
+      version:
+        required: true
+        type: string
+
+env:
+  HF_HOME: /mnt/cache
+  TRANSFORMERS_IS_CI: yes
+  OMP_NUM_THREADS: 8
+  MKL_NUM_THREADS: 8
+  RUN_SLOW: yes
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
+  TF_FORCE_GPU_ALLOW_GROWTH: true
+  RUN_PT_TF_CROSS_TESTS: 1
+
+jobs:
+  setup:
+    name: Setup
+    runs-on: ubuntu-latest
+    outputs:
+      matrix: ${{ steps.set-matrix.outputs.matrix }}
+    steps:
+      - name: Checkout transformers
+        uses: actions/checkout@v2
+        with:
+          fetch-depth: 2
+
+      - name: Cleanup
+        run: |
+          rm -rf tests/__pycache__
+          rm -rf tests/models/__pycache__
+          rm -rf reports
+
+      - id: set-matrix
+        name: Identify models to test
+        run: |
+          cd tests
+          echo "::set-output name=matrix::$(python3 -c 'import os; tests = os.getcwd(); model_tests = os.listdir(os.path.join(tests, "models")); d1 = sorted(list(filter(os.path.isdir, os.listdir(tests)))); d2 = sorted(list(filter(os.path.isdir, [f"models/{x}" for x in model_tests]))); d1.remove("models"); d = d2 + d1; print(d)')"
+
+  run_tests_single_gpu:
+    name: Model tests
+    strategy:
+      fail-fast: false
+      matrix:
+        folders: ${{ fromJson(needs.setup.outputs.matrix) }}
+        machine_type: [single-gpu]
+    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    container:
+      image: huggingface/transformers-${{ inputs.framework }}-past-${{ inputs.version }}-gpu
+      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    needs: setup
+    steps:
+      - name: Update clone
+        working-directory: /transformers
+        run: git fetch && git checkout ${{ github.sha }}
+
+      - 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 }}"
+          matrix_folders=${{ matrix.folders }}
+          matrix_folders=${matrix_folders/'models/'/'models_'}
+          echo "$matrix_folders"
+          echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
+
+      - name: NVIDIA-SMI
+        run: |
+          nvidia-smi
+
+      - name: Environment
+        working-directory: /transformers
+        run: |
+          python3 utils/print_env.py
+
+      - name: Run all tests on GPU
+        working-directory: /transformers
+        run: python3 -m pytest -v --make-reports=${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }} tests/${{ 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@v2
+        with:
+          name: ${{ matrix.machine_type }}_run_all_tests_gpu_${{ env.matrix_folders }}_test_reports
+          path: /transformers/reports/${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }}
+
+  run_tests_multi_gpu:
+    name: Model tests
+    strategy:
+      fail-fast: false
+      matrix:
+        folders: ${{ fromJson(needs.setup.outputs.matrix) }}
+        machine_type: [multi-gpu]
+    runs-on: ${{ format('{0}-{1}', matrix.machine_type, 'docker-past-ci') }}
+    container:
+      image: huggingface/transformers-${{ inputs.framework }}-past-${{ inputs.version }}-gpu
+      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    needs: setup
+    steps:
+      - name: Update clone
+        working-directory: /transformers
+        run: git fetch && git checkout ${{ github.sha }}
+
+      - 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 }}"
+          matrix_folders=${{ matrix.folders }}
+          matrix_folders=${matrix_folders/'models/'/'models_'}
+          echo "$matrix_folders"
+          echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
+
+      - name: NVIDIA-SMI
+        run: |
+          nvidia-smi
+
+      - name: Environment
+        working-directory: /transformers
+        run: |
+          python3 utils/print_env.py
+
+      - name: Run all tests on GPU
+        working-directory: /transformers
+        run: python3 -m pytest -v --make-reports=${{ matrix.machine_type }}_tests_gpu_${{ matrix.folders }} tests/${{ 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@v2
+        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: [setup, run_tests_single_gpu, run_tests_multi_gpu]
+    steps:
+      - uses: actions/checkout@v2
+      - uses: actions/download-artifact@v2
+
+      # Create a directory to store test failure tables in the next step
+      - name: Create directory
+        run: mkdir test_failure_tables
+
+      - 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_DUMMY_TESTS: ${{ secrets.CI_SLACK_CHANNEL_DUMMY_TESTS }}
+          CI_SLACK_REPORT_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID_PAST_FUTURE }}
+          CI_EVENT: Past CI - ${{ inputs.framework }}-${{ inputs.version }}
+        # We pass `needs.setup.outputs.matrix` as the argument. A processing in `notification_service.py` to change
+        # `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
+        run: |
+          pip install slack_sdk
+          python utils/notification_service.py "${{ needs.setup.outputs.matrix }}"
+
+      # Upload complete failure tables, as they might be big and only truncated versions could be sent to Slack.
+      - name: Failure table artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: test_failure_tables_${{ inputs.framework }}-${{ inputs.version }}
+          path: test_failure_tables
--- a/.github/workflows/self-push-caller.yml
+++ b/.github/workflows/self-push-caller.yml
@@ -1,3 +1,4 @@
+# Used to trigger self-push CI
 name: Self-hosted runner (push-caller)

 on:
@@ -12,18 +13,40 @@ on:
      - "utils/**"

 jobs:
-  run_push_ci:
-    name: Run Push CI
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout transformers
-        uses: actions/checkout@v2
-        with:
-          fetch-depth: 2
-          ssh-key: "${{ secrets.COMMIT_KEY }}"
+  check-for-setup:
+      runs-on: ubuntu-latest
+      name: Check if setup was changed
+      outputs:
+        changed: ${{ steps.was_changed.outputs.changed }}
+      steps:
+        - uses: actions/checkout@v3
+          with: 
+            fetch-depth: "2"
+        
+        - name: Get changed files
+          id: changed-files
+          uses: tj-actions/changed-files@v22.2
+        
+        - name: Was setup changed 
+          id: was_changed
+          run: |
+            for file in ${{ steps.changed-files.outputs.all_changed_files }}; do
+              if [ `basename "${file}"` = "setup.py" ]; then
+                echo ::set-output name=changed::"1"
+              fi
+            done

-      - name: Checkout to branch push-ci
-        # A more strict way to make sure`push-ci` is exactly the same as `main` at the push event commit.
-        run: |
-          git checkout -b push-ci
-          git push -u origin push-ci --force
+  build-docker-containers:
+    needs: check-for-setup
+    if: (github.event_name == 'push') && (needs.check-for-setup.outputs.changed == '1')
+    uses: ./.github/workflows/build-docker-images.yml
+    secrets: inherit
+
+  run_push_ci:
+    name: Trigger Push CI
+    runs-on: ubuntu-latest
+    if: ${{ always() }}
+    needs: build-docker-containers
+    steps:
+      - name: Trigger push CI via workflow_run
+        run: echo "Trigger push CI via workflow_run"
--- a/.github/workflows/self-push.yml
+++ b/.github/workflows/self-push.yml
@@ -1,9 +1,12 @@
 name: Self-hosted runner (push)

 on:
+  workflow_run:
+    workflows: ["Self-hosted runner (push-caller)"]
+    branches: ["main"]
+    types: [completed]
  push:
    branches:
-      - push-ci
      - ci_*
      - ci-*
    paths:
@@ -31,11 +34,47 @@ jobs:
      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: Checkout transformers
        uses: actions/checkout@v2
        with:
          fetch-depth: 2

+      - 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 = $(git log -n 1)"
+
      - name: Cleanup
        run: |
          rm -rf tests/__pycache__
@@ -87,6 +126,38 @@ jobs:
      image: huggingface/transformers-all-latest-gpu
      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
    steps:
+      # Necessary to get the correct branch name and commit SHA for `workflow_run` event
+      # We also take into account the `push` event (we might want to test some changes in a branch)
+      - name: Prepare custom environment variables
+        shell: bash
+        # For the meaning of these environment variables, see the job `Setup`
+        run: |
+          CI_BRANCH_PUSH=${{ 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: Echo folder ${{ matrix.folders }}
        shell: bash
        # For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
@@ -99,10 +170,6 @@ jobs:
          echo "$matrix_folders"
          echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV

-      - name: Update clone
-        working-directory: /transformers
-        run: git fetch && git checkout ${{ github.sha }}
-
      - name: NVIDIA-SMI
        run: |
          nvidia-smi
@@ -144,6 +211,38 @@ jobs:
      image: huggingface/transformers-all-latest-gpu
      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
    steps:
+      # Necessary to get the correct branch name and commit SHA for `workflow_run` event
+      # We also take into account the `push` event (we might want to test some changes in a branch)
+      - name: Prepare custom environment variables
+        shell: bash
+        # For the meaning of these environment variables, see the job `Setup`
+        run: |
+          CI_BRANCH_PUSH=${{ 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: Echo folder ${{ matrix.folders }}
        shell: bash
        # For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
@@ -156,10 +255,6 @@ jobs:
          echo "$matrix_folders"
          echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV

-      - name: Update clone
-        working-directory: /transformers
-        run: git fetch && git checkout ${{ github.sha }}
-
      - name: NVIDIA-SMI
        run: |
          nvidia-smi
@@ -201,13 +296,41 @@ jobs:
      image: huggingface/transformers-pytorch-deepspeed-latest-gpu
      options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
    steps:
-      - name: Update clone
+      # Necessary to get the correct branch name and commit SHA for `workflow_run` event
+      # We also take into account the `push` event (we might want to test some changes in a branch)
+      - name: Prepare custom environment variables
+        shell: bash
+        # For the meaning of these environment variables, see the job `Setup`
+        run: |
+          CI_BRANCH_PUSH=${{ 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: /workspace/transformers
-        run: git fetch && git checkout ${{ github.sha }}
+        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)"

      # To avoid unknown test failures
      - name: Pre build DeepSpeed *again*
-        working-directory: /workspace/transformers
+        working-directory: /workspace
        run: |
          python3 -m pip uninstall -y deepspeed
          DS_BUILD_CPU_ADAM=1 DS_BUILD_AIO=1 DS_BUILD_UTILS=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check
@@ -217,10 +340,12 @@ jobs:
          nvidia-smi

      - name: Environment
+        working-directory: /workspace/transformers
        run: |
          python utils/print_env.py

      - name: Run all non-slow selected tests on GPU
+        working-directory: /workspace/transformers
        # TODO: Here we pass all tests in the 2 folders for simplicity. It's better to pass only the identified tests.
        run: |
          python -m pytest -n 1 --dist=loadfile -v --make-reports=${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu tests/deepspeed tests/extended
@@ -228,14 +353,14 @@ jobs:
      - name: Failure short reports
        if: ${{ failure() }}
        continue-on-error: true
-        run: cat reports/${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu/failures_short.txt
+        run: cat /workspace/transformers/reports/${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu/failures_short.txt

      - name: Test suite reports artifacts
        if: ${{ always() }}
        uses: actions/upload-artifact@v2
        with:
          name: ${{ matrix.machine_type }}_run_tests_torch_cuda_extensions_gpu_test_reports
-          path: reports/${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu
+          path: /workspace/transformers/reports/${{ matrix.machine_type }}_tests_torch_cuda_extensions_gpu

  run_tests_torch_cuda_extensions_multi_gpu:
    name: Torch CUDA extension tests
@@ -250,13 +375,41 @@ jobs:
      image: huggingface/transformers-pytorch-deepspeed-latest-gpu
      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
    steps:
-      - name: Update clone
+      # 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: /workspace/transformers
-        run: git fetch && git checkout ${{ github.sha }}
+        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)"

      # To avoid unknown test failures
      - name: Pre build DeepSpeed *again*
-        working-directory: /workspace/transformers
+        working-directory: /workspace
        run: |
          python3 -m pip uninstall -y deepspeed
          DS_BUILD_CPU_ADAM=1 DS_BUILD_AIO=1 DS_BUILD_UTILS=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check
@@ -300,7 +453,39 @@ jobs:
        run_tests_torch_cuda_extensions_multi_gpu
    ]
    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 }}"
+
      - uses: actions/checkout@v2
+
+      - 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@v2
      - name: Send message to Slack
        env:
@@ -310,8 +495,9 @@ jobs:
          CI_SLACK_CHANNEL_DUMMY_TESTS: ${{ secrets.CI_SLACK_CHANNEL_DUMMY_TESTS }}
          CI_SLACK_REPORT_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID }}
          CI_EVENT: push
-          CI_TITLE: ${{ github.event.head_commit.message }}
-          CI_COMMIT_URL: ${{ github.event.head_commit.url }}
+          CI_TITLE_PUSH: ${{ github.event.head_commit.message }}
+          CI_TITLE_WORKFLOW_RUN: ${{ github.event.workflow_run.head_commit.message }}
+          CI_SHA: ${{ env.CI_SHA }}
        # We pass `needs.setup.outputs.matrix` as the argument. A processing in `notification_service.py` to change
        # `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
        run: |
--- a/.github/workflows/self-scheduled.yml
+++ b/.github/workflows/self-scheduled.yml
@@ -308,7 +308,7 @@ jobs:

      # To avoid unknown test failures
      - name: Pre build DeepSpeed *again*
-        working-directory: /workspace/transformers
+        working-directory: /workspace
        run: |
          python3 -m pip uninstall -y deepspeed
          DS_BUILD_CPU_ADAM=1 DS_BUILD_AIO=1 DS_BUILD_UTILS=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check
--- a/.github/workflows/update_metdata.yml
+++ b/.github/workflows/update_metdata.yml
@@ -21,7 +21,7 @@ jobs:
        id: cache
        with:
          path: ~/venv/
-          key: v2-metadata-${{ hashFiles('setup.py') }}
+          key: v3-metadata-${{ hashFiles('setup.py') }}

      - name: Create virtual environment on cache miss
        if: steps.cache.outputs.cache-hit != 'true'
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -128,7 +128,7 @@ You will need basic `git` proficiency to be able to contribute to
 manual. Type `git --help` in a shell and enjoy. If you prefer books, [Pro
 Git](https://git-scm.com/book/en/v2) is a very good reference.

-Follow these steps to start contributing:
+Follow these steps to start contributing ([supported Python versions](https://github.com/huggingface/transformers/blob/main/setup.py#L426)):

 1. Fork the [repository](https://github.com/huggingface/transformers) by
   clicking on the 'Fork' button on the repository's page. This creates a copy of the code
--- a/2
+++ b/2
@@ -51,6 +51,7 @@ quality:
 	python utils/sort_auto_mappings.py --check_only
 	flake8 $(check_dirs)
 	doc-builder style src/transformers docs/source --max_len 119 --check_only --path_to_docs docs/source
+	python utils/check_doc_toc.py

 # Format source code automatically and check is there are any problems left that need manual fixing

@@ -58,6 +59,7 @@ extra_style_checks:
 	python utils/custom_init_isort.py
 	python utils/sort_auto_mappings.py
 	doc-builder style src/transformers docs/source --max_len 119 --path_to_docs docs/source
+	python utils/check_doc_toc.py --fix_and_overwrite

 # this target runs checks on all files and potentially modifies some of them

--- a/README.md
+++ b/README.md
@@ -116,22 +116,46 @@ To immediately use a model on a given input (text, image, audio, ...), we provid

 The second line of code downloads and caches the pretrained model used by the pipeline, while the third evaluates it on the given text. Here the answer is "positive" with a confidence of 99.97%.

-Many NLP tasks have a pre-trained `pipeline` ready to go. For example, we can easily extract question answers given context:
+Many tasks have a pre-trained `pipeline` ready to go, in NLP but also in computer vision and speech. For example, we can easily extract detected objects in an image:

 ``` python
+>>> import requests
+>>> from PIL import Image
 >>> from transformers import pipeline

-# Allocate a pipeline for question-answering
->>> question_answerer = pipeline('question-answering')
->>> question_answerer({
-...     'question': 'What is the name of the repository ?',
-...     'context': 'Pipeline has been included in the huggingface/transformers repository'
-... })
-{'score': 0.30970096588134766, 'start': 34, 'end': 58, 'answer': 'huggingface/transformers'}
+# Download an image with cute cats
+>>> 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)

+# Allocate a pipeline for object detection
+>>> 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}}]
 ```

-In addition to the answer, the pretrained model used here returned its confidence score, along with the start position and end position of the answer in the tokenized sentence. You can learn more about the tasks supported by the `pipeline` API in [this tutorial](https://huggingface.co/docs/transformers/task_summary).
+Here we get a list of objects detected in the image, with a box surrounding the object and a confidence score. Here is the original image on the right, with the predictions displayed on the left:
+
+<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>
+
+You can learn more about the tasks supported by the `pipeline` API in [this tutorial](https://huggingface.co/docs/transformers/task_summary).

 To download and use any of the pretrained models on your given task, all it takes is three lines of code. Here is the PyTorch version:
 ```python
@@ -143,6 +167,7 @@ To download and use any of the pretrained models on your given task, all it take
 >>> inputs = tokenizer("Hello world!", return_tensors="pt")
 >>> outputs = model(**inputs)
 ```
+
 And here is the equivalent code for TensorFlow:
 ```python
 >>> from transformers import AutoTokenizer, TFAutoModel
@@ -240,18 +265,19 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 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. **[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. **[BLOOM](https://huggingface.co/docs/transformers/main/model_doc/bloom)** (from BigScience workshop) released by the [BigSicence Workshop](https://bigscience.huggingface.co/).
+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)** (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. **[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. **[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. **[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. **[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/main/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. **[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. **[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. **[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/main/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/main/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. **[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.
@@ -265,42 +291,48 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 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. **[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. **[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/main/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. **[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. **[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. **[GLPN](https://huggingface.co/docs/transformers/main/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. **[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/main/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](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-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. **[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. **[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. **[ImageGPT](https://huggingface.co/docs/transformers/main/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. **[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. **[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/main/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. **[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/layoutlmv2)** (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/main/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. **[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. **[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/main/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. **[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/main/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. **[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. **[MaskFormer](https://huggingface.co/docs/transformers/main/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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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.
@@ -311,7 +343,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 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/main/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. **[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.
@@ -327,10 +359,11 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 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. **[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/main/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. **[Trajectory Transformer](https://huggingface.co/docs/transformers/main/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. **[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. **[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. **[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. **[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. **[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.
@@ -339,7 +372,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 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. **[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. **[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/main/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. **[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. **[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.
@@ -350,7 +383,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 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/main/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. **[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.

--- a/README_ko.md
+++ b/README_ko.md
@@ -221,18 +221,19 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 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. **[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. **[BLOOM](https://huggingface.co/docs/transformers/main/model_doc/bloom)** (from BigScience workshop) released by the [BigSicence Workshop](https://bigscience.huggingface.co/).
+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)** (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. **[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. **[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. **[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. **[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/main/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. **[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. **[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. **[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/main/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/main/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. **[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.
@@ -246,42 +247,48 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 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. **[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. **[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/main/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. **[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. **[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. **[GLPN](https://huggingface.co/docs/transformers/main/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. **[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/main/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](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-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. **[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. **[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. **[ImageGPT](https://huggingface.co/docs/transformers/main/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. **[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. **[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/main/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. **[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/layoutlmv2)** (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/main/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. **[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. **[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/main/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. **[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/main/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. **[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. **[MaskFormer](https://huggingface.co/docs/transformers/main/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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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.
@@ -292,7 +299,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 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/main/model_doc/regnet)** (from META Research) 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. **[RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)** (from META Research) 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/pdf/2010.12821.pdf) 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 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.
@@ -308,10 +315,11 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 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. **[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/main/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. **[Trajectory Transformer](https://huggingface.co/docs/transformers/main/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. **[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. **[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. **[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. **[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. **[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/pdf/2202.09741.pdf) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
@@ -320,7 +328,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 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. **[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. **[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/main/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. **[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. **[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. 
@@ -331,7 +339,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 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/main/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. **[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. 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. 

@@ -366,4 +374,4 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
    url = "https://www.aclweb.org/anthology/2020.emnlp-demos.6",
    pages = "38--45"
 }
-```
+```
--- a/README_zh-hans.md
+++ b/README_zh-hans.md
@@ -245,18 +245,19 @@ conda install -c huggingface transformers
 1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (来自 Google Research) 伴随论文 [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) 由 Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed 发布。
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (来自 Facebook) 伴随论文 [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) 由 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 发布。
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (来自 Facebook) 伴随论文 [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) 由 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 发布。
-1. **[BLOOM](https://huggingface.co/docs/transformers/main/model_doc/bloom)** (from BigScience workshop) released by the [BigSicence Workshop](https://bigscience.huggingface.co/). 
+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)** (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. **[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。
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (来自 OpenAI) 伴随论文 [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) 由 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 发布。
+1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (来自 Salesforce) 伴随论文 [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) 由 Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong 发布。
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (来自 YituTech) 伴随论文 [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) 由 Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan 发布。
-1. **[ConvNeXT](https://huggingface.co/docs/transformers/main/model_doc/convnext)** (来自 Facebook AI) 伴随论文 [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) 由 Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie 发布。
+1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (来自 Facebook AI) 伴随论文 [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) 由 Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie 发布。
 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (来自 Tsinghua University) 伴随论文 [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) 由 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 发布。
 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (来自 Salesforce) 伴随论文 [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) 由 Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher 发布。
-1. **[CvT](https://huggingface.co/docs/transformers/main/model_doc/cvt)** (来自 Microsoft) 伴随论文 [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) 由 Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang 发布。
-1. **[Data2Vec](https://huggingface.co/docs/transformers/main/model_doc/data2vec)** (来自 Facebook) 伴随论文 [Data2Vec:  A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) 由 Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli 发布。
+1. **[CvT](https://huggingface.co/docs/transformers/model_doc/cvt)** (来自 Microsoft) 伴随论文 [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) 由 Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang 发布。
+1. **[Data2Vec](https://huggingface.co/docs/transformers/model_doc/data2vec)** (来自 Facebook) 伴随论文 [Data2Vec:  A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) 由 Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli 发布。
 1. **[DeBERTa](https://huggingface.co/docs/transformers/model_doc/deberta)** (来自 Microsoft) 伴随论文 [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) 由 Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen 发布。
 1. **[DeBERTa-v2](https://huggingface.co/docs/transformers/model_doc/deberta-v2)** (来自 Microsoft) 伴随论文 [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) 由 Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen 发布。
 1. **[Decision Transformer](https://huggingface.co/docs/transformers/model_doc/decision_transformer)** (来自 Berkeley/Facebook/Google) 伴随论文 [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://arxiv.org/abs/2106.01345) 由 Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch 发布。
@@ -270,42 +271,48 @@ conda install -c huggingface transformers
 1. **[ELECTRA](https://huggingface.co/docs/transformers/model_doc/electra)** (来自 Google Research/Stanford University) 伴随论文 [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) 由 Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning 发布。
 1. **[EncoderDecoder](https://huggingface.co/docs/transformers/model_doc/encoder-decoder)** (来自 Google Research) 伴随论文 [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) 由 Sascha Rothe, Shashi Narayan, Aliaksei Severyn 发布。
 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (来自 CNRS) 伴随论文 [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) 由 Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab 发布。
-1. **[FLAVA](https://huggingface.co/docs/transformers/main/model_doc/flava)** (来自 Facebook AI) 伴随论文 [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) 由 Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela 发布。
+1. **[FLAVA](https://huggingface.co/docs/transformers/model_doc/flava)** (来自 Facebook AI) 伴随论文 [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) 由 Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela 发布。
 1. **[FNet](https://huggingface.co/docs/transformers/model_doc/fnet)** (来自 Google Research) 伴随论文 [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) 由 James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon 发布。
 1. **[Funnel Transformer](https://huggingface.co/docs/transformers/model_doc/funnel)** (来自 CMU/Google Brain) 伴随论文 [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) 由 Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le 发布。
-1. **[GLPN](https://huggingface.co/docs/transformers/main/model_doc/glpn)** (来自 KAIST) 伴随论文 [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) 由 Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim 发布。
+1. **[GLPN](https://huggingface.co/docs/transformers/model_doc/glpn)** (来自 KAIST) 伴随论文 [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) 由 Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim 发布。
 1. **[GPT](https://huggingface.co/docs/transformers/model_doc/openai-gpt)** (来自 OpenAI) 伴随论文 [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) 由 Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever 发布。
 1. **[GPT Neo](https://huggingface.co/docs/transformers/model_doc/gpt_neo)** (来自 EleutherAI) 随仓库 [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) 发布。作者为 Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy 发布。
-1. **[GPT NeoX](https://huggingface.co/docs/transformers/main/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](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-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (来自 OpenAI) 伴随论文 [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) 由 Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever** 发布。
 1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (来自 EleutherAI) 伴随论文 [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) 由 Ben Wang and Aran Komatsuzaki 发布。
+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. **[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. **[ImageGPT](https://huggingface.co/docs/transformers/main/model_doc/imagegpt)** (来自 OpenAI) 伴随论文 [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) 由 Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever 发布。
+1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (来自 OpenAI) 伴随论文 [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) 由 Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever 发布。
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (来自 Microsoft Research Asia) 伴随论文 [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) 由 Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou 发布。
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (来自 Microsoft Research Asia) 伴随论文 [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) 由 Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou 发布。
-1. **[LayoutLMv3](https://huggingface.co/docs/transformers/main/model_doc/layoutlmv3)** (来自 Microsoft Research Asia) 伴随论文 [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) 由 Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei 发布。
+1. **[LayoutLMv3](https://huggingface.co/docs/transformers/model_doc/layoutlmv3)** (来自 Microsoft Research Asia) 伴随论文 [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) 由 Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei 发布。
 1. **[LayoutXLM](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (来自 Microsoft Research Asia) 伴随论文 [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) 由 Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei 发布。
 1. **[LED](https://huggingface.co/docs/transformers/model_doc/led)** (来自 AllenAI) 伴随论文 [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) 由 Iz Beltagy, Matthew E. Peters, Arman Cohan 发布。
-1. **[LeViT](https://huggingface.co/docs/transformers/main/model_doc/levit)** (来自 Meta AI) 伴随论文 [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) 由 Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze 发布。
+1. **[LeViT](https://huggingface.co/docs/transformers/model_doc/levit)** (来自 Meta AI) 伴随论文 [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) 由 Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze 发布。
 1. **[Longformer](https://huggingface.co/docs/transformers/model_doc/longformer)** (来自 AllenAI) 伴随论文 [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) 由 Iz Beltagy, Matthew E. Peters, Arman Cohan 发布。
-1. **[LongT5](https://huggingface.co/docs/transformers/main/model_doc/longt5)** (来自 Google AI) released 伴随论文 [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) 由 Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang 发布。
+1. **[LongT5](https://huggingface.co/docs/transformers/model_doc/longt5)** (来自 Google AI) released 伴随论文 [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) 由 Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang 发布。
 1. **[LUKE](https://huggingface.co/docs/transformers/model_doc/luke)** (来自 Studio Ousia) 伴随论文 [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) 由 Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto 发布。
 1. **[LXMERT](https://huggingface.co/docs/transformers/model_doc/lxmert)** (来自 UNC Chapel Hill) 伴随论文 [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) 由 Hao Tan and Mohit Bansal 发布。
-1. **[M-CTC-T](https://huggingface.co/docs/transformers/main/model_doc/mctct)** (来自 Facebook) 伴随论文 [Pseudo-Labeling For Massively Multilingual Speech Recognition](https://arxiv.org/abs/2111.00161) 由 Loren Lugosch, Tatiana Likhomanenko, Gabriel Synnaeve, and Ronan Collobert 发布。
+1. **[M-CTC-T](https://huggingface.co/docs/transformers/model_doc/mctct)** (来自 Facebook) 伴随论文 [Pseudo-Labeling For Massively Multilingual Speech Recognition](https://arxiv.org/abs/2111.00161) 由 Loren Lugosch, Tatiana Likhomanenko, Gabriel Synnaeve, and Ronan Collobert 发布。
 1. **[M2M100](https://huggingface.co/docs/transformers/model_doc/m2m_100)** (来自 Facebook) 伴随论文 [Beyond English-Centric Multilingual Machine Translation](https://arxiv.org/abs/2010.11125) 由 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 发布。
 1. **[MarianMT](https://huggingface.co/docs/transformers/model_doc/marian)** 用 [OPUS](http://opus.nlpl.eu/) 数据训练的机器翻译模型由 Jörg Tiedemann 发布。[Marian Framework](https://marian-nmt.github.io/) 由微软翻译团队开发。
-1. **[MaskFormer](https://huggingface.co/docs/transformers/main/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. **[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. **[mBART](https://huggingface.co/docs/transformers/model_doc/mbart)** (来自 Facebook) 伴随论文 [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) 由 Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer 发布。
 1. **[mBART-50](https://huggingface.co/docs/transformers/model_doc/mbart)** (来自 Facebook) 伴随论文 [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) 由 Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, Angela Fan 发布。
 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. **[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. **[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。
+1. **[MobileViT](https://huggingface.co/docs/transformers/model_doc/mobilevit)** (来自 Apple) 伴随论文 [MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer](https://arxiv.org/abs/2110.02178) 由 Sachin Mehta and Mohammad Rastegari 发布。
 1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (来自 Microsoft Research) 伴随论文 [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) 由 Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu 发布。
 1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (来自 Google AI) 伴随论文 [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) 由 Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel 发布。
+1. **[MVP](https://huggingface.co/docs/transformers/model_doc/mvp)** (来自 中国人民大学 AI Box) 伴随论文 [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) 由 Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen 发布。 
+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. **[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. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (来自 Meta AI) 伴随论文 [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) 由 Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al 发布。
+1. **[OWL-ViT](https://huggingface.co/docs/transformers/model_doc/owlvit)** (来自 Google AI) 伴随论文 [Simple Open-Vocabulary Object Detection with Vision Transformers](https://arxiv.org/abs/2205.06230) 由 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 发布。
 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. **[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. **[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。
@@ -316,7 +323,7 @@ conda install -c huggingface transformers
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (来自 Facebook) 伴随论文 [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) 由 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 发布。
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (来自 Google Research) 伴随论文 [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) 由 Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang 发布。
 1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (来自 Google Research) 伴随论文 [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) 由 Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya 发布。
-1. **[RegNet](https://huggingface.co/docs/transformers/main/model_doc/regnet)** (from META Research) 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. **[RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)** (from META Research) 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)** (来自 Google Research) 伴随论文 [Rethinking embedding coupling in pre-trained language models](https://arxiv.org/pdf/2010.12821.pdf) 由 Hyung Won Chung, Thibault Févry, Henry Tsai, M. Johnson, Sebastian Ruder 发布。
 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)** (来自 Facebook), 伴随论文 [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) 由 Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov 发布。
@@ -332,10 +339,11 @@ conda install -c huggingface transformers
 1. **[T5](https://huggingface.co/docs/transformers/model_doc/t5)** (来自 Google AI) 伴随论文 [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) 由 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 发布。
 1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (来自 Google AI) 伴随论文 [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) 由 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 发布。
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (来自 Google AI) 伴随论文 [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) 由 Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos 发布。
-1. **[TAPEX](https://huggingface.co/docs/transformers/main/model_doc/tapex)** (来自 Microsoft Research) 伴随论文 [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) 由 Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou 发布。
-1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/main/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. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (来自 Microsoft Research) 伴随论文 [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) 由 Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou 发布。
+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)** (来自 Google/CMU) 伴随论文 [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) 由 Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov 发布。
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (来自 Microsoft) 伴随论文 [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) 由 Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei 发布。
+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. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (来自 Microsoft Research) 伴随论文 [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) 由 Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang 发布。
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (来自 Microsoft Research) 伴随论文 [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) 由 Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu 发布。
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (来自 Tsinghua University and Nankai University) 伴随论文 [Visual Attention Network](https://arxiv.org/pdf/2202.09741.pdf) 由 Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu 发布。
@@ -344,7 +352,7 @@ conda install -c huggingface transformers
 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. **[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. **[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/main/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。
+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。
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/transformers/model_doc/wav2vec2_phoneme)** (来自 Facebook AI) 伴随论文 [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) 由 Qiantong Xu, Alexei Baevski, Michael Auli 发布。
 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. **[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. 
@@ -355,7 +363,7 @@ conda install -c huggingface transformers
 1. **[XLNet](https://huggingface.co/docs/transformers/model_doc/xlnet)** (来自 Google/CMU) 伴随论文 [XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) 由 Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le 发布。
 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/main/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. **[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. 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。

@@ -371,7 +379,7 @@ conda install -c huggingface transformers
 | [文档](https://huggingface.co/transformers/) | 完整的 API 文档和教程 |
 | [任务总结](https://huggingface.co/docs/transformers/task_summary) | 🤗 Transformers 支持的任务 |
 | [预处理教程](https://huggingface.co/docs/transformers/preprocessing) | 使用 `Tokenizer` 来为模型准备数据 |
-| [训练和微调](https://huggingface.co/docstransformers/training) | 在 PyTorch/TensorFlow 的训练循环或 `Trainer` API 中使用 🤗 Transformers 提供的模型 |
+| [训练和微调](https://huggingface.co/docs/transformers/training) | 在 PyTorch/TensorFlow 的训练循环或 `Trainer` API 中使用 🤗 Transformers 提供的模型 |
 | [快速上手：微调和用例脚本](https://github.com/huggingface/transformers/tree/main/examples) | 为各种任务提供的用例脚本 |
 | [模型分享和上传](https://huggingface.co/docs/transformers/model_sharing) | 和社区上传和分享你微调的模型 |
 | [迁移](https://huggingface.co/docs/transformers/migration) | 从 `pytorch-transformers` 或 `pytorch-pretrained-bert` 迁移到 🤗 Transformers |
@@ -391,4 +399,4 @@ conda install -c huggingface transformers
    url = "https://www.aclweb.org/anthology/2020.emnlp-demos.6",
    pages = "38--45"
 }
-```
+```
--- a/README_zh-hant.md
+++ b/README_zh-hant.md
@@ -257,18 +257,19 @@ conda install -c huggingface transformers
 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. **[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. **[BLOOM](https://huggingface.co/docs/transformers/main/model_doc/bloom)** (from BigScience workshop) released by the [BigSicence Workshop](https://bigscience.huggingface.co/). 
+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)** (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. **[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. **[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. **[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. **[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/main/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. **[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. **[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. **[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/main/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/main/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. **[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.
@@ -282,42 +283,48 @@ conda install -c huggingface transformers
 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. **[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. **[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/main/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. **[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. **[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. **[GLPN](https://huggingface.co/docs/transformers/main/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. **[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/main/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](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-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 with the paper [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
+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. **[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. **[ImageGPT](https://huggingface.co/docs/transformers/main/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. **[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. **[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/main/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. **[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/layoutlmv2)** (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/main/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. **[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. **[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/main/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. **[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/main/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. **[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. **[MaskFormer](https://huggingface.co/docs/transformers/main/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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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. **[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.
@@ -328,7 +335,7 @@ conda install -c huggingface transformers
 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/main/model_doc/regnet)** (from META Research) 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. **[RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)** (from META Research) 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/pdf/2010.12821.pdf) 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 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.
@@ -344,10 +351,11 @@ conda install -c huggingface transformers
 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 with the paper [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. **[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/main/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. **[Trajectory Transformer](https://huggingface.co/docs/transformers/main/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. **[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. **[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 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. **[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. **[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. **[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/pdf/2202.09741.pdf) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
@@ -356,7 +364,7 @@ conda install -c huggingface transformers
 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. **[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. **[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/main/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. **[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. **[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. 
@@ -367,7 +375,7 @@ conda install -c huggingface transformers
 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/main/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. **[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. 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。

@@ -403,4 +411,4 @@ conda install -c huggingface transformers
    url = "https://www.aclweb.org/anthology/2020.emnlp-demos.6",
    pages = "38--45"
 }
-```
+```
--- a/docker/transformers-all-latest-gpu/Dockerfile
+++ b/docker/transformers-all-latest-gpu/Dockerfile
@@ -3,10 +3,13 @@ LABEL maintainer="Hugging Face"

 ARG DEBIAN_FRONTEND=noninteractive

+# Use login shell to read variables from `~/.profile` (to pass dynamic created variables between RUN commands)
+SHELL ["sh", "-lc"]
+
 # The following `ARG` are mainly used to specify the versions explicitly & directly in this docker file, and not meant
 # to be used as arguments for docker build (so far).

-ARG PYTORCH='1.11.0'
+ARG PYTORCH='1.12.0'
 # (not always a valid torch version)
 ARG INTEL_TORCH_EXT='1.11.0'
 # Example: `cu102`, `cu113`, etc.
@@ -21,11 +24,20 @@ ARG REF=main
 RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
 RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime]

-RUN python3 -m pip install --no-cache-dir -U torch==$PYTORCH torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/$CUDA
+# TODO: Handle these in a python utility script
+RUN [ ${#PYTORCH} -gt 0 -a "$PYTORCH" != "pre" ] && VERSION='torch=='$PYTORCH'.*' ||  VERSION='torch'; echo "export VERSION='$VERSION'" >> ~/.profile
+RUN echo torch=$VERSION
+# `torchvision` and `torchaudio` should be installed along with `torch`, especially for nightly build.
+# Currently, let's just use their latest releases (when `torch` is installed with a release version)
+# TODO: We might need to specify proper versions that work with a specific torch version (especially for past CI).
+RUN [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/$CUDA || python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
+
 RUN python3 -m pip install --no-cache-dir -U tensorflow
 RUN python3 -m pip uninstall -y flax jax

-RUN python3 -m pip install --no-cache-dir torch-scatter -f https://data.pyg.org/whl/torch-$PYTORCH+$CUDA.html
+# Use installed torch version for `torch-scatter` to avid to deal with PYTORCH='pre'.
+# If torch is nightly version, the link is likely to be invalid, but the installation falls back to the latest torch-scatter
+RUN python3 -m pip install --no-cache-dir torch-scatter -f https://data.pyg.org/whl/torch-$(python3 -c "from torch import version; print(version.__version__.split('+')[0])")+$CUDA.html
 RUN python3 -m pip install --no-cache-dir intel_extension_for_pytorch==$INTEL_TORCH_EXT+cpu -f https://software.intel.com/ipex-whl-stable

 RUN python3 -m pip install --no-cache-dir git+https://github.com/facebookresearch/detectron2.git pytesseract https://github.com/kpu/kenlm/archive/master.zip
--- a/docker/transformers-past-gpu/Dockerfile
+++ b/docker/transformers-past-gpu/Dockerfile
@@ -0,0 +1,43 @@
+ARG BASE_DOCKER_IMAGE="nvidia/cuda:11.2.2-cudnn8-devel-ubuntu20.04"
+FROM $BASE_DOCKER_IMAGE
+LABEL maintainer="Hugging Face"
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# Use login shell to read variables from `~/.profile` (to pass dynamic created variables between RUN commands)
+SHELL ["sh", "-lc"]
+
+RUN apt update
+RUN apt install -y git libsndfile1-dev tesseract-ocr espeak-ng python3 python3-pip ffmpeg git-lfs
+RUN git lfs install
+RUN python3 -m pip install --no-cache-dir --upgrade pip
+
+ARG REF=main
+RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
+RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime]
+
+# 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
+
+ARG FRAMEWORK
+ARG VERSION
+
+# Remove all frameworks
+# (`accelerate` requires `torch`, and this causes import issues for TF-only testing)
+RUN python3 -m pip uninstall -y torch torchvision torchaudio accelerate tensorflow jax flax
+
+# Get the libraries and their versions to install, and write installation command to `~/.profile`.
+RUN python3 ./transformers/utils/past_ci_versions.py --framework $FRAMEWORK --version $VERSION
+
+# Install the target framework
+RUN echo "INSTALL_CMD = $INSTALL_CMD"
+RUN $INSTALL_CMD
+
+# Having installation problems for torch-scatter with torch <= 1.6. Disable so we have the same set of tests.
+# (This part will be removed once the logic of using `past_ci_versions.py` is used in other Dockerfile files.)
+# # Use installed torch version for `torch-scatter`.
+# # (The env. variable $CUDA is defined in `past_ci_versions.py`)
+# RUN [ "$FRAMEWORK" = "pytorch" ] && python3 -m pip install --no-cache-dir torch-scatter -f https://data.pyg.org/whl/torch-$(python3 -c "from torch import version; print(version.__version__.split('+')[0])")+$CUDA.html || echo "torch-scatter not to be installed"
+
+RUN python3 -m pip install -U "itsdangerous<2.1.0"
--- a/docker/transformers-pytorch-deepspeed-latest-gpu/Dockerfile
+++ b/docker/transformers-pytorch-deepspeed-latest-gpu/Dockerfile
@@ -3,6 +3,10 @@ LABEL maintainer="Hugging Face"

 ARG DEBIAN_FRONTEND=noninteractive

+ARG PYTORCH='1.12.0'
+# Example: `cu102`, `cu113`, etc.
+ARG CUDA='cu113'
+
 RUN apt -y update
 RUN apt install -y libaio-dev
 RUN python3 -m pip install --no-cache-dir --upgrade pip
@@ -13,13 +17,16 @@ RUN git clone https://github.com/huggingface/transformers && cd transformers &&
 # Install latest release PyTorch
 # (PyTorch must be installed before pre-compiling any DeepSpeed c++/cuda ops.)
 # (https://www.deepspeed.ai/tutorials/advanced-install/#pre-install-deepspeed-ops)
-RUN python3 -m pip install --no-cache-dir -U torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
+RUN python3 -m pip install --no-cache-dir -U torch==$PYTORCH torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/$CUDA

 RUN python3 -m pip install --no-cache-dir ./transformers[deepspeed-testing]

-# Pre-build DeepSpeed, so it would be ready for testing (otherwise, the 1st deepspeed test will timeout)
+# Pre-build **latest** DeepSpeed, so it would be ready for testing (otherwise, the 1st deepspeed test will timeout)
 RUN python3 -m pip uninstall -y deepspeed
-RUN DS_BUILD_CPU_ADAM=1 DS_BUILD_AIO=1 DS_BUILD_UTILS=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check
+# This has to be run (again) inside the GPU VMs running the tests.
+# The installation works here, but some tests fail, if we don't pre-build deepspeed again in the VMs running the tests.
+# TODO: Find out why test fail.
+RUN DS_BUILD_CPU_ADAM=1 DS_BUILD_AIO=1 DS_BUILD_UTILS=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check 2>&1

 # When installing in editable mode, `transformers` is not recognized as a package.
 # this line must be added in order for python to be aware of transformers.
--- a/docker/transformers-pytorch-deepspeed-nightly-gpu/Dockerfile
+++ b/docker/transformers-pytorch-deepspeed-nightly-gpu/Dockerfile
@@ -0,0 +1,35 @@
+FROM nvcr.io/nvidia/pytorch:21.03-py3
+LABEL maintainer="Hugging Face"
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+# Example: `cu102`, `cu113`, etc.
+ARG CUDA='cu113'
+
+RUN apt -y update
+RUN apt install -y libaio-dev
+RUN python3 -m pip install --no-cache-dir --upgrade pip
+
+ARG REF=main
+RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
+
+# Install **nightly** release PyTorch (flag `--pre`)
+# (PyTorch must be installed before pre-compiling any DeepSpeed c++/cuda ops.)
+# (https://www.deepspeed.ai/tutorials/advanced-install/#pre-install-deepspeed-ops)
+RUN python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
+
+RUN python3 -m pip install --no-cache-dir ./transformers[deepspeed-testing]
+
+# Pre-build **nightly** release of DeepSpeed, so it would be ready for testing (otherwise, the 1st deepspeed test will timeout)
+RUN python3 -m pip uninstall -y deepspeed
+# This has to be run inside the GPU VMs running the tests. (So far, it fails here due to GPU checks during compilation.)
+# Issue: https://github.com/microsoft/DeepSpeed/issues/2010
+# RUN git clone https://github.com/microsoft/DeepSpeed && cd DeepSpeed && rm -rf build && \
+#    DS_BUILD_CPU_ADAM=1 DS_BUILD_AIO=1 DS_BUILD_UTILS=1 python3 -m pip install . --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check 2>&1
+
+# When installing in editable mode, `transformers` is not recognized as a package.
+# this line must be added in order for python to be aware of transformers.
+RUN cd transformers && python3 setup.py develop
+
+# Disable for now as deepspeed is not installed above. To be enabled once the issue is fixed.
+# RUN python3 -c "from deepspeed.launcher.runner import main"
--- a/docker/transformers-pytorch-gpu/Dockerfile
+++ b/docker/transformers-pytorch-gpu/Dockerfile
@@ -12,7 +12,7 @@ RUN git clone https://github.com/huggingface/transformers && cd transformers &&
 RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-torch,testing]

 # If set to nothing, will install the latest version
-ARG PYTORCH=''
+ARG PYTORCH='1.12.0'
 ARG TORCH_VISION=''
 ARG TORCH_AUDIO=''

--- a/docs/README.md
+++ b/docs/README.md
@@ -50,11 +50,32 @@ You can adapt the `--build_dir` to set any temporary folder that you prefer. Thi
 the MDX files that will be rendered as the documentation on the main website. You can inspect them in your favorite
 Markdown editor.

+## Previewing the documentation
+
+To preview the docs, first install the `watchdog` module with:
+
+```bash
+pip install watchdog
+```
+
+Then run the following command:
+
+```bash
+doc-builder preview {package_name} {path_to_docs}
+```
+
+For example:
+
+```bash
+doc-builder preview transformers docs/source/en/
+```
+
+The docs will be viewable at [http://localhost:3000](http://localhost:3000). You can also preview the docs once you have opened a PR. You will see a bot add a comment to a link where the documentation with your changes lives.
+
 ---
 **NOTE**

-It's not possible to see locally how the final documentation will look like for now. Once you have opened a PR, you
-will see a bot add a comment to a link where the documentation with your changes lives.
+The `preview` command only works with existing doc files. When you add a completely new file, you need to update `_toctree.yml` & restart `preview` command (`ctrl-c` to stop it & call `doc-builder preview ...` again).

 ---

--- a/docs/source/en/_toctree.yml
+++ b/docs/source/en/_toctree.yml
@@ -22,7 +22,7 @@
  title: Tutorials
 - sections:
  - local: fast_tokenizers
-    title: "Use tokenizers from 🤗 Tokenizers"
+    title: Use tokenizers from 🤗 Tokenizers
  - local: create_a_model
    title: Create a custom architecture
  - local: custom_models
@@ -59,7 +59,29 @@
    title: Converting TensorFlow Checkpoints
  - local: serialization
    title: Export 🤗 Transformers models
-  - local: performance
+  - sections:
+    - local: performance
+      title: Overview
+    - local: perf_train_gpu_one
+      title: Training on one GPU
+    - local: perf_train_gpu_many
+      title: Training on many GPUs
+    - local: perf_train_cpu
+      title: Training on CPU
+    - local: perf_train_tpu
+      title: Training on TPUs
+    - local: perf_train_special
+      title: Training on Specialized Hardware
+    - local: perf_infer_cpu
+      title: Inference on CPU
+    - local: perf_infer_gpu_one
+      title: Inference on one GPU
+    - local: perf_infer_gpu_many
+      title: Inference on many GPUs
+    - local: perf_infer_special
+      title: Inference on Specialized Hardware
+    - local: perf_hardware
+      title: Custom hardware for training
    title: Performance and scalability
  - local: big_models
    title: Instantiating a big model
@@ -72,25 +94,15 @@
  - local: debugging
    title: Debugging
  - local: notebooks
-    title: "🤗 Transformers Notebooks"
+    title: 🤗 Transformers Notebooks
  - local: community
    title: Community
  - local: contributing
    title: How to contribute to transformers?
  - local: add_new_model
-    title: "How to add a model to 🤗 Transformers?"
+    title: How to add a model to 🤗 Transformers?
  - local: add_new_pipeline
-    title: "How to add a pipeline to 🤗 Transformers?"
-  - local: perf_train_gpu_one
-    title: Training on one GPU
-  - local: perf_train_gpu_many
-    title: Training on many GPUs
-  - local: perf_train_cpu
-    title: Training on CPU
-  - local: perf_infer_cpu
-    title: Inference on CPU
-  - local: perf_hardware
-    title: Custom hardware for training
+    title: How to create a custom pipeline?
  - local: testing
    title: Testing
  - local: pr_checks
@@ -164,12 +176,12 @@
      title: BEiT
    - local: model_doc/bert
      title: BERT
-    - local: model_doc/bertweet
-      title: Bertweet
    - local: model_doc/bert-generation
      title: BertGeneration
    - local: model_doc/bert-japanese
      title: BertJapanese
+    - local: model_doc/bertweet
+      title: Bertweet
    - local: model_doc/big_bird
      title: BigBird
    - local: model_doc/bigbird_pegasus
@@ -188,12 +200,14 @@
      title: CamemBERT
    - local: model_doc/canine
      title: CANINE
-    - local: model_doc/convnext
-      title: ConvNeXT
    - local: model_doc/clip
      title: CLIP
+    - local: model_doc/codegen
+      title: CodeGen
    - local: model_doc/convbert
      title: ConvBERT
+    - local: model_doc/convnext
+      title: ConvNeXT
    - local: model_doc/cpm
      title: CPM
    - local: model_doc/ctrl
@@ -236,14 +250,20 @@
      title: FSMT
    - local: model_doc/funnel
      title: Funnel Transformer
-    - local: model_doc/gptj
-      title: GPT-J
    - local: model_doc/glpn
      title: GLPN
+    - local: model_doc/openai-gpt
+      title: GPT
    - local: model_doc/gpt_neo
      title: GPT Neo
    - local: model_doc/gpt_neox
      title: GPT NeoX
+    - local: model_doc/gptj
+      title: GPT-J
+    - local: model_doc/gpt2
+      title: GPT2
+    - local: model_doc/groupvit
+      title: GroupViT
    - local: model_doc/herbert
      title: HerBERT
    - local: model_doc/hubert
@@ -272,12 +292,12 @@
      title: LUKE
    - local: model_doc/lxmert
      title: LXMERT
+    - local: model_doc/m2m_100
+      title: M2M100
    - local: model_doc/marian
      title: MarianMT
    - local: model_doc/maskformer
      title: MaskFormer
-    - local: model_doc/m2m_100
-      title: M2M100
    - local: model_doc/mbart
      title: MBart and MBart-50
    - local: model_doc/mctct
@@ -290,18 +310,24 @@
      title: mLUKE
    - local: model_doc/mobilebert
      title: MobileBERT
+    - local: model_doc/mobilevit
+      title: MobileViT
    - local: model_doc/mpnet
      title: MPNet
    - local: model_doc/mt5
      title: MT5
+    - local: model_doc/mvp
+      title: MVP
+    - local: model_doc/nezha
+      title: NEZHA
+    - local: model_doc/nllb
+      title: NLLB
    - local: model_doc/nystromformer
      title: Nyströmformer
-    - local: model_doc/openai-gpt
-      title: OpenAI GPT
-    - local: model_doc/gpt2
-      title: OpenAI GPT2
    - local: model_doc/opt
      title: OPT
+    - local: model_doc/owlvit
+      title: OWL-ViT
    - local: model_doc/pegasus
      title: Pegasus
    - local: model_doc/perceiver
@@ -322,10 +348,10 @@
      title: REALM
    - local: model_doc/reformer
      title: Reformer
-    - local: model_doc/rembert
-      title: RemBERT
    - local: model_doc/regnet
      title: RegNet
+    - local: model_doc/rembert
+      title: RemBERT
    - local: model_doc/resnet
      title: ResNet
    - local: model_doc/retribert
@@ -366,6 +392,8 @@
      title: Transformer XL
    - local: model_doc/trocr
      title: TrOCR
+    - local: model_doc/ul2
+      title: UL2
    - local: model_doc/unispeech
      title: UniSpeech
    - local: model_doc/unispeech-sat
@@ -380,10 +408,10 @@
      title: Vision Text Dual Encoder
    - local: model_doc/vit
      title: Vision Transformer (ViT)
-    - local: model_doc/vit_mae
-      title: ViTMAE
    - local: model_doc/visual_bert
      title: VisualBERT
+    - local: model_doc/vit_mae
+      title: ViTMAE
    - local: model_doc/wav2vec2
      title: Wav2Vec2
    - local: model_doc/wav2vec2-conformer
@@ -404,10 +432,10 @@
      title: XLM-RoBERTa-XL
    - local: model_doc/xlnet
      title: XLNet
-    - local: model_doc/xlsr_wav2vec2
-      title: XLSR-Wav2Vec2
    - local: model_doc/xls_r
      title: XLS-R
+    - local: model_doc/xlsr_wav2vec2
+      title: XLSR-Wav2Vec2
    - local: model_doc/yolos
      title: YOLOS
    - local: model_doc/yoso
--- a/docs/source/en/add_new_pipeline.mdx
+++ b/docs/source/en/add_new_pipeline.mdx
@@ -9,7 +9,10 @@ Unless required by applicable law or agreed to in writing, software distributed
 an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
 -->

-# How to add a pipeline to 🤗 Transformers?
+# How to create a custom pipeline?
+
+In this guide, we will see how to create a custom pipeline and share it on the [Hub](hf.co/models) or add it to the
+Transformers library.

 First and foremost, you need to decide the raw entries the pipeline will be able to take. It can be strings, raw bytes,
 dictionaries or whatever seems to be the most likely desired input. Try to keep these inputs as pure Python as possible
@@ -99,7 +102,7 @@ def _sanitize_parameters(self, **kwargs):

    postprocess_kwargs = {}
    if "top_k" in kwargs:
-        preprocess_kwargs["top_k"] = kwargs["top_k"]
+        postprocess_kwargs["top_k"] = kwargs["top_k"]
    return preprocess_kwargs, {}, postprocess_kwargs
 ```

@@ -111,12 +114,123 @@ of arguments for ease of use (audio files, can be filenames, URLs or pure bytes)

 ## Adding it to the list of supported tasks

-Go to `src/transformers/pipelines/__init__.py` and fill in `SUPPORTED_TASKS` with your newly created pipeline.
-If possible it should provide a default model.
+To register your `new-task` to the list of supported tasks, you have to add it to the `PIPELINE_REGISTRY`:

-## Adding tests
+```python
+from transformers.pipelines import PIPELINE_REGISTRY

-Create a new file `tests/test_pipelines_MY_PIPELINE.py` with example with the other tests.
+PIPELINE_REGISTRY.register_pipeline(
+    "new-task",
+    pipeline_class=MyPipeline,
+    pt_model=AutoModelForSequenceClassification,
+)
+```
+
+You can specify a default model if you want, in which case it should come with a specific revision (which can be the name of a branch or a commit hash, here we took `"abcdef"`) as well was the type:
+
+```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
+)
+```
+
+## Share your pipeline on the Hub
+
+To share your custom pipeline on the Hub, you just have to save the custom code of your `Pipeline` subclass in a
+python file. For instance, let's say we want to use a custom pipeline for sentence pair classification like this:
+
+```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}
+```
+
+The implementation is framework agnostic, and will work for PyTorch and TensorFlow models. If we have saved this in
+a file named `pair_classification.py`, we can then import it and register it like this:
+
+```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,
+)
+```
+
+Once this is done, we can use it with a pretrained model. For instance `sgugger/finetuned-bert-mrpc` has been
+fine-tuned on the MRPC dataset, which classifies pairs of sentences as paraphrases or not.
+
+```py
+from transformers import pipeline
+
+classifier = pipeline("pair-classification", model="sgugger/finetuned-bert-mrpc")
+```
+
+Then we can share it on the Hub by using the `save_pretrained` method in a `Repository`:
+
+```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()
+```
+
+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
+`trust_remote_code=True`:
+
+```py
+from transformers import pipeline
+
+classifier = pipeline(model="{your_username}/test-dynamic-pipeline", trust_remote_code=True)
+```
+
+## 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`.
+
+Then you will need to add tests. Create a new file `tests/test_pipelines_MY_PIPELINE.py` with example with 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`.
--- a/docs/source/en/big_models.mdx
+++ b/docs/source/en/big_models.mdx
@@ -114,15 +114,6 @@ If you want to directly load such a sharded checkpoint inside a model without us

 ## Low memory loading

-Sharded checkpoints reduce the memory usage during step 2 of the worflow mentioned above, but when loadin a pretrained model, why keep the random weights in memory? The option `low_cpu_mem_usage` will destroy the weights of the randomly initialized model, then progressively load the weights inside, then perform a random initialization for potential missing weights (if you are loadding a model with a newly initialized head for a fine-tuning task for instance).
-
-It's very easy to use, just add `low_cpu_mem_usage=True` to your call to [`~PreTrainedModel.from_pretrained`]:
-
-```py
-from transformers import AutoModelForSequenceClas
-
-model = AutoModel.from_pretrained("bert-base-cased", low_cpu_mem_usage=True)
-```
-
-This can be used in conjunction with a sharded checkpoint.
+Sharded checkpoints reduce the memory usage during step 2 of the workflow mentioned above, but in order to use that model in a low memory setting, we recommend leveraging our tools based on the Accelerate library.

+Please read the following guide for more information: [Large model loading using Accelerate](./main_classes/model#large-model-loading)
--- a/docs/source/en/custom_models.mdx
+++ b/docs/source/en/custom_models.mdx
@@ -289,7 +289,7 @@ from huggingface_hub import notebook_login
 notebook_login()
 ```

-You can then push to to your own namespace (or an organization you are a member of) like this:
+You can then push to your own namespace (or an organization you are a member of) like this:

 ```py
 resnet50d.push_to_hub("custom-resnet50d")
--- a/docs/source/en/index.mdx
+++ b/docs/source/en/index.mdx
@@ -69,6 +69,7 @@ The library currently contains JAX, PyTorch and TensorFlow implementations, pret
 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.
 1. **[CLIP](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. **[CodeGen](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. **[ConvBERT](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](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. **[CPM](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.
@@ -97,6 +98,7 @@ The library currently contains JAX, PyTorch and TensorFlow implementations, pret
 1. **[GPT NeoX](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-2](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](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. **[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. **[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. **[ImageGPT](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.
@@ -120,10 +122,15 @@ The library currently contains JAX, PyTorch and TensorFlow implementations, pret
 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. **[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. **[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.
+1. **[MobileViT](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. **[MPNet](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. **[MT5](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. **[MVP](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. **[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. **[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. **[OPT](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](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](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. **[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. **[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.
@@ -154,6 +161,7 @@ The library currently contains JAX, PyTorch and TensorFlow implementations, pret
 1. **[Trajectory Transformer](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](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](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. **[UL2](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. **[UniSpeech](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](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. **[VAN](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.
@@ -200,6 +208,7 @@ Flax), PyTorch, and/or TensorFlow.
 |          CamemBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |           CANINE            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            CLIP             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|           CodeGen           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |          ConvBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |          ConvNeXT           |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            CTRL             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
@@ -210,7 +219,7 @@ Flax), PyTorch, and/or TensorFlow.
 |           DeBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |         DeBERTa-v2          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |    Decision Transformer     |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
-|            DeiT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            DeiT             |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            DETR             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         DistilBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |             DPR             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
@@ -226,6 +235,7 @@ Flax), PyTorch, and/or TensorFlow.
 |           GPT Neo           |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
 |          GPT NeoX           |       ❌       |       ✅       |       ✅        |         ❌         |      ❌      |
 |            GPT-J            |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
+|          GroupViT           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           Hubert            |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |           I-BERT            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |          ImageGPT           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
@@ -245,12 +255,16 @@ Flax), PyTorch, and/or TensorFlow.
 |            mBART            |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |        Megatron-BERT        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         MobileBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|          MobileViT          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            MPNet            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |             MT5             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|             MVP             |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|            Nezha            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |        Nyströmformer        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         OpenAI GPT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |        OpenAI GPT-2         |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |             OPT             |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
+|           OWL-ViT           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           Pegasus           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |          Perceiver          |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           PLBart            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
@@ -260,13 +274,13 @@ Flax), PyTorch, and/or TensorFlow.
 |             RAG             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            REALM            |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |          Reformer           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
-|           RegNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           RegNet            |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |           RemBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
-|           ResNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           ResNet            |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |          RetriBERT          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RoBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |          RoFormer           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
-|          SegFormer          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|          SegFormer          |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |             SEW             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            SEW-D            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |   Speech Encoder decoder    |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
--- a/docs/source/en/installation.mdx
+++ b/docs/source/en/installation.mdx
@@ -34,11 +34,16 @@ Start by creating a virtual environment in your project directory:
 python -m venv .env
 ```

-Activate the virtual environment:
+Activate the virtual environment. On Linux and MacOs:

 ```bash
 source .env/bin/activate
 ```
+Activate Virtual environment on Windows
+
+```bash
+.env/Scripts/activate
+```

 Now you're ready to install 🤗 Transformers with the following command:

--- a/docs/source/en/model_doc/bert.mdx
+++ b/docs/source/en/model_doc/bert.mdx
@@ -58,6 +58,10 @@ This model was contributed by [thomwolf](https://huggingface.co/thomwolf). The o

 [[autodoc]] BertTokenizerFast

+## TFBertTokenizer
+
+[[autodoc]] TFBertTokenizer
+
 ## Bert specific outputs

 [[autodoc]] models.bert.modeling_bert.BertForPreTrainingOutput
--- a/docs/source/en/model_doc/bloom.mdx
+++ b/docs/source/en/model_doc/bloom.mdx
@@ -23,7 +23,7 @@ Several smaller versions of the models have been trained on the same dataset. BL
 - [bloom-1b3](https://huggingface.co/bigscience/bloom-1b3)
 - [bloom-2b5](https://huggingface.co/bigscience/bloom-2b5)
 - [bloom-6b3](https://huggingface.co/bigscience/bloom-6b3)
- [bloom](https://huggingface.co/bigscience/bloom) (175B parameters)
+- [bloom](https://huggingface.co/bigscience/bloom) (176B parameters)


 ## BloomConfig
--- a/docs/source/en/model_doc/codegen.mdx
+++ b/docs/source/en/model_doc/codegen.mdx
@@ -0,0 +1,81 @@
+<!--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.
+-->
+
+# CodeGen
+
+## Overview
+
+The CodeGen model was proposed in [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, and Caiming Xiong.
+
+CodeGen is an autoregressive language model for program synthesis trained sequentially on [The Pile](https://pile.eleuther.ai/), BigQuery, and BigPython.
+
+The abstract from the paper is the following:
+
+*Program synthesis strives to generate a computer program as a solution to a given problem specification. We propose a conversational program synthesis approach via large language models, which addresses the challenges of searching over a vast program space and user intent specification faced in prior approaches. Our new approach casts the process of writing a specification and program as a multi-turn conversation between a user and a system. It treats program synthesis as a sequence prediction problem, in which the specification is expressed in natural language and the desired program is conditionally sampled. We train a family of large language models, called CodeGen, on natural language and programming language data. With weak supervision in the data and the scaling up of data size and model size, conversational capacities emerge from the simple autoregressive language modeling. To study the model behavior on conversational program synthesis, we develop a multi-turn programming benchmark (MTPB), where solving each problem requires multi-step synthesis via multi-turn conversation between the user and the model. Our findings show the emergence of conversational capabilities and the effectiveness of the proposed conversational program synthesis paradigm. In addition, our model CodeGen (with up to 16B parameters trained on TPU-v4) outperforms OpenAI's Codex on the HumanEval benchmark. We make the training library JaxFormer including checkpoints available as open source contribution: [this https URL](https://github.com/salesforce/codegen).* 
+
+This model was contributed by [Hiroaki Hayashi](https://huggingface.co/rooa).
+The original code can be found [here](https://github.com/salesforce/codegen).
+
+## Checkpoint Naming
+
+* CodeGen model [checkpoints](https://huggingface.co/models?other=codegen) are available on different pre-training data with variable sizes.
+* The format is: `Salesforce/codegen-{size}-{data}`, where
+  * `size`: `350M`, `2B`, `6B`, `16B`
+  * `data`: 
+    * `nl`: Pre-trained on the Pile
+    * `multi`: Initialized with `nl`, then further pre-trained on multiple programming languages data
+    * `mono`: Initialized with `multi`, then further pre-trained on Python data
+* For example, `Salesforce/codegen-350M-mono` offers a 350 million-parameter checkpoint pre-trained sequentially on the Pile, multiple programming languages, and Python.
+
+## How to use
+
+```python
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
+
+>>> checkpoint = "Salesforce/codegen-350M-mono"
+>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
+>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
+
+>>> text = "def hello_world():"
+
+>>> completion = model.generate(**tokenizer(text, return_tensors="pt"))
+
+>>> print(tokenizer.decode(completion[0]))
+def hello_world():
+    print("Hello World")
+
+hello_world()
+```
+
+## CodeGenConfig
+
+[[autodoc]] CodeGenConfig
+    - all
+
+## CodeGenTokenizer
+
+[[autodoc]] CodeGenTokenizer
+    - save_vocabulary
+
+## CodeGenTokenizerFast
+
+[[autodoc]] CodeGenTokenizerFast
+
+## CodeGenModel
+
+[[autodoc]] CodeGenModel
+    - forward
+
+## CodeGenForCausalLM
+
+[[autodoc]] CodeGenForCausalLM
+    - forward
--- a/docs/source/en/model_doc/deit.mdx
+++ b/docs/source/en/model_doc/deit.mdx
@@ -69,7 +69,7 @@ Tips:
  *facebook/deit-base-patch16-384*. Note that one should use [`DeiTFeatureExtractor`] in order to
  prepare images for the model.

-This model was contributed by [nielsr](https://huggingface.co/nielsr).
+This model was contributed by [nielsr](https://huggingface.co/nielsr). The TensorFlow version of this model was added by [amyeroberts](https://huggingface.co/amyeroberts).


 ## DeiTConfig
@@ -100,3 +100,23 @@ This model was contributed by [nielsr](https://huggingface.co/nielsr).

 [[autodoc]] DeiTForImageClassificationWithTeacher
    - forward
+
+## TFDeiTModel
+
+[[autodoc]] TFDeiTModel
+    - call
+
+## TFDeiTForMaskedImageModeling
+
+[[autodoc]] TFDeiTForMaskedImageModeling
+    - call
+
+## TFDeiTForImageClassification
+
+[[autodoc]] TFDeiTForImageClassification
+    - call
+
+## TFDeiTForImageClassificationWithTeacher
+
+[[autodoc]] TFDeiTForImageClassificationWithTeacher
+    - call
--- a/docs/source/en/model_doc/detr.mdx
+++ b/docs/source/en/model_doc/detr.mdx
@@ -113,6 +113,28 @@ Tips:
 - The size of the images will determine the amount of memory being used, and will thus determine the `batch_size`.
  It is advised to use a batch size of 2 per GPU. See [this Github thread](https://github.com/facebookresearch/detr/issues/150) for more info.

+There are three ways to instantiate a DETR model (depending on what you prefer):
+  
+Option 1: Instantiate DETR with pre-trained weights for entire model
+```py
+>>> from transformers import DetrForObjectDetection
+
+>>> model = DetrForObjectDetection.from_pretrained("facebook/resnet-50")
+```
+
+Option 2: Instantiate DETR with randomly initialized weights for Transformer, but pre-trained weights for backbone
+```py
+>>> from transformers import DetrConfig, DetrForObjectDetection
+
+>>> config = DetrConfig()
+>>> model = DetrForObjectDetection(config)
+```
+Option 3: Instantiate DETR with randomly initialized weights for backbone + Transformer
+```py
+>>> config = DetrConfig(use_pretrained_backbone=False)
+>>> model = DetrForObjectDetection(config)
+```
+
 As a summary, consider the following table:

 | Task | Object detection | Instance segmentation | Panoptic segmentation |
@@ -166,4 +188,4 @@ mean Average Precision (mAP) and Panoptic Quality (PQ). The latter objects are i
 ## DetrForSegmentation

 [[autodoc]] DetrForSegmentation
-    - forward
+    - forward
--- a/docs/source/en/model_doc/encoder-decoder.mdx
+++ b/docs/source/en/model_doc/encoder-decoder.mdx
@@ -12,6 +12,8 @@ specific language governing permissions and limitations under the License.

 # Encoder Decoder Models

+## Overview
+
 The [`EncoderDecoderModel`] can be used to initialize a sequence-to-sequence model with any
 pretrained autoencoding model as the encoder and any pretrained autoregressive model as the decoder.

@@ -25,15 +27,77 @@ any other models (see the examples for more information).
 An application of this architecture could be to leverage two pretrained [`BertModel`] as the encoder
 and decoder for a summarization model as was shown in: [Text Summarization with Pretrained Encoders](https://arxiv.org/abs/1908.08345) by Yang Liu and Mirella Lapata.

-The [`~TFEncoderDecoderModel.from_pretrained`] currently doesn't support initializing the model from a
+## Randomly initializing `EncoderDecoderModel` from model configurations.
+
+[`EncoderDecoderModel`] can be randomly initialized from an encoder and a decoder config. In the following example, we show how to do this using the default [`BertModel`] configuration for the encoder and the default [`BertForCausalLM`] configuration for the decoder.
+
+```python
+>>> from transformers import BertConfig, EncoderDecoderConfig, EncoderDecoderModel
+
+>>> config_encoder = BertConfig()
+>>> config_decoder = BertConfig()
+
+>>> config = EncoderDecoderConfig.from_encoder_decoder_configs(config_encoder, config_decoder)
+>>> model = EncoderDecoderModel(config=config)
+```
+
+## Initialising `EncoderDecoderModel` from a pretrained encoder and a pretrained decoder.
+
+[`EncoderDecoderModel`] can be initialized from a pretrained encoder checkpoint and a pretrained decoder checkpoint. Note that any pretrained auto-encoding model, *e.g.* BERT, can serve as the encoder and both pretrained auto-encoding models, *e.g.* BERT, pretrained causal language models, *e.g.* GPT2, as well as the pretrained decoder part of sequence-to-sequence models, *e.g.* decoder of BART, can be used as the decoder.
+Depending on which architecture you choose as the decoder, the cross-attention layers might be randomly initialized.
+Initializing [`EncoderDecoderModel`] from a pretrained encoder and decoder checkpoint requires the model to be fine-tuned on a downstream task, as has been shown in [the *Warm-starting-encoder-decoder blog post*](https://huggingface.co/blog/warm-starting-encoder-decoder).
+To do so, the `EncoderDecoderModel` class provides a [`EncoderDecoderModel.from_encoder_decoder_pretrained`] method.
+
+```python
+>>> from transformers import EncoderDecoderModel, BertTokenizer
+
+>>> tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+>>> model = EncoderDecoderModel.from_encoder_decoder_pretrained("bert-base-uncased", "bert-base-uncased")
+```
+
+## Loading an existing `EncoderDecoderModel` checkpoint and perform inference.
+
+To load fine-tuned checkpoints of the `EncoderDecoderModel` class, [`EncoderDecoderModel`] provides the `from_pretrained(...)` method just like any other model architecture in Transformers.
+
+To perform inference, one uses the [`generate`] method, which allows to autoregressively generate text. This method supports various forms of decoding, such as greedy, beam search and multinomial sampling.
+
+```python
+>>> from transformers import AutoTokenizer, EncoderDecoderModel
+
+>>> # load a fine-tuned seq2seq model and corresponding tokenizer
+>>> model = EncoderDecoderModel.from_pretrained("patrickvonplaten/bert2bert_cnn_daily_mail")
+>>> tokenizer = AutoTokenizer.from_pretrained("patrickvonplaten/bert2bert_cnn_daily_mail")
+
+>>> # let's perform inference on a long piece of text
+>>> ARTICLE_TO_SUMMARIZE = (
+...     "PG&E stated it scheduled the blackouts in response to forecasts for high winds "
+...     "amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were "
+...     "scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow."
+... )
+>>> input_ids = tokenizer(ARTICLE_TO_SUMMARIZE, return_tensors="pt").input_ids
+
+>>> # autoregressively generate summary (uses greedy decoding by default)
+>>> generated_ids = model.generate(input_ids)
+>>> generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+>>> print(generated_text)
+nearly 800 thousand customers were affected by the shutoffs. the aim is to reduce the risk of wildfires. nearly 800, 000 customers were expected to be affected by high winds amid dry conditions. pg & e said it scheduled the blackouts to last through at least midday tomorrow.
+```
+
+## Loading a PyTorch checkpoint into `TFEncoderDecoderModel`.
+
+[`TFEncoderDecoderModel.from_pretrained`] currently doesn't support initializing the model from a
 pytorch checkpoint. Passing `from_pt=True` to this method will throw an exception. If there are only pytorch
 checkpoints for a particular encoder-decoder model, a workaround is:

 ```python
 >>> # a workaround to load from pytorch checkpoint
+>>> from transformers import EncoderDecoderModel, TFEncoderDecoderModel
+
 >>> _model = EncoderDecoderModel.from_pretrained("patrickvonplaten/bert2bert-cnn_dailymail-fp16")
+
 >>> _model.encoder.save_pretrained("./encoder")
 >>> _model.decoder.save_pretrained("./decoder")
+
 >>> model = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
 ...     "./encoder", "./decoder", encoder_from_pt=True, decoder_from_pt=True
 ... )
@@ -41,6 +105,38 @@ checkpoints for a particular encoder-decoder model, a workaround is:
 >>> model.config = _model.config
 ```

+## Training
+
+Once the model is created, it can be fine-tuned similar to BART, T5 or any other encoder-decoder model.
+As you can see, only 2 inputs are required for the model in order to compute a loss: `input_ids` (which are the
+`input_ids` of the encoded input sequence) and `labels` (which are the `input_ids` of the encoded
+target sequence).
+
+```python
+>>> from transformers import BertTokenizer, EncoderDecoderModel
+
+>>> tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+>>> model = EncoderDecoderModel.from_encoder_decoder_pretrained("bert-base-uncased", "bert-base-uncased")
+
+>>> model.config.decoder_start_token_id = tokenizer.cls_token_id
+>>> model.config.pad_token_id = tokenizer.pad_token_id
+
+>>> input_ids = tokenizer(
+...     "The tower is 324 metres (1,063 ft) tall, about the same height as an 81-storey building, and the tallest structure in Paris. Its base is square, measuring 125 metres (410 ft) on each side.During its construction, the Eiffel Tower surpassed the Washington Monument to become the tallest man-made structure in the world, 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, it is now taller than the Chrysler Building by 5.2 metres (17 ft).Excluding transmitters, the Eiffel Tower is the second tallest free-standing structure in France after the Millau Viaduct.",
+...     return_tensors="pt",
+... ).input_ids
+
+>>> labels = tokenizer(
+...     "the eiffel tower surpassed the washington monument to become the tallest structure in the world. it was the first structure to reach a height of 300 metres in paris in 1930. it is now taller than the chrysler building by 5. 2 metres ( 17 ft ) and is the second tallest free - standing structure in paris.",
+...     return_tensors="pt",
+... ).input_ids
+
+>>> # the forward function automatically creates the correct decoder_input_ids
+>>> loss = model(input_ids=input_ids, labels=labels).loss
+```
+
+Detailed [colab](https://colab.research.google.com/drive/1WIk2bxglElfZewOHboPFNj8H44_VAyKE?usp=sharing#scrollTo=ZwQIEhKOrJpl) for training.
+
 This model was contributed by [thomwolf](https://github.com/thomwolf). This model's TensorFlow and Flax versions
 were contributed by [ydshieh](https://github.com/ydshieh).

--- a/docs/source/en/model_doc/groupvit.mdx
+++ b/docs/source/en/model_doc/groupvit.mdx
@@ -0,0 +1,61 @@
+<!--Copyright 2022 NVIDIA 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.
+-->
+
+# GroupViT
+
+## Overview
+
+The GroupViT model was proposed in [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.
+Inspired by [CLIP](clip), GroupViT is a vision-language model that can perform zero-shot semantic segmentation on any given vocabulary categories.
+
+The abstract from the paper is the following:
+
+*Grouping and recognition are important components of visual scene understanding, e.g., for object detection and semantic segmentation. With end-to-end deep learning systems, grouping of image regions usually happens implicitly via top-down supervision from pixel-level recognition labels. Instead, in this paper, we propose to bring back the grouping mechanism into deep networks, which allows semantic segments to emerge automatically with only text supervision. We propose a hierarchical Grouping Vision Transformer (GroupViT), which goes beyond the regular grid structure representation and learns to group image regions into progressively larger arbitrary-shaped segments. We train GroupViT jointly with a text encoder on a large-scale image-text dataset via contrastive losses. With only text supervision and without any pixel-level annotations, GroupViT learns to group together semantic regions and successfully transfers to the task of semantic segmentation in a zero-shot manner, i.e., without any further fine-tuning. It achieves a zero-shot accuracy of 52.3% mIoU on the PASCAL VOC 2012 and 22.4% mIoU on PASCAL Context datasets, and performs competitively to state-of-the-art transfer-learning methods requiring greater levels of supervision.*
+
+Tips:
+
+- You may specify `output_segmentation=True` in the forward of `GroupViTModel` to get the segmentation logits of input texts. 
+- The quickest way to get started with GroupViT is by checking the [example notebooks](https://github.com/xvjiarui/GroupViT/blob/main/demo/GroupViT_hf_inference_notebook.ipynb) (which showcase zero-shot segmentation inference). One can also check out the [HuggingFace Spaces demo](https://huggingface.co/spaces/xvjiarui/GroupViT) to play with GroupViT. 
+
+This model was contributed by [xvjiarui](https://huggingface.co/xvjiarui).
+The original code can be found [here](https://github.com/NVlabs/GroupViT).
+
+
+## GroupViTConfig
+
+[[autodoc]] GroupViTConfig
+    - from_text_vision_configs
+
+## GroupViTTextConfig
+
+[[autodoc]] GroupViTTextConfig
+
+## GroupViTVisionConfig
+
+[[autodoc]] GroupViTVisionConfig
+
+## GroupViTModel
+
+[[autodoc]] GroupViTModel
+    - forward
+    - get_text_features
+    - get_image_features
+
+## GroupViTTextModel
+
+[[autodoc]] GroupViTTextModel
+    - forward
+
+## GroupViTVisionModel
+
+[[autodoc]] GroupViTVisionModel
+    - forward
--- a/docs/source/en/model_doc/mobilevit.mdx
+++ b/docs/source/en/model_doc/mobilevit.mdx
@@ -0,0 +1,55 @@
+<!--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.
+-->
+
+# MobileViT
+
+## Overview
+
+The MobileViT model was proposed in [MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer](https://arxiv.org/abs/2110.02178) by Sachin Mehta and Mohammad Rastegari. MobileViT introduces a new layer that replaces local processing in convolutions with global processing using transformers. 
+
+The abstract from the paper is the following:
+
+*Light-weight convolutional neural networks (CNNs) are the de-facto for mobile vision tasks. Their spatial inductive biases allow them to learn representations with fewer parameters across different vision tasks. However, these networks are spatially local. To learn global representations, self-attention-based vision trans-formers (ViTs) have been adopted. Unlike CNNs, ViTs are heavy-weight. In this paper, we ask the following question: is it possible to combine the strengths of CNNs and ViTs to build a light-weight and low latency network for mobile vision tasks? Towards this end, we introduce MobileViT, a light-weight and general-purpose vision transformer for mobile devices. MobileViT presents a different perspective for the global processing of information with transformers, i.e., transformers as convolutions. Our results show that MobileViT significantly outperforms CNN- and ViT-based networks across different tasks and datasets. On the ImageNet-1k dataset, MobileViT achieves top-1 accuracy of 78.4% with about 6 million parameters, which is 3.2% and 6.2% more accurate than MobileNetv3 (CNN-based) and DeIT (ViT-based) for a similar number of parameters. On the MS-COCO object detection task, MobileViT is 5.7% more accurate than MobileNetv3 for a similar number of parameters.*
+
+Tips:
+
+- MobileViT is more like a CNN than a Transformer model. It does not work on sequence data but on batches of images. Unlike ViT, there are no embeddings. The backbone model outputs a feature map.
+- One can use [`MobileViTFeatureExtractor`] to prepare images for the model. Note that if you do your own preprocessing, the pretrained checkpoints expect images to be in BGR pixel order (not RGB).
+- The available image classification checkpoints are pre-trained on [ImageNet-1k](https://huggingface.co/datasets/imagenet-1k) (also referred to as ILSVRC 2012, a collection of 1.3 million images and 1,000 classes).
+- The segmentation model uses a [DeepLabV3](https://arxiv.org/abs/1706.05587) head. The available semantic segmentation checkpoints are pre-trained on [PASCAL VOC](http://host.robots.ox.ac.uk/pascal/VOC/). 
+
+This model was contributed by [matthijs](https://huggingface.co/Matthijs). The original code and weights can be found [here](https://github.com/apple/ml-cvnets).
+
+
+## MobileViTConfig
+
+[[autodoc]] MobileViTConfig
+
+## MobileViTFeatureExtractor
+
+[[autodoc]] MobileViTFeatureExtractor
+    - __call__
+
+## MobileViTModel
+
+[[autodoc]] MobileViTModel
+    - forward
+
+## MobileViTForImageClassification
+
+[[autodoc]] MobileViTForImageClassification
+    - forward
+
+## MobileViTForSemanticSegmentation
+
+[[autodoc]] MobileViTForSemanticSegmentation
+    - forward
--- a/docs/source/en/model_doc/mt5.mdx
+++ b/docs/source/en/model_doc/mt5.mdx
@@ -96,3 +96,7 @@ See [`T5TokenizerFast`] for all details.
 ## FlaxMT5ForConditionalGeneration

 [[autodoc]] FlaxMT5ForConditionalGeneration
+
+## FlaxMT5EncoderModel
+
+[[autodoc]] FlaxMT5EncoderModel
--- a/docs/source/en/model_doc/mvp.mdx
+++ b/docs/source/en/model_doc/mvp.mdx
@@ -0,0 +1,138 @@
+<!--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.
+-->
+
+# MVP
+
+## Overview
+
+The MVP model was proposed in [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.
+
+
+According to the abstract,
+
+- MVP follows a standard Transformer encoder-decoder architecture.
+- MVP is supervised pre-trained using labeled datasets.
+- MVP also has task-specific soft prompts to stimulate the model's capacity in performing a certain task.
+- MVP is specially designed for natural language generation and can be adapted to a wide range of generation tasks, including but not limited to summarization, data-to-text generation, open-ended dialogue system, story generation, question answering, question generation, task-oriented dialogue system, commonsense generation, paraphrase generation, text style transfer, and text simplification. Our model can also be adapted to natural language understanding tasks such as sequence classification and (extractive) question answering.
+
+Tips:
+- We have released a series of models [here](https://huggingface.co/models?filter=mvp), including MVP, MVP with task-specific prompts, and multi-task pre-trained variants.
+- If you want to use a model without prompts (standard Transformer), you can load it through `MvpForConditionalGeneration.from_pretrained('RUCAIBox/mvp')`.
+- If you want to use a model with task-specific prompts, such as summarization, you can load it through `MvpForConditionalGeneration.from_pretrained('RUCAIBox/mvp-summarization')`.
+- Our model supports lightweight prompt tuning following [Prefix-tuning](https://arxiv.org/abs/2101.00190) with method `set_lightweight_tuning()`.
+
+This model was contributed by [Tianyi Tang](https://huggingface.co/StevenTang). The detailed information and instructions can be found [here](https://github.com/RUCAIBox/MVP).
+
+## Examples
+For summarization, it is an example to use MVP and MVP with summarization-specific prompts.
+
+```python
+>>> from transformers import MvpTokenizer, MvpForConditionalGeneration
+
+>>> tokenizer = MvpTokenizer.from_pretrained("RUCAIBox/mvp")
+>>> model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp")
+>>> model_with_prompt = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp-summarization")
+
+>>> inputs = tokenizer(
+...     "Summarize: You may want to stick it to your boss and leave your job, but don't do it if these are your reasons.",
+...     return_tensors="pt",
+... )
+>>> generated_ids = model.generate(**inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+["Why You Shouldn't Quit Your Job"]
+
+>>> generated_ids = model_with_prompt.generate(**inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+["Don't do it if these are your reasons"]
+```
+
+For data-to-text generation, it is an example to use MVP and multi-task pre-trained variants.
+```python
+>>> from transformers import MvpTokenizerFast, MvpForConditionalGeneration
+
+>>> tokenizer = MvpTokenizerFast.from_pretrained("RUCAIBox/mvp")
+>>> model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp")
+>>> model_with_mtl = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mtl-data-to-text")
+
+>>> inputs = tokenizer(
+...     "Describe the following data: Iron Man | instance of | Superhero [SEP] Stan Lee | creator | Iron Man",
+...     return_tensors="pt",
+... )
+>>> generated_ids = model.generate(**inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+['Stan Lee created the character of Iron Man, a fictional superhero appearing in American comic']
+
+>>> generated_ids = model_with_mtl.generate(**inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+['Iron Man is a fictional superhero appearing in American comic books published by Marvel Comics.']
+```
+
+For lightweight tuning, *i.e.*, fixing the model and only tuning prompts, you can load MVP with randomly initialized prompts or with task-specific prompts. Our code also supports Prefix-tuning with BART following the [original paper](https://arxiv.org/abs/2101.00190).
+
+```python
+>>> from transformers import MvpForConditionalGeneration
+
+>>> model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp", use_prompt=True)
+>>> # the number of trainable parameters (full tuning)
+>>> sum(p.numel() for p in model.parameters() if p.requires_grad)
+468116832
+
+>>> # lightweight tuning with randomly initialized prompts
+>>> model.set_lightweight_tuning()
+>>> # the number of trainable parameters (lightweight tuning)
+>>> sum(p.numel() for p in model.parameters() if p.requires_grad)
+61823328
+
+>>> # lightweight tuning with task-specific prompts
+>>> model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mtl-data-to-text")
+>>> model.set_lightweight_tuning()
+>>> # original lightweight Prefix-tuning
+>>> model = MvpForConditionalGeneration.from_pretrained("facebook/bart-large", use_prompt=True)
+>>> model.set_lightweight_tuning()
+```
+
+## MvpConfig
+
+[[autodoc]] MvpConfig
+
+## MvpTokenizer
+
+[[autodoc]] MvpTokenizer
+
+## MvpTokenizerFast
+
+[[autodoc]] MvpTokenizerFast
+
+## MvpModel
+
+[[autodoc]] MvpModel
+    - forward
+
+## MvpForConditionalGeneration
+
+[[autodoc]] MvpForConditionalGeneration
+    - forward
+
+## MvpForSequenceClassification
+
+[[autodoc]] MvpForSequenceClassification
+    - forward
+
+## MvpForQuestionAnswering
+
+[[autodoc]] MvpForQuestionAnswering
+    - forward
+
+## MvpForCausalLM
+
+[[autodoc]] MvpForCausalLM
+    - forward
--- a/docs/source/en/model_doc/nezha.mdx
+++ b/docs/source/en/model_doc/nezha.mdx
@@ -0,0 +1,76 @@
+<!--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.
+-->
+
+# Nezha
+
+## Overview
+
+The Nezha model was proposed in [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei et al.
+
+The abstract from the paper is the following:
+
+*The pre-trained language models have achieved great successes in various natural language understanding (NLU) tasks
+due to its capacity to capture the deep contextualized information in text by pre-training on large-scale corpora.
+In this technical report, we present our practice of pre-training language models named NEZHA (NEural contextualiZed
+representation for CHinese lAnguage understanding) on Chinese corpora and finetuning for the Chinese NLU tasks. 
+The current version of NEZHA is based on BERT with a collection of proven improvements, which include Functional 
+Relative Positional Encoding as an effective positional encoding scheme, Whole Word Masking strategy,
+Mixed Precision Training and the LAMB Optimizer in training the models. The experimental results show that NEZHA
+achieves the state-of-the-art performances when finetuned on several representative Chinese tasks, including
+named entity recognition (People's Daily NER), sentence matching (LCQMC), Chinese sentiment classification (ChnSenti)
+and natural language inference (XNLI).*
+
+This model was contributed by [sijunhe](https://huggingface.co/sijunhe). The original code can be found [here](https://github.com/huawei-noah/Pretrained-Language-Model/tree/master/NEZHA-PyTorch).
+
+## NezhaConfig
+
+[[autodoc]] NezhaConfig
+
+## NezhaModel
+
+[[autodoc]] NezhaModel
+    - forward
+
+## NezhaForPreTraining
+
+[[autodoc]] NezhaForPreTraining
+    - forward
+
+## NezhaForMaskedLM
+
+[[autodoc]] NezhaForMaskedLM
+    - forward
+
+## NezhaForNextSentencePrediction
+
+[[autodoc]] NezhaForNextSentencePrediction
+    - forward
+
+## NezhaForSequenceClassification
+
+[[autodoc]] NezhaForSequenceClassification
+    - forward
+
+## NezhaForMultipleChoice
+
+[[autodoc]] NezhaForMultipleChoice
+    - forward
+
+## NezhaForTokenClassification
+
+[[autodoc]] NezhaForTokenClassification
+    - forward
+
+## NezhaForQuestionAnswering
+
+[[autodoc]] NezhaForQuestionAnswering
+    - forward
--- a/docs/source/en/model_doc/nllb.mdx
+++ b/docs/source/en/model_doc/nllb.mdx
@@ -0,0 +1,99 @@
+<!--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.
+-->
+
+# NLLB
+
+**DISCLAIMER:** If you see something strange, file a [Github Issue](https://github.com/huggingface/transformers/issues/new?assignees=&labels=bug&template=bug-report.yml) and assign
+@LysandreJik
+
+## Overview of NLLB
+
+The NLLB model was presented in [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by Marta R. Costa-jussà, James Cross, Onur Çelebi,
+Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula,
+Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, John Hoffman, Semarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau Tran, Pierre Andrews,
+Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, Philipp Koehn, Alexandre Mourachko, Christophe Ropers,
+Safiyyah Saleem, Holger Schwenk, and Jeff Wang.
+
+The abstract of the paper is the following:
+
+*Driven by the goal of eradicating language barriers on a global scale, machine translation has solidified itself as a key focus of artificial intelligence research today.
+However, such efforts have coalesced around a small subset of languages, leaving behind the vast majority of mostly low-resource languages. What does it take to break the
+200 language barrier while ensuring safe, high quality results, all while keeping ethical considerations in mind? In No Language Left Behind, we took on this challenge by
+first contextualizing the need for low-resource language translation support through exploratory interviews with native speakers. Then, we created datasets and models aimed
+at narrowing the performance gap between low and high-resource languages. More specifically, we developed a conditional compute model based on Sparsely Gated Mixture of
+Experts that is trained on data obtained with novel and effective data mining techniques tailored for low-resource languages. We propose multiple architectural and training
+improvements to counteract overfitting while training on thousands of tasks. Critically, we evaluated the performance of over 40,000 different translation directions using
+a human-translated benchmark, Flores-200, and combined human evaluation with a novel toxicity benchmark covering all languages in Flores-200 to assess translation safety.
+Our model achieves an improvement of 44% BLEU relative to the previous state-of-the-art, laying important groundwork towards realizing a universal translation system.*
+
+This implementation contains the dense models available on release. Let us know via a GitHub issue if you would like to see the MoE models as well.
+
+This model was contributed by [Lysandre](https://huggingface.co/lysandre). The authors' code can be found [here](https://github.com/facebookresearch/fairseq/tree/nllb).
+
+## Generating with NLLB
+
+While generating the target text set the `forced_bos_token_id` to the target language id. The following
+example shows how to translate English to French using the *facebook/nllb-200-distilled-600M* model.
+
+Note that we're using the BCP-47 code for French `fra_Latn`. See [here](https://github.com/facebookresearch/flores/blob/main/flores200/README.md#languages-in-flores-200)
+for the list of all BCP-47 in the Flores 200 dataset.
+
+```python
+>>> from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
+
+>>> tokenizer = AutoTokenizer.from_pretrained("facebook/nllb-200-distilled-600M")
+>>> model = AutoModelForSeq2SeqLM.from_pretrained("facebook/nllb-200-distilled-600M")
+
+>>> article = "UN Chief says there is no military solution in Syria"
+>>> inputs = tokenizer(article, return_tensors="pt")
+
+>>> translated_tokens = model.generate(
+...     **inputs, forced_bos_token_id=tokenizer.lang_code_to_id["fra_Latn"], max_length=30
+... )
+>>> tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
+Le chef de l'ONU dit qu'il n'y a pas de solution militaire en Syrie
+```
+
+### Generating from any other language than English
+
+English (`eng_Latn`) is set as the default language from which to translate. In order to specify that you'd like to translate from a different language,
+you should specify the BCP-47 code in the `src_lang` keyword argument of the tokenizer initialization.
+
+See example below for a translation from romanian to german:
+
+```py
+>>> from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
+
+>>> tokenizer = AutoTokenizer.from_pretrained(
+...     "facebook/nllb-200-distilled-600M", use_auth_token=True, src_lang="ron_Latn"
+... )
+>>> model = AutoModelForSeq2SeqLM.from_pretrained("facebook/nllb-200-distilled-600M", use_auth_token=True)
+
+>>> article = "Şeful ONU spune că nu există o soluţie militară în Siria"
+>>> inputs = tokenizer(article, return_tensors="pt")
+
+>>> translated_tokens = model.generate(
+...     **inputs, forced_bos_token_id=tokenizer.lang_code_to_id["deu_Latn"], max_length=30
+... )
+>>> tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
+UN-Chef sagt, es gibt keine militärische Lösung in Syrien
+```
+
+## NllbTokenizer
+
+[[autodoc]] NllbTokenizer
+    - as_target_tokenizer
+    - build_inputs_with_special_tokens
+
+## NllbTokenizerFast
+
+[[autodoc]] NllbTokenizerFast
--- a/docs/source/en/model_doc/opt.mdx
+++ b/docs/source/en/model_doc/opt.mdx
@@ -54,6 +54,11 @@ The original code can be found [here](https://github.com/facebookresearch/metase
 [[autodoc]] TFOPTForCausalLM
    - call

+## OPTForSequenceClassification
+
+[[autodoc]] OPTForSequenceClassification
+    - forward
+
 ## FlaxOPTModel

 [[autodoc]] FlaxOPTModel
--- a/docs/source/en/model_doc/owlvit.mdx
+++ b/docs/source/en/model_doc/owlvit.mdx
@@ -0,0 +1,108 @@
+<!--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.
+-->
+
+# OWL-ViT
+
+## Overview
+
+The OWL-ViT (short for Vision Transformer for Open-World Localization) was proposed in [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. OWL-ViT is an open-vocabulary object detection network trained on a variety of (image, text) pairs. It can be used to query an image with one or multiple text queries to search for and detect target objects described in text.
+
+The abstract from the paper is the following:
+
+*Combining simple architectures with large-scale pre-training has led to massive improvements in image classification. For object detection, pre-training and scaling approaches are less well established, especially in the long-tailed and open-vocabulary setting, where training data is relatively scarce. In this paper, we propose a strong recipe for transferring image-text models to open-vocabulary object detection. We use a standard Vision Transformer architecture with minimal modifications, contrastive image-text pre-training, and end-to-end detection fine-tuning. Our analysis of the scaling properties of this setup shows that increasing image-level pre-training and model size yield consistent improvements on the downstream detection task. We provide the adaptation strategies and regularizations needed to attain very strong performance on zero-shot text-conditioned and one-shot image-conditioned object detection. Code and models are available on GitHub.*
+
+## Usage
+
+OWL-ViT is a zero-shot text-conditioned object detection model. OWL-ViT uses [CLIP](clip) as its multi-modal backbone, with a ViT-like Transformer to get visual features and a causal language model to get the text features. To use CLIP for detection, OWL-ViT removes the final token pooling layer of the vision model and attaches a lightweight classification and box head to each transformer output token. Open-vocabulary classification is enabled by replacing the fixed classification layer weights with the class-name embeddings obtained from the text model. The authors first train CLIP from scratch and fine-tune it end-to-end with the classification and box heads on standard detection datasets using a bipartite matching loss. One or multiple text queries per image can be used to perform zero-shot text-conditioned object detection. 
+
+[`OwlViTFeatureExtractor`] can be used to resize (or rescale) and normalize images for the model and [`CLIPTokenizer`] is used to encode the text. [`OwlViTProcessor`] wraps [`OwlViTFeatureExtractor`] and [`CLIPTokenizer`] into a single instance to both encode the text and prepare the images. The following example shows how to perform object detection using [`OwlViTProcessor`] and [`OwlViTForObjectDetection`].
+
+
+```python
+>>> import requests
+>>> from PIL import Image
+>>> import torch
+
+>>> from transformers import OwlViTProcessor, OwlViTForObjectDetection
+
+>>> processor = OwlViTProcessor.from_pretrained("google/owlvit-base-patch32")
+>>> model = OwlViTForObjectDetection.from_pretrained("google/owlvit-base-patch32")
+
+>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+>>> image = Image.open(requests.get(url, stream=True).raw)
+>>> texts = [["a photo of a cat", "a photo of a dog"]]
+>>> inputs = processor(text=texts, images=image, return_tensors="pt")
+>>> outputs = model(**inputs)
+
+>>> # Target image sizes (height, width) to rescale box predictions [batch_size, 2]
+>>> target_sizes = torch.Tensor([image.size[::-1]])
+>>> # Convert outputs (bounding boxes and class logits) to COCO API
+>>> results = processor.post_process(outputs=outputs, target_sizes=target_sizes)
+
+>>> i = 0  # Retrieve predictions for the first image for the corresponding text queries
+>>> text = texts[i]
+>>> boxes, scores, labels = results[i]["boxes"], results[i]["scores"], results[i]["labels"]
+
+>>> score_threshold = 0.1
+>>> for box, score, label in zip(boxes, scores, labels):
+...     box = [round(i, 2) for i in box.tolist()]
+...     if score >= score_threshold:
+...         print(f"Detected {text[label]} with confidence {round(score.item(), 3)} at location {box}")
+Detected a photo of a cat with confidence 0.243 at location [1.42, 50.69, 308.58, 370.48]
+Detected a photo of a cat with confidence 0.298 at location [348.06, 20.56, 642.33, 372.61]
+```
+
+This model was contributed by [adirik](https://huggingface.co/adirik). The original code can be found [here](https://github.com/google-research/scenic/tree/main/scenic/projects/owl_vit).
+
+## OwlViTConfig
+
+[[autodoc]] OwlViTConfig
+    - from_text_vision_configs
+
+## OwlViTTextConfig
+
+[[autodoc]] OwlViTTextConfig
+
+## OwlViTVisionConfig
+
+[[autodoc]] OwlViTVisionConfig
+
+## OwlViTFeatureExtractor
+
+[[autodoc]] OwlViTFeatureExtractor
+    - __call__
+
+## OwlViTProcessor
+
+[[autodoc]] OwlViTProcessor
+
+## OwlViTModel
+
+[[autodoc]] OwlViTModel
+    - forward
+    - get_text_features
+    - get_image_features
+
+## OwlViTTextModel
+
+[[autodoc]] OwlViTTextModel
+    - forward
+
+## OwlViTVisionModel
+
+[[autodoc]] OwlViTVisionModel
+    - forward
+
+## OwlViTForObjectDetection
+
+[[autodoc]] OwlViTForObjectDetection
+    - forward
--- a/docs/source/en/model_doc/regnet.mdx
+++ b/docs/source/en/model_doc/regnet.mdx
@@ -27,7 +27,8 @@ Tips:
 - One can use [`AutoFeatureExtractor`] to prepare images for the model.
 - The huge 10B model from [Self-supervised Pretraining of Visual Features in the Wild](https://arxiv.org/abs/2103.01988), trained on one billion Instagram images, is available on the [hub](https://huggingface.co/facebook/regnet-y-10b-seer)

-This model was contributed by [Francesco](https://huggingface.co/Francesco).
+This model was contributed by [Francesco](https://huggingface.co/Francesco). The TensorFlow version of the model
+was contributed by [sayakpaul](https://huggingface.com/sayakpaul) and [ariG23498](https://huggingface.com/ariG23498).
 The original code can be found [here](https://github.com/facebookresearch/pycls).


@@ -45,4 +46,15 @@ The original code can be found [here](https://github.com/facebookresearch/pycls)
 ## RegNetForImageClassification

 [[autodoc]] RegNetForImageClassification
-    - forward
+    - forward
+
+## TFRegNetModel
+
+[[autodoc]] TFRegNetModel
+    - call
+
+
+## TFRegNetForImageClassification
+
+[[autodoc]] TFRegNetForImageClassification
+    - call
--- a/docs/source/en/model_doc/resnet.mdx
+++ b/docs/source/en/model_doc/resnet.mdx
@@ -31,7 +31,7 @@ The figure below illustrates the architecture of ResNet. Taken from the [origina

 <img width="600" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/resnet_architecture.png"/>

-This model was contributed by [Francesco](https://huggingface.co/Francesco). The original code can be found [here](https://github.com/KaimingHe/deep-residual-networks).
+This model was contributed by [Francesco](https://huggingface.co/Francesco). The TensorFlow version of this model was added by [amyeroberts](https://huggingface.co/amyeroberts). The original code can be found [here](https://github.com/KaimingHe/deep-residual-networks).

 ## ResNetConfig

@@ -47,4 +47,16 @@ This model was contributed by [Francesco](https://huggingface.co/Francesco). The
 ## ResNetForImageClassification

 [[autodoc]] ResNetForImageClassification
-    - forward
+    - forward
+
+
+## TFResNetModel
+
+[[autodoc]] TFResNetModel
+    - call
+
+
+## TFResNetForImageClassification
+
+[[autodoc]] TFResNetForImageClassification
+    - call
--- a/docs/source/en/model_doc/segformer.mdx
+++ b/docs/source/en/model_doc/segformer.mdx
@@ -36,13 +36,14 @@ The figure below illustrates the architecture of SegFormer. Taken from the [orig

 <img width="600" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/segformer_architecture.png"/>

-This model was contributed by [nielsr](https://huggingface.co/nielsr). The original code can be found [here](https://github.com/NVlabs/SegFormer).
+This model was contributed by [nielsr](https://huggingface.co/nielsr). The TensorFlow version 
+of the model was contributed by [sayakpaul](https://huggingface.co/sayakpaul). The original code can be found [here](https://github.com/NVlabs/SegFormer).

 Tips:

- SegFormer consists of a hierarchical Transformer encoder, and a lightweight all-MLP decode head.
+- SegFormer consists of a hierarchical Transformer encoder, and a lightweight all-MLP decoder head.
  [`SegformerModel`] is the hierarchical Transformer encoder (which in the paper is also referred to
-  as Mix Transformer or MiT). [`SegformerForSemanticSegmentation`] adds the all-MLP decode head on
+  as Mix Transformer or MiT). [`SegformerForSemanticSegmentation`] adds the all-MLP decoder head on
  top to perform semantic segmentation of images. In addition, there's
  [`SegformerForImageClassification`] which can be used to - you guessed it - classify images. The
  authors of SegFormer first pre-trained the Transformer encoder on ImageNet-1k to classify images. Next, they throw
@@ -51,6 +52,9 @@ Tips:
  found on the [hub](https://huggingface.co/models?other=segformer).
 - The quickest way to get started with SegFormer is by checking the [example notebooks](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/SegFormer) (which showcase both inference and
  fine-tuning on custom data). One can also check out the [blog post](https://huggingface.co/blog/fine-tune-segformer) introducing SegFormer and illustrating how it can be fine-tuned on custom data.
+- TensorFlow users should refer to [this repository](https://github.com/deep-diver/segformer-tf-transformers) that shows off-the-shelf inference and fine-tuning.
+- One can also check out [this interactive demo on Hugging Face Spaces](https://huggingface.co/spaces/chansung/segformer-tf-transformers)
+  to try out a SegFormer model on custom images.
 - SegFormer works on any input size, as it pads the input to be divisible by `config.patch_sizes`. 
 - One can use [`SegformerFeatureExtractor`] to prepare images and corresponding segmentation maps
  for the model. Note that this feature extractor is fairly basic and does not include all data augmentations used in
@@ -65,7 +69,8 @@ Tips:
  used by [`SegformerForSemanticSegmentation`]). However, other datasets use the 0 index as
  background class and include this class as part of all labels. In that case, `reduce_labels` should be set to
  `False`, as loss should also be computed for the background class.
- As most models, SegFormer comes in different sizes, the details of which can be found in the table below.
+- As most models, SegFormer comes in different sizes, the details of which can be found in the table below
+  (taken from Table 7 of the [original paper](https://arxiv.org/abs/2105.15203)).

 | **Model variant** | **Depths**    | **Hidden sizes**    | **Decoder hidden size** | **Params (M)** | **ImageNet-1k Top 1** |
 | :---------------: | ------------- | ------------------- | :---------------------: | :------------: | :-------------------: |
@@ -76,6 +81,10 @@ Tips:
 | MiT-b4            | [3, 8, 27, 3] | [64, 128, 320, 512] | 768                     | 62.6           | 83.6                  |
 | MiT-b5            | [3, 6, 40, 3] | [64, 128, 320, 512] | 768                     | 82.0           | 83.8                  |

+Note that MiT in the above table refers to the Mix Transformer encoder backbone introduced in SegFormer. For
+SegFormer's results on the segmentation datasets like ADE20k, refer to the [paper](https://arxiv.org/abs/2105.15203).
+
+
 ## SegformerConfig

 [[autodoc]] SegformerConfig
@@ -104,3 +113,23 @@ Tips:

 [[autodoc]] SegformerForSemanticSegmentation
    - forward
+
+## TFSegformerDecodeHead
+
+[[autodoc]] TFSegformerDecodeHead
+    - call
+
+## TFSegformerModel
+
+[[autodoc]] TFSegformerModel
+    - call 
+
+## TFSegformerForImageClassification
+
+[[autodoc]] TFSegformerForImageClassification
+    - call 
+
+## TFSegformerForSemanticSegmentation
+
+[[autodoc]] TFSegformerForSemanticSegmentation
+    - call 
--- a/docs/source/en/model_doc/speech-encoder-decoder.mdx
+++ b/docs/source/en/model_doc/speech-encoder-decoder.mdx
@@ -12,7 +12,7 @@ specific language governing permissions and limitations under the License.

 # Speech Encoder Decoder Models

-The [`SpeechEncoderDecoderModel`] can be used to initialize a speech-sequence-to-text-sequence model
+The [`SpeechEncoderDecoderModel`] can be used to initialize a speech-to-text model
 with any pretrained speech autoencoding model as the encoder (*e.g.* [Wav2Vec2](wav2vec2), [Hubert](hubert)) and any pretrained autoregressive model as the decoder.

 The effectiveness of initializing speech-sequence-to-text-sequence models with pretrained checkpoints for speech
@@ -20,9 +20,95 @@ recognition and speech translation has *e.g.* been shown in [Large-Scale Self- a
 Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli,
 Alexis Conneau.

-An example of how to use a [`SpeechEncoderDecoderModel`] for inference can be seen in
-[Speech2Text2](speech_to_text_2).
+An example of how to use a [`SpeechEncoderDecoderModel`] for inference can be seen in [Speech2Text2](speech_to_text_2).

+## Randomly initializing `SpeechEncoderDecoderModel` from model configurations.
+
+[`SpeechEncoderDecoderModel`] can be randomly initialized from an encoder and a decoder config. In the following example, we show how to do this using the default [`Wav2Vec2Model`] configuration for the encoder
+and the default [`BertForCausalLM`] configuration for the decoder.
+
+```python
+>>> from transformers import BertConfig, Wav2Vec2Config, SpeechEncoderDecoderConfig, SpeechEncoderDecoderModel
+
+>>> config_encoder = Wav2Vec2Config()
+>>> config_decoder = BertConfig()
+
+>>> config = SpeechEncoderDecoderConfig.from_encoder_decoder_configs(config_encoder, config_decoder)
+>>> model = SpeechEncoderDecoderModel(config=config)
+```
+
+## Initialising `SpeechEncoderDecoderModel` from a pretrained encoder and a pretrained decoder.
+
+[`SpeechEncoderDecoderModel`] can be initialized from a pretrained encoder checkpoint and a pretrained decoder checkpoint. Note that any pretrained Transformer-based speech model, *e.g.* [Wav2Vec2](wav2vec2), [Hubert](hubert) can serve as the encoder and both pretrained auto-encoding models, *e.g.* BERT, pretrained causal language models, *e.g.* GPT2, as well as the pretrained decoder part of sequence-to-sequence models, *e.g.* decoder of BART, can be used as the decoder.
+Depending on which architecture you choose as the decoder, the cross-attention layers might be randomly initialized.
+Initializing [`SpeechEncoderDecoderModel`] from a pretrained encoder and decoder checkpoint requires the model to be fine-tuned on a downstream task, as has been shown in [the *Warm-starting-encoder-decoder blog post*](https://huggingface.co/blog/warm-starting-encoder-decoder).
+To do so, the `SpeechEncoderDecoderModel` class provides a [`SpeechEncoderDecoderModel.from_encoder_decoder_pretrained`] method.
+
+```python
+>>> from transformers import SpeechEncoderDecoderModel
+
+>>> model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
+...     "facebook/hubert-large-ll60k", "bert-base-uncased"
+... )
+```
+
+## Loading an existing `SpeechEncoderDecoderModel` checkpoint and perform inference.
+
+To load fine-tuned checkpoints of the `SpeechEncoderDecoderModel` class, [`SpeechEncoderDecoderModel`] provides the `from_pretrained(...)` method just like any other model architecture in Transformers.
+
+To perform inference, one uses the [`generate`] method, which allows to autoregressively generate text. This method supports various forms of decoding, such as greedy, beam search and multinomial sampling.
+
+```python
+>>> from transformers import Wav2Vec2Processor, SpeechEncoderDecoderModel
+>>> from datasets import load_dataset
+>>> import torch
+
+>>> # load a fine-tuned speech translation model and corresponding processor
+>>> model = SpeechEncoderDecoderModel.from_pretrained("facebook/wav2vec2-xls-r-300m-en-to-15")
+>>> processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-xls-r-300m-en-to-15")
+
+>>> # let's perform inference on a piece of English speech (which we'll translate to German)
+>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+>>> input_values = processor(ds[0]["audio"]["array"], return_tensors="pt").input_values
+
+>>> # autoregressively generate transcription (uses greedy decoding by default)
+>>> generated_ids = model.generate(input_values)
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+>>> print(generated_text)
+Mr. Quilter ist der Apostel der Mittelschicht und wir freuen uns, sein Evangelium willkommen heißen zu können.
+```
+
+## Training
+
+Once the model is created, it can be fine-tuned similar to BART, T5 or any other encoder-decoder model on a dataset of (speech, text) pairs.
+As you can see, only 2 inputs are required for the model in order to compute a loss: `input_values` (which are the
+speech inputs) and `labels` (which are the `input_ids` of the encoded target sequence).
+
+```python
+>>> from transformers import Wav2Vec2Processor, SpeechEncoderDecoderModel
+>>> from datasets import load_dataset
+
+>>> processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
+>>> tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+>>> model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
+...     "facebook/wav2vec2-base-960h", "bert-base-uncased"
+... )
+
+>>> model.config.decoder_start_token_id = processor.tokenizer.cls_token_id
+>>> model.config.pad_token_id = processor.tokenizer.pad_token_id
+
+>>> # load a speech input
+>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+>>> input_values = processor(ds[0]["audio"]["array"], return_tensors="pt").input_values
+
+>>> # load its corresponding transcription
+>>> with processor.as_target_processor():
+...     labels = processor(ds[0]["text"], return_tensors="pt").input_ids
+
+>>> # the forward function automatically creates the correct decoder_input_ids
+>>> loss = model(input_values, labels=labels).loss
+>>> loss.backward()
+```

 ## SpeechEncoderDecoderConfig

--- a/docs/source/en/model_doc/speech_to_text.mdx
+++ b/docs/source/en/model_doc/speech_to_text.mdx
@@ -37,7 +37,7 @@ predicted token ids.

 The feature extractor depends on `torchaudio` and the tokenizer depends on `sentencepiece` so be sure to
 install those packages before running the examples. You could either install those as extra speech dependencies with
-`pip install transformers"[speech, sentencepiece]"` or install the packages seperately with `pip install torchaudio sentencepiece`. Also `torchaudio` requires the development version of the [libsndfile](http://www.mega-nerd.com/libsndfile/) package which can be installed via a system package manager. On Ubuntu it can
+`pip install transformers"[speech, sentencepiece]"` or install the packages separately with `pip install torchaudio sentencepiece`. Also `torchaudio` requires the development version of the [libsndfile](http://www.mega-nerd.com/libsndfile/) package which can be installed via a system package manager. On Ubuntu it can
 be installed as follows: `apt install libsndfile1-dev`


--- a/docs/source/en/model_doc/t5.mdx
+++ b/docs/source/en/model_doc/t5.mdx
@@ -371,3 +371,8 @@ T5 is supported by several example scripts, both for pre-training and fine-tunin
    - __call__
    - encode
    - decode
+
+## FlaxT5EncoderModel
+
+[[autodoc]] FlaxT5EncoderModel
+    - __call__
--- a/docs/source/en/model_doc/ul2.mdx
+++ b/docs/source/en/model_doc/ul2.mdx
@@ -0,0 +1,31 @@
+<!--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.
+-->
+
+# UL2
+
+## Overview
+
+The T5 model was presented in [Unifying Language Learning Paradigms](https://arxiv.org/pdf/2205.05131v1.pdf) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler.
+
+The abstract from the paper is the following:
+
+*Existing pre-trained models are generally geared towards a particular class of problems. To date, there seems to be still no consensus on what the right architecture and pre-training setup should be. This paper presents a unified framework for pre-training models that are universally effective across datasets and setups. We begin by disentangling architectural archetypes with pre-training objectives -- two concepts that are commonly conflated. Next, we present a generalized and unified perspective for self-supervision in NLP and show how different pre-training objectives can be cast as one another and how interpolating between different objectives can be effective. We then propose Mixture-of-Denoisers (MoD), a pre-training objective that combines diverse pre-training paradigms together. We furthermore introduce a notion of mode switching, wherein downstream fine-tuning is associated with specific pre-training schemes. We conduct extensive ablative experiments to compare multiple pre-training objectives and find that our method pushes the Pareto-frontier by outperforming T5 and/or GPT-like models across multiple diverse setups. Finally, by scaling our model up to 20B parameters, we achieve SOTA performance on 50 well-established supervised NLP tasks ranging from language generation (with automated and human evaluation), language understanding, text classification, question answering, commonsense reasoning, long text reasoning, structured knowledge grounding and information retrieval. Our model also achieve strong results at in-context learning, outperforming 175B GPT-3 on zero-shot SuperGLUE and tripling the performance of T5-XXL on one-shot summarization.*
+
+Tips:
+
+- UL2 is an encoder-decoder model pre-trained on a mixture of denoising functions as well as fine-tuned on an array of downstream tasks.
+- UL2 has the same architecture as [T5v1.1](t5v1.1) but uses the Gated-SiLU activation function instead of Gated-GELU.
+- The authors release checkpoints of one architecture which can be seen [here](https://huggingface.co/google/ul2)
+
+The original code can be found [here](https://github.com/google-research/google-research/tree/master/ul2).
+
+This model was contributed by [DanielHesslow](https://huggingface.co/Seledorn).
--- a/docs/source/en/model_doc/vilt.mdx
+++ b/docs/source/en/model_doc/vilt.mdx
@@ -87,3 +87,8 @@ This model was contributed by [nielsr](https://huggingface.co/nielsr). The origi

 [[autodoc]] ViltForImageAndTextRetrieval
    - forward
+
+## ViltForTokenClassification
+
+[[autodoc]] ViltForTokenClassification
+    - forward
--- a/docs/source/en/model_doc/vision-encoder-decoder.mdx
+++ b/docs/source/en/model_doc/vision-encoder-decoder.mdx
@@ -12,16 +12,136 @@ specific language governing permissions and limitations under the License.

 # Vision Encoder Decoder Models

-The [`VisionEncoderDecoderModel`] can be used to initialize an image-to-text-sequence model with any
-pretrained Transformer-based vision autoencoding model as the encoder (*e.g.* [ViT](vit), [BEiT](beit), [DeiT](deit), [Swin](swin))
+## Overview
+
+The [`VisionEncoderDecoderModel`] can be used to initialize an image-to-text model with any
+pretrained Transformer-based vision model as the encoder (*e.g.* [ViT](vit), [BEiT](beit), [DeiT](deit), [Swin](swin))
 and any pretrained language model as the decoder (*e.g.* [RoBERTa](roberta), [GPT2](gpt2), [BERT](bert), [DistilBERT](distilbert)).

 The effectiveness of initializing image-to-text-sequence models with pretrained checkpoints has been shown in (for
 example) [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.

-An example of how to use a [`VisionEncoderDecoderModel`] for inference can be seen in [TrOCR](trocr).
+After such a [`VisionEncoderDecoderModel`] has been trained/fine-tuned, it can be saved/loaded just like any other models (see the examples below
+for more information).

+An example application is image captioning, in which the encoder is used to encode the image, after which an autoregressive language model generates
+the caption. Another example is optical character recognition. Refer to [TrOCR](trocr), which is an instance of [`VisionEncoderDecoderModel`].
+
+## Randomly initializing `VisionEncoderDecoderModel` from model configurations.
+
+[`VisionEncoderDecoderModel`] can be randomly initialized from an encoder and a decoder config. In the following example, we show how to do this using the default [`ViTModel`] configuration for the encoder
+and the default [`BertForCausalLM`] configuration for the decoder.
+
+```python
+>>> from transformers import BertConfig, ViTConfig, VisionEncoderDecoderConfig, VisionEncoderDecoderModel
+
+>>> config_encoder = ViTConfig()
+>>> config_decoder = BertConfig()
+
+>>> config = VisionEncoderDecoderConfig.from_encoder_decoder_configs(config_encoder, config_decoder)
+>>> model = VisionEncoderDecoderModel(config=config)
+```
+
+## Initialising `VisionEncoderDecoderModel` from a pretrained encoder and a pretrained decoder.
+
+[`VisionEncoderDecoderModel`] can be initialized from a pretrained encoder checkpoint and a pretrained decoder checkpoint. Note that any pretrained Transformer-based vision model, *e.g.* [Swin](swin), can serve as the encoder and both pretrained auto-encoding models, *e.g.* BERT, pretrained causal language models, *e.g.* GPT2, as well as the pretrained decoder part of sequence-to-sequence models, *e.g.* decoder of BART, can be used as the decoder.
+Depending on which architecture you choose as the decoder, the cross-attention layers might be randomly initialized.
+Initializing [`VisionEncoderDecoderModel`] from a pretrained encoder and decoder checkpoint requires the model to be fine-tuned on a downstream task, as has been shown in [the *Warm-starting-encoder-decoder blog post*](https://huggingface.co/blog/warm-starting-encoder-decoder).
+To do so, the `VisionEncoderDecoderModel` class provides a [`VisionEncoderDecoderModel.from_encoder_decoder_pretrained`] method.
+
+```python
+>>> from transformers import VisionEncoderDecoderModel
+
+>>> model = VisionEncoderDecoderModel.from_encoder_decoder_pretrained(
+...     "microsoft/swin-base-patch4-window7-224-in22k", "bert-base-uncased"
+... )
+```
+
+## Loading an existing `VisionEncoderDecoderModel` checkpoint and perform inference.
+
+To load fine-tuned checkpoints of the `VisionEncoderDecoderModel` class, [`VisionEncoderDecoderModel`] provides the `from_pretrained(...)` method just like any other model architecture in Transformers.
+
+To perform inference, one uses the [`generate`] method, which allows to autoregressively generate text. This method supports various forms of decoding, such as greedy, beam search and multinomial sampling.
+
+```python
+>>> import requests
+>>> from PIL import Image
+
+>>> from transformers import GPT2TokenizerFast, ViTFeatureExtractor, VisionEncoderDecoderModel
+
+>>> # load a fine-tuned image captioning model and corresponding tokenizer and feature extractor
+>>> model = VisionEncoderDecoderModel.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
+>>> tokenizer = GPT2TokenizerFast.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
+>>> feature_extractor = ViTFeatureExtractor.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
+
+>>> # let's perform inference on an image
+>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+>>> image = Image.open(requests.get(url, stream=True).raw)
+>>> pixel_values = feature_extractor(image, return_tensors="pt").pixel_values
+
+>>> # autoregressively generate caption (uses greedy decoding by default)
+>>> generated_ids = model.generate(pixel_values)
+>>> generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+>>> print(generated_text)
+a cat laying on a blanket next to a cat laying on a bed
+```
+
+## Loading a PyTorch checkpoint into `TFVisionEncoderDecoderModel`.
+
+[`TFVisionEncoderDecoderModel.from_pretrained`] currently doesn't support initializing the model from a
+PyTorch checkpoint. Passing `from_pt=True` to this method will throw an exception. If there are only PyTorch
+checkpoints for a particular vision encoder-decoder model, a workaround is:
+
+```python
+>>> from transformers import VisionEncoderDecoderModel, TFVisionEncoderDecoderModel
+
+>>> _model = VisionEncoderDecoderModel.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
+
+>>> _model.encoder.save_pretrained("./encoder")
+>>> _model.decoder.save_pretrained("./decoder")
+
+>>> model = TFVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
+...     "./encoder", "./decoder", encoder_from_pt=True, decoder_from_pt=True
+... )
+>>> # This is only for copying some specific attributes of this particular model.
+>>> model.config = _model.config
+```
+
+## Training
+
+Once the model is created, it can be fine-tuned similar to BART, T5 or any other encoder-decoder model on a dataset of (image, text) pairs.
+As you can see, only 2 inputs are required for the model in order to compute a loss: `pixel_values` (which are the
+images) and `labels` (which are the `input_ids` of the encoded target sequence).
+
+```python
+>>> from transformers import ViTFeatureExtractor, BertTokenizer, VisionEncoderDecoderModel
+>>> from datasets import load_dataset
+
+>>> feature_extractor = ViTFeatureExtractor.from_pretrained("google/vit-base-patch16-224-in21k")
+>>> tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+>>> model = VisionEncoderDecoderModel.from_encoder_decoder_pretrained(
+...     "google/vit-base-patch16-224-in21k", "bert-base-uncased"
+... )
+
+>>> model.config.decoder_start_token_id = tokenizer.cls_token_id
+>>> model.config.pad_token_id = tokenizer.pad_token_id
+
+>>> dataset = load_dataset("huggingface/cats-image")
+>>> image = dataset["test"]["image"][0]
+>>> pixel_values = feature_extractor(image, return_tensors="pt").pixel_values
+
+>>> labels = tokenizer(
+...     "an image of two cats chilling on a couch",
+...     return_tensors="pt",
+... ).input_ids
+
+>>> # the forward function automatically creates the correct decoder_input_ids
+>>> loss = model(pixel_values=pixel_values, labels=labels).loss
+```
+
+This model was contributed by [nielsr](https://github.com/nielsrogge). This model's TensorFlow and Flax versions
+were contributed by [ydshieh](https://github.com/ydshieh).

 ## VisionEncoderDecoderConfig

--- a/docs/source/en/perf_infer_gpu_many.mdx
+++ b/docs/source/en/perf_infer_gpu_many.mdx
@@ -0,0 +1,14 @@
+<!--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
+-->
+
+# Efficient Inference on a Multiple GPUs
+
+This document will be completed soon with information on how to infer on a multiple GPUs. In the meantime you can check out [the guide for training on a single GPU](perf_train_gpu_one) and [the guide for inference on CPUs](perf_infer_cpu).
--- a/docs/source/en/perf_infer_gpu_one.mdx
+++ b/docs/source/en/perf_infer_gpu_one.mdx
@@ -0,0 +1,14 @@
+<!--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
+-->
+
+# Efficient Inference on a Single GPU
+
+This document will be completed soon with information on how to infer on a single GPU. In the meantime you can check out [the guide for training on a single GPU](perf_train_gpu_one) and [the guide for inference on CPUs](perf_infer_cpu).
--- a/docs/source/en/perf_infer_special.mdx
+++ b/docs/source/en/perf_infer_special.mdx
@@ -0,0 +1,14 @@
+<!--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
+-->
+
+# Inference on Specialized Hardware
+
+This document will be completed soon with information on how to infer on specialized hardware. In the meantime you can check out [the guide for inference on CPUs](perf_infer_cpu).
--- a/docs/source/en/perf_train_gpu_many.mdx
+++ b/docs/source/en/perf_train_gpu_many.mdx
@@ -13,6 +13,12 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o

 When training on a single GPU is too slow or the model weights don't fit in a single GPUs memory we use a mutli-GPU setup. Switching from a single GPU to multiple requires some form of parallelism as the work needs to be distributed. There are several techniques to achieve parallism such as data, tensor, or pipeline parallism. However, there is no one solution to fit them all and which settings works best depends on the hardware you are running on. While the main concepts most likely will apply to any other framework, this article is focused on PyTorch-based implementations.

+<Tip>
+
+ Note: Most of the strategies introduced in the [single GPU section](perf_train_gpu_one) (such as mixed precision training or gradient accumulation) are generic and apply to training models in general so make sure to have a look at it before diving into the following sections such as multi-GPU or CPU training.
+
+</Tip>
+
 We will first discuss in depth various 1D parallelism techniques and their pros and cons and then look at how they can be combined into 2D and 3D parallelism to enable an even faster training and to support even bigger models. Various other powerful alternative approaches will be presented.

 ## Concepts
--- a/docs/source/en/perf_train_gpu_one.mdx
+++ b/docs/source/en/perf_train_gpu_one.mdx
@@ -11,7 +11,7 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o

 # Efficient Training on a Single GPU

-This guide focuses on training large models efficiently on a single GPU. These approaches are still valid if you have access to a machine with multiple GPUs but you will also have access to additional methods outlined in the [multi-GPU section](perf_train_gpu_many). 
+This guide focuses on training large models efficiently on a single GPU. These approaches are still valid if you have access to a machine with multiple GPUs but you will also have access to additional methods outlined in the [multi-GPU section](perf_train_gpu_many).

 In this section we have a look at a few tricks to reduce the memory footprint and speed up training for large models and how they are integrated in the [`Trainer`] and [🤗 Accelerate](https://huggingface.co/docs/accelerate/). Each method can improve speed or memory usage which is summarized in the table below:

@@ -33,7 +33,7 @@ pip install transformers datasets accelerate nvidia-ml-py3

 The `nvidia-ml-py3` library allows us to monitor the memory usage of the models from within Python. You might be familiar with the `nvidia-smi` command in the terminal - this library allows to access the same information in Python directly.

-Then we create some dummy data. We create random token IDs between 100 and 30000 and binary labels for a classifier. In total we get 512 sequences each with length 512 and store them in a [`Dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=dataset#datasets.Dataset) with PyTorch format.
+Then we create some dummy data. We create random token IDs between 100 and 30000 and binary labels for a classifier. In total we get 512 sequences each with length 512 and store them in a [`~datasets.Dataset`] with PyTorch format.


 ```py
@@ -367,7 +367,7 @@ Samples/second: 10.09
 GPU memory occupied: 7275 MB.
 ```

-We can see that with these tweaks we use about half the GPU memory as at the beginning while also being slightly faster. 
+We can see that with these tweaks we use about half the GPU memory as at the beginning while also being slightly faster.

 ### BF16
 If you have access to a Ampere or newer hardware you can use bf16 for your training and evaluation. While bf16 has a worse precision than fp16, it has a much much bigger dynamic range. Therefore, if in the past you were experiencing overflow issues while training the model, bf16 will prevent this from happening most of the time. Remember that in fp16 the biggest number you can have is `65535` and any number above that will overflow. A bf16 number can be as large as `3.39e+38` (!) which is about the same as fp32 - because both have 8-bits used for the numerical range.
@@ -394,13 +394,13 @@ Like all cases with reduced precision this may or may not be satisfactory for yo

 If you're already using fp16 or bf16 mixed precision it may help with the throughput as well.

-You can enable this mode in the 🤗 Trainer with: 
+You can enable this mode in the 🤗 Trainer with:
 ```python
 TrainingArguments(tf32=True)
 ```
 By default the PyTorch default is used.

-Note: tf32 mode is internal to CUDA and can't be accessed directly via `tensor.to(dtype=torch.tf32)` as `torch.tf32` doesn't exit.
+Note: tf32 mode is internal to CUDA and can't be accessed directly via `tensor.to(dtype=torch.tf32)` as `torch.tf32` doesn't exist.

 Note: you need `torch>=1.7` to enjoy this feature.

@@ -654,7 +654,7 @@ https://github.com/huggingface/transformers/blob/master/src/transformers/trainer


 ## Choice of GPU
-Sometimes, even when applying all the above tweaks the throughput on a given GPU might still not be good enough. One easy solution is to change the type of GPU. For example switching from let's say a K80 (which you typically get on Google Colab) to a fancier GPU such as the V100 or A100. Although they are more expensive they are usually more cost effective than cheaper GPUs due to their larger memory and faster architecture. 
+Sometimes, even when applying all the above tweaks the throughput on a given GPU might still not be good enough. One easy solution is to change the type of GPU. For example switching from let's say a K80 (which you typically get on Google Colab) to a fancier GPU such as the V100 or A100. Although they are more expensive they are usually more cost effective than cheaper GPUs due to their larger memory and faster architecture.

 Now, let's take a step back and discuss what we should optimize for when scaling the training of large models.

@@ -718,3 +718,15 @@ For some applications, such as pretraining large language models, applying all t

 Another use case for training on many GPUs is if the model does not fit on a single GPU with all the mentioned tricks. There are still more methods we can apply although life starts to get a bit more complicated. This usually involves some form of pipeline or tensor parallelism where the model itself is distributed across several GPUs. One can also make use of DeepSpeed which implements some of these parallelism strategies along with some more optimization to reduce the memory footprint such as partitioning the optimizer states. You can read more about this in the ["Multi-GPU training" section](perf_train_gpu_many).

+## Inference with torchdynamo
+TorchDynamo is a new tracer that uses Python’s frame evaluation API to automatically create FX traces from existing PyTorch programs. After capturing the FX graph, different backends can be deployed to lower the graph to an optimized engine. One solution is using the [TensorRT](https://developer.nvidia.com/tensorrt) or NVFuser as backend. You can choose one option below for performance boost.
+```
+TrainingArguments(torchdynamo="eager")      #enable eager model GPU. No performance boost
+TrainingArguments(torchdynamo="nvfuser")    #enable nvfuser
+TrainingArguments(torchdynamo="fx2trt")     #enable tensorRT fp32
+TrainingArguments(torchdynamo="fx2trt-f16") #enable tensorRT fp16
+```
+This feature involves 3 different libraries. To install them, please follow the instructions below:  
+- [Torchdynamo installation](https://github.com/pytorch/torchdynamo#requirements-and-setup)  
+- [Functorch installation](https://github.com/pytorch/functorch#install)  
+- [Torch-TensorRT(FX) installation](https://github.com/pytorch/TensorRT/blob/master/docsrc/tutorials/getting_started_with_fx_path.rst#installation)  
--- a/docs/source/en/perf_train_special.mdx
+++ b/docs/source/en/perf_train_special.mdx
@@ -0,0 +1,20 @@
+<!--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
+-->
+
+# Training on Specialized Hardware
+
+<Tip>
+
+ Note: Most of the strategies introduced in the [single GPU section](perf_train_gpu_one) (such as mixed precision training or gradient accumulation) and [mutli-GPU section](perf_train_gpu_many) are generic and apply to training models in general so make sure to have a look at it before diving into this section.
+
+</Tip>
+
+This document will be completed soon with information on how to train on specialized hardware.
--- a/docs/source/en/perf_train_tpu.mdx
+++ b/docs/source/en/perf_train_tpu.mdx
@@ -0,0 +1,20 @@
+<!--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
+-->
+
+# Training on TPUs
+
+<Tip>
+
+ Note: Most of the strategies introduced in the [single GPU section](perf_train_gpu_one) (such as mixed precision training or gradient accumulation) and [mutli-GPU section](perf_train_gpu_many) are generic and apply to training models in general so make sure to have a look at it before diving into this section.
+
+</Tip>
+
+This document will be completed soon with information on how to train on TPUs.
--- a/docs/source/en/performance.mdx
+++ b/docs/source/en/performance.mdx
@@ -24,7 +24,13 @@ This document serves as an overview and entry point for the methods that could b

 ## Training

-Training transformer models efficiently requires an accelerator such as a GPU or TPU. The most common case is where you only have a single GPU.
+Training transformer models efficiently requires an accelerator such as a GPU or TPU. The most common case is where you only have a single GPU, but there is also a section about mutli-GPU and CPU training (with more coming soon).
+
+<Tip>
+
+ Note: Most of the strategies introduced in the single GPU sections (such as mixed precision training or gradient accumulation) are generic and apply to training models in general so make sure to have a look at it before diving into the following sections such as multi-GPU or CPU training.
+
+</Tip>

 ### Single GPU

@@ -46,11 +52,11 @@ In some cases training on a single GPU is still too slow or won't fit the large

 ### TPU

-_Coming soon_
+[_Coming soon_](perf_train_tpu)

 ### Specialized Hardware

-_Coming soon_
+[_Coming soon_](perf_train_special)

 ## Inference

@@ -58,19 +64,19 @@ Efficient inference with large models in a production environment can be as chal

 ### CPU

-[Go to CPU inference section](perf_infer_cpu.mdx)
+[Go to CPU inference section](perf_infer_cpu)

 ### Single GPU

-_Coming soon_
+[Go to single GPU inference section](perf_infer_gpu_one)

 ### Multi-GPU

-_Coming soon_
+[Go to multi-GPU inference section](perf_infer_gpu_many)

 ### Specialized Hardware

-_Coming soon_
+[_Coming soon_](perf_infer_special)

 ## Hardware

--- a/docs/source/en/preprocessing.mdx
+++ b/docs/source/en/preprocessing.mdx
@@ -244,7 +244,7 @@ For example, the [MInDS-14](https://huggingface.co/datasets/PolyAI/minds14) data
 'sampling_rate': 8000}
 ```

-1. Use 🤗 Datasets' [`cast_column`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.cast_column) method to upsample the sampling rate to 16kHz:
+1. Use 🤗 Datasets' [`~datasets.Dataset.cast_column`] method to upsample the sampling rate to 16kHz:

 ```py
 >>> dataset = dataset.cast_column("audio", Audio(sampling_rate=16_000))
--- a/docs/source/en/quicktour.mdx
+++ b/docs/source/en/quicktour.mdx
@@ -389,3 +389,42 @@ One particularly cool 🤗 Transformers feature is the ability to save a model a
 ```
 </tf>
 </frameworkcontent>
+
+## Custom model builds
+
+You can modify the model's configuration class to change how a model is built. The configuration specifies a model's attributes, such as the number of hidden layers or attention heads. You start from scratch when you initialize a model from a custom configuration class. The model attributes are randomly initialized, and you'll need to train the model before you can use it to get meaningful results.
+
+Start by importing [`AutoConfig`], and then load the pretrained model you want to modify. Within [`AutoConfig.from_pretrained`], you can specify the attribute you want to change, such as the number of attention heads:
+
+```py
+>>> from transformers import AutoConfig
+
+>>> my_config = AutoConfig.from_pretrained("distilbert-base-uncased", n_heads=12)
+```
+
+<frameworkcontent>
+<pt>
+Create a model from your custom configuration with [`AutoModel.from_config`]:
+
+```py
+>>> from transformers import AutoModel
+
+>>> my_model = AutoModel.from_config(my_config)
+```
+</pt>
+<tf>
+Create a model from your custom configuration with [`TFAutoModel.from_config`]:
+
+```py
+>>> from transformers import TFAutoModel
+
+>>> my_model = TFAutoModel.from_config(my_config)
+```
+</tf>
+</frameworkcontent>
+
+Take a look at the [Create a custom architecture](./create_a_model) guide for more information about building custom configurations.
+
+## What's next?
+
+Now that you've completed the 🤗 Transformers quick tour, check out our guides and learn how to do more specific things like writing a custom model, fine-tuning a model for a task, and how to train a model with a script. If you're interested in learning more about 🤗 Transformers core concepts, grab a cup of coffee and take a look at our Conceptual Guides!
--- a/docs/source/en/serialization.mdx
+++ b/docs/source/en/serialization.mdx
@@ -53,12 +53,17 @@ Ready-made configurations include the following architectures:
 - BigBird-Pegasus
 - Blenderbot
 - BlenderbotSmall
+- BLOOM
 - CamemBERT
+- CodeGen
 - ConvBERT
 - ConvNeXT
 - Data2VecText
 - Data2VecVision
+- DeBERTa
+- DeBERTa-v2
 - DeiT
+- DETR
 - DistilBERT
 - ELECTRA
 - FlauBERT
@@ -66,11 +71,14 @@ Ready-made configurations include the following architectures:
 - GPT-J
 - I-BERT
 - LayoutLM
+- LayoutLMv3
+- LeViT
 - LongT5
 - M2M100
 - Marian
 - mBART
 - MobileBERT
+- MobileViT
 - OpenAI GPT-2
 - Perceiver
 - PLBart
@@ -83,6 +91,7 @@ Ready-made configurations include the following architectures:
 - XLM
 - XLM-RoBERTa
 - XLM-RoBERTa-XL
+- YOLOS

 In the next two sections, we'll show you how to:

@@ -673,4 +682,4 @@ torch.neuron.trace(model, [token_tensor, segments_tensors])
 This change enables Neuron SDK to trace the model and optimize it to run in Inf1 instances.

 To learn more about AWS Neuron SDK features, tools, example tutorials and latest updates,
-please see the [AWS NeuronSDK documentation](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/index.html).
+please see the [AWS NeuronSDK documentation](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/index.html).
--- a/docs/source/en/tasks/asr.mdx
+++ b/docs/source/en/tasks/asr.mdx
@@ -117,7 +117,7 @@ The preprocessing function needs to:
 ...     return batch
 ```

-Use 🤗 Datasets [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) function to apply the preprocessing function over the entire dataset. You can speed up the map function by increasing the number of processes with `num_proc`. Remove the columns you don't need:
+Use 🤗 Datasets [`~datasets.Dataset.map`] function to apply the preprocessing function over the entire dataset. You can speed up the map function by increasing the number of processes with `num_proc`. Remove the columns you don't need:

 ```py
 >>> encoded_minds = minds.map(prepare_dataset, remove_columns=minds.column_names["train"], num_proc=4)
--- a/docs/source/en/tasks/audio_classification.mdx
+++ b/docs/source/en/tasks/audio_classification.mdx
@@ -129,7 +129,7 @@ The preprocessing function needs to:
 ...     return inputs
 ```

-Use 🤗 Datasets [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once. Remove the columns you don't need, and rename `intent_class` to `label` because that is what the model expects:
+Use 🤗 Datasets [`~datasets.Dataset.map`] function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once. Remove the columns you don't need, and rename `intent_class` to `label` because that is what the model expects:

 ```py
 >>> encoded_minds = minds.map(preprocess_function, remove_columns="audio", batched=True)
--- a/docs/source/en/tasks/image_classification.mdx
+++ b/docs/source/en/tasks/image_classification.mdx
@@ -95,7 +95,7 @@ Create a preprocessing function that will apply the transforms and return the `p
 ...     return examples
 ```

-Use 🤗 Dataset's [`with_transform`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?#datasets.Dataset.with_transform) method to apply the transforms over the entire dataset. The transforms are applied on-the-fly when you load an element of the dataset:
+Use 🤗 Dataset's [`~datasets.Dataset.with_transform`] method to apply the transforms over the entire dataset. The transforms are applied on-the-fly when you load an element of the dataset:

 ```py
 >>> food = food.with_transform(transforms)
--- a/docs/source/en/tasks/language_modeling.mdx
+++ b/docs/source/en/tasks/language_modeling.mdx
@@ -118,7 +118,7 @@ Here is how you can create a preprocessing function to convert the list to a str
 ...     return tokenizer([" ".join(x) for x in examples["answers.text"]], truncation=True)
 ```

-Use 🤗 Datasets [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once and increasing the number of processes with `num_proc`. Remove the columns you don't need:
+Use 🤗 Datasets [`~datasets.Dataset.map`] function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once and increasing the number of processes with `num_proc`. Remove the columns you don't need:

 ```py
 >>> tokenized_eli5 = eli5.map(
@@ -141,6 +141,7 @@ Now you need a second preprocessing function to capture text truncated from any
 >>> def group_texts(examples):
 ...     concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
 ...     total_length = len(concatenated_examples[list(examples.keys())[0]])
+...     total_length = (total_length // block_size) * block_size
 ...     result = {
 ...         k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
 ...         for k, t in concatenated_examples.items()
@@ -244,7 +245,7 @@ At this point, only three steps remain:
 ```
 </pt>
 <tf>
-To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`to_tf_dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.to_tf_dataset). Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:
+To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`~datasets.Dataset.to_tf_dataset`]. Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:

 ```py
 >>> tf_train_set = lm_dataset["train"].to_tf_dataset(
@@ -351,7 +352,7 @@ At this point, only three steps remain:
 ```
 </pt>
 <tf>
-To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`to_tf_dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.to_tf_dataset). Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:
+To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`~datasets.Dataset.to_tf_dataset`]. Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:

 ```py
 >>> tf_train_set = lm_dataset["train"].to_tf_dataset(
--- a/docs/source/en/tasks/multiple_choice.mdx
+++ b/docs/source/en/tasks/multiple_choice.mdx
@@ -79,7 +79,7 @@ The preprocessing function needs to do:
 ...     return {k: [v[i : i + 4] for i in range(0, len(v), 4)] for k, v in tokenized_examples.items()}
 ```

-Use 🤗 Datasets [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:
+Use 🤗 Datasets [`~datasets.Dataset.map`] function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:

 ```py
 tokenized_swag = swag.map(preprocess_function, batched=True)
@@ -224,7 +224,7 @@ At this point, only three steps remain:
 ```
 </pt>
 <tf>
-To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`to_tf_dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.to_tf_dataset). Specify inputs in `columns`, targets in `label_cols`, whether to shuffle the dataset order, batch size, and the data collator:
+To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`~datasets.Dataset.to_tf_dataset`]. Specify inputs in `columns`, targets in `label_cols`, whether to shuffle the dataset order, batch size, and the data collator:

 ```py
 >>> data_collator = DataCollatorForMultipleChoice(tokenizer=tokenizer)
--- a/docs/source/en/tasks/question_answering.mdx
+++ b/docs/source/en/tasks/question_answering.mdx
@@ -126,7 +126,7 @@ Here is how you can create a function to truncate and map the start and end toke
 ...     return inputs
 ```

-Use 🤗 Datasets [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once. Remove the columns you don't need:
+Use 🤗 Datasets [`~datasets.Dataset.map`] function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once. Remove the columns you don't need:

 ```py
 >>> tokenized_squad = squad.map(preprocess_function, batched=True, remove_columns=squad["train"].column_names)
@@ -199,7 +199,7 @@ At this point, only three steps remain:
 ```
 </pt>
 <tf>
-To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`to_tf_dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.to_tf_dataset). Specify inputs and the start and end positions of an answer in `columns`, whether to shuffle the dataset order, batch size, and the data collator:
+To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`~datasets.Dataset.to_tf_dataset`]. Specify inputs and the start and end positions of an answer in `columns`, whether to shuffle the dataset order, batch size, and the data collator:

 ```py
 >>> tf_train_set = tokenized_squad["train"].to_tf_dataset(
--- a/docs/source/en/tasks/sequence_classification.mdx
+++ b/docs/source/en/tasks/sequence_classification.mdx
@@ -66,7 +66,7 @@ Create a preprocessing function to tokenize `text` and truncate sequences to be
 ...     return tokenizer(examples["text"], truncation=True)
 ```

-Use 🤗 Datasets [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:
+Use 🤗 Datasets [`~datasets.Dataset.map`] function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:

 ```py
 tokenized_imdb = imdb.map(preprocess_function, batched=True)
@@ -144,7 +144,7 @@ At this point, only three steps remain:
 </Tip>
 </pt>
 <tf>
-To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`to_tf_dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.to_tf_dataset). Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:
+To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`~datasets.Dataset.to_tf_dataset`]. Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:

 ```py
 >>> tf_train_set = tokenized_imdb["train"].to_tf_dataset(
--- a/docs/source/en/tasks/summarization.mdx
+++ b/docs/source/en/tasks/summarization.mdx
@@ -85,7 +85,7 @@ The preprocessing function needs to:
 ...     return model_inputs
 ```

-Use 🤗 Datasets [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:
+Use 🤗 Datasets [`~datasets.Dataset.map`] function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:

 ```py
 >>> tokenized_billsum = billsum.map(preprocess_function, batched=True)
@@ -160,7 +160,7 @@ At this point, only three steps remain:
 ```
 </pt>
 <tf>
-To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`to_tf_dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.to_tf_dataset). Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:
+To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`~datasets.Dataset.to_tf_dataset`]. Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:

 ```py
 >>> tf_train_set = tokenized_billsum["train"].to_tf_dataset(
--- a/docs/source/en/tasks/token_classification.mdx
+++ b/docs/source/en/tasks/token_classification.mdx
@@ -126,7 +126,7 @@ Here is how you can create a function to realign the tokens and labels, and trun
 ...     return tokenized_inputs
 ```

-Use 🤗 Datasets [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) function to tokenize and align the labels over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:
+Use 🤗 Datasets [`~datasets.Dataset.map`] function to tokenize and align the labels over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:

 ```py
 >>> tokenized_wnut = wnut.map(tokenize_and_align_labels, batched=True)
@@ -199,7 +199,7 @@ At this point, only three steps remain:
 ```
 </pt>
 <tf>
-To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`to_tf_dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.to_tf_dataset). Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:
+To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`~datasets.Dataset.to_tf_dataset`]. Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:

 ```py
 >>> tf_train_set = tokenized_wnut["train"].to_tf_dataset(
--- a/docs/source/en/tasks/translation.mdx
+++ b/docs/source/en/tasks/translation.mdx
@@ -87,16 +87,38 @@ The preprocessing function needs to:
 ...     return model_inputs
 ```

-Use 🤗 Datasets [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:
+Use 🤗 Datasets [`~datasets.Dataset.map`] function to apply the preprocessing function over the entire dataset. You can speed up the `map` function by setting `batched=True` to process multiple elements of the dataset at once:

 ```py
 >>> tokenized_books = books.map(preprocess_function, batched=True)
 ```

+<frameworkcontent>
+<pt>
+Load T5 with [`AutoModelForSeq2SeqLM`]:
+
+```py
+>>> from transformers import AutoModelForSeq2SeqLM
+
+>>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-small")
+```
+</pt>
+<tf>
+Load T5 with [`TFAutoModelForSeq2SeqLM`]:
+
+```py
+>>> from transformers import TFAutoModelForSeq2SeqLM
+
+>>> model = TFAutoModelForSeq2SeqLM.from_pretrained("t5-small")
+```
+</tf>
+</frameworkcontent>
+
 Use [`DataCollatorForSeq2Seq`] to create a batch of examples. It will also *dynamically pad* your text and labels to the length of the longest element in its batch, so they are a uniform length. While it is possible to pad your text in the `tokenizer` function by setting `padding=True`, dynamic padding is more efficient.

 <frameworkcontent>
 <pt>
+
 ```py
 >>> from transformers import DataCollatorForSeq2Seq

@@ -104,6 +126,7 @@ Use [`DataCollatorForSeq2Seq`] to create a batch of examples. It will also *dyna
 ```
 </pt>
 <tf>
+
 ```py
 >>> from transformers import DataCollatorForSeq2Seq

@@ -116,13 +139,6 @@ Use [`DataCollatorForSeq2Seq`] to create a batch of examples. It will also *dyna

 <frameworkcontent>
 <pt>
-Load T5 with [`AutoModelForSeq2SeqLM`]:
-
-```py
->>> from transformers import AutoModelForSeq2SeqLM, Seq2SeqTrainingArguments, Seq2SeqTrainer
-
->>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-small")
-```

 <Tip>

@@ -137,6 +153,8 @@ At this point, only three steps remain:
 3. Call [`~Trainer.train`] to fine-tune your model.

 ```py
+>>> from transformers import Seq2SeqTrainingArguments, Seq2SeqTrainer
+
 >>> training_args = Seq2SeqTrainingArguments(
 ...     output_dir="./results",
 ...     evaluation_strategy="epoch",
@@ -162,7 +180,7 @@ At this point, only three steps remain:
 ```
 </pt>
 <tf>
-To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`to_tf_dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.to_tf_dataset). Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:
+To fine-tune a model in TensorFlow, start by converting your datasets to the `tf.data.Dataset` format with [`~datasets.Dataset.to_tf_dataset`]. Specify inputs and labels in `columns`, whether to shuffle the dataset order, batch size, and the data collator:

 ```py
 >>> tf_train_set = tokenized_books["train"].to_tf_dataset(
@@ -194,14 +212,6 @@ Set up an optimizer function, learning rate schedule, and some training hyperpar
 >>> optimizer = AdamWeightDecay(learning_rate=2e-5, weight_decay_rate=0.01)
 ```

-Load T5 with [`TFAutoModelForSeq2SeqLM`]:
-
-```py
->>> from transformers import TFAutoModelForSeq2SeqLM
-
->>> model = TFAutoModelForSeq2SeqLM.from_pretrained("t5-small")
-```
-
 Configure the model for training with [`compile`](https://keras.io/api/models/model_training_apis/#compile-method):

 ```py
@@ -222,4 +232,4 @@ For a more in-depth example of how to fine-tune a model for translation, take a
 [PyTorch notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/translation.ipynb)
 or [TensorFlow notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/translation-tf.ipynb).

-</Tip>
+</Tip>
--- a/docs/source/en/testing.mdx
+++ b/docs/source/en/testing.mdx
@@ -1226,7 +1226,7 @@ This whole process would have been much easier if we only could set something li
 experimental step, and let it fail without impacting the overall status of PRs. But as mentioned earlier CircleCI and
 Github Actions don't support it at the moment.

-You can vote for this feature and see where it is at at these CI-specific threads:
+You can vote for this feature and see where it is at these CI-specific threads:

 - [Github Actions:](https://github.com/actions/toolkit/issues/399)
 - [CircleCI:](https://ideas.circleci.com/ideas/CCI-I-344)
--- a/docs/source/en/training.mdx
+++ b/docs/source/en/training.mdx
@@ -169,7 +169,7 @@ The [`DefaultDataCollator`] assembles tensors into a batch for the model to trai

 </Tip>

-Next, convert the tokenized datasets to TensorFlow datasets with the [`to_tf_dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.to_tf_dataset) method. Specify your inputs in `columns`, and your label in `label_cols`:
+Next, convert the tokenized datasets to TensorFlow datasets with the [`~datasets.Dataset.to_tf_dataset`] method. Specify your inputs in `columns`, and your label in `label_cols`:

 ```py
 >>> tf_train_dataset = small_train_dataset.to_tf_dataset(
--- a/docs/source/es/_toctree.yml
+++ b/docs/source/es/_toctree.yml
@@ -25,6 +25,8 @@
    title: Usa tokenizadores de 🤗 Tokenizers
  - local: create_a_model
    title: Crea una arquitectura personalizada
+  - local: custom_models
+    title: Compartir modelos personalizados
  - sections:
    - local: tasks/language_modeling
      title: Modelado de lenguaje
--- a/docs/source/es/custom_models.mdx
+++ b/docs/source/es/custom_models.mdx
@@ -0,0 +1,351 @@
+<!--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.
+-->
+
+# Compartir modelos personalizados
+
+La biblioteca 🤗 Transformers está diseñada para ser fácilmente ampliable. Cada modelo está completamente codificado 
+sin abstracción en una subcarpeta determinada del repositorio, por lo que puedes copiar fácilmente un archivo del modelo 
+y ajustarlo según tus necesidades.
+
+Si estás escribiendo un modelo completamente nuevo, podría ser más fácil comenzar desde cero. En este tutorial, te mostraremos 
+cómo escribir un modelo personalizado y su configuración para que pueda usarse dentro de Transformers, y cómo puedes compartirlo 
+con la comunidad (con el código en el que se basa) para que cualquiera pueda usarlo, incluso si no está presente en la biblioteca 
+🤗 Transformers.
+
+Ilustraremos todo esto con un modelo ResNet, envolviendo la clase ResNet de la [biblioteca timm](https://github.com/rwightman/pytorch-image-models/tree/master/timm) en un [`PreTrainedModel`].
+
+## Escribir una configuración personalizada
+
+Antes de adentrarnos en el modelo, primero escribamos su configuración. La configuración de un modelo es un objeto que
+contendrá toda la información necesaria para construir el modelo. Como veremos en la siguiente sección, el modelo solo puede
+tomar un `config` para ser inicializado, por lo que realmente necesitamos que ese objeto esté lo más completo posible.
+
+En nuestro ejemplo, tomaremos un par de argumentos de la clase ResNet que tal vez queramos modificar. Las diferentes 
+configuraciones nos darán los diferentes tipos de ResNet que son posibles. Luego simplemente almacenamos esos argumentos 
+después de verificar la validez de algunos de ellos.
+
+```python
+from transformers import PretrainedConfig
+from typing import List
+
+
+class ResnetConfig(PretrainedConfig):
+    model_type = "resnet"
+
+    def __init__(
+        self,
+        block_type="bottleneck",
+        layers: List[int] = [3, 4, 6, 3],
+        num_classes: int = 1000,
+        input_channels: int = 3,
+        cardinality: int = 1,
+        base_width: int = 64,
+        stem_width: int = 64,
+        stem_type: str = "",
+        avg_down: bool = False,
+        **kwargs,
+    ):
+        if block_type not in ["basic", "bottleneck"]:
+            raise ValueError(f"`block` must be 'basic' or bottleneck', got {block}.")
+        if stem_type not in ["", "deep", "deep-tiered"]:
+            raise ValueError(f"`stem_type` must be '', 'deep' or 'deep-tiered', got {block}.")
+
+        self.block_type = block_type
+        self.layers = layers
+        self.num_classes = num_classes
+        self.input_channels = input_channels
+        self.cardinality = cardinality
+        self.base_width = base_width
+        self.stem_width = stem_width
+        self.stem_type = stem_type
+        self.avg_down = avg_down
+        super().__init__(**kwargs)
+```
+
+Las tres cosas importantes que debes recordar al escribir tu propia configuración son las siguientes:
+- tienes que heredar de `PretrainedConfig`,
+- el `__init__` de tu `PretrainedConfig` debe aceptar cualquier `kwargs`,
+- esos `kwargs` deben pasarse a la superclase `__init__`.
+
+La herencia es para asegurarte de obtener toda la funcionalidad de la biblioteca 🤗 Transformers, mientras que las otras dos 
+restricciones provienen del hecho de que una `PretrainedConfig` tiene más campos que los que estás configurando. Al recargar una 
+`config` con el método `from_pretrained`, esos campos deben ser aceptados por tu `config` y luego enviados a la superclase.
+
+Definir un `model_type` para tu configuración (en este caso `model_type="resnet"`) no es obligatorio, a menos que quieras
+registrar tu modelo con las clases automáticas (ver la última sección).
+
+Una vez hecho esto, puedes crear y guardar fácilmente tu configuración como lo harías con cualquier otra configuración de un 
+modelo de la biblioteca. Así es como podemos crear una configuración resnet50d y guardarla:
+
+```py
+resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
+resnet50d_config.save_pretrained("custom-resnet")
+```
+
+Esto guardará un archivo llamado `config.json` dentro de la carpeta `custom-resnet`. Luego puedes volver a cargar tu configuración 
+con el método `from_pretrained`:
+
+```py
+resnet50d_config = ResnetConfig.from_pretrained("custom-resnet")
+```
+
+También puedes usar cualquier otro método de la clase [`PretrainedConfig`], como [`~PretrainedConfig.push_to_hub`], para cargar 
+directamente tu configuración en el Hub.
+
+## Escribir un modelo personalizado
+
+Ahora que tenemos nuestra configuración de ResNet, podemos seguir escribiendo el modelo. En realidad escribiremos dos: una que
+extrae las características ocultas de un grupo de imágenes (como [`BertModel`]) y una que es adecuada para clasificación de
+imagenes (como [`BertForSequenceClassification`]).
+
+Como mencionamos antes, solo escribiremos un envoltura (_wrapper_) libre del modelo para simplificar este ejemplo. Lo único que debemos 
+hacer antes de escribir esta clase es un mapeo entre los tipos de bloques y las clases de bloques reales. Luego se define el 
+modelo desde la configuración pasando todo a la clase `ResNet`:
+
+```py
+from transformers import PreTrainedModel
+from timm.models.resnet import BasicBlock, Bottleneck, ResNet
+from .configuration_resnet import ResnetConfig
+
+
+BLOCK_MAPPING = {"basic": BasicBlock, "bottleneck": Bottleneck}
+
+
+class ResnetModel(PreTrainedModel):
+    config_class = ResnetConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        block_layer = BLOCK_MAPPING[config.block_type]
+        self.model = ResNet(
+            block_layer,
+            config.layers,
+            num_classes=config.num_classes,
+            in_chans=config.input_channels,
+            cardinality=config.cardinality,
+            base_width=config.base_width,
+            stem_width=config.stem_width,
+            stem_type=config.stem_type,
+            avg_down=config.avg_down,
+        )
+
+    def forward(self, tensor):
+        return self.model.forward_features(tensor)
+```
+
+Para el modelo que clasificará las imágenes, solo cambiamos el método de avance (es decir, el método `forward`):
+
+```py
+class ResnetModelForImageClassification(PreTrainedModel):
+    config_class = ResnetConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        block_layer = BLOCK_MAPPING[config.block_type]
+        self.model = ResNet(
+            block_layer,
+            config.layers,
+            num_classes=config.num_classes,
+            in_chans=config.input_channels,
+            cardinality=config.cardinality,
+            base_width=config.base_width,
+            stem_width=config.stem_width,
+            stem_type=config.stem_type,
+            avg_down=config.avg_down,
+        )
+
+    def forward(self, tensor, labels=None):
+        logits = self.model(tensor)
+        if labels is not None:
+            loss = torch.nn.cross_entropy(logits, labels)
+            return {"loss": loss, "logits": logits}
+        return {"logits": logits}
+```
+
+En ambos casos, observa cómo heredamos de `PreTrainedModel` y llamamos a la inicialización de la superclase con `config` 
+(un poco como cuando escribes `torch.nn.Module`). La línea que establece `config_class` no es obligatoria, a menos 
+que quieras registrar tu modelo con las clases automáticas (consulta la última sección).
+
+<Tip>
+
+Si tu modelo es muy similar a un modelo dentro de la biblioteca, puedes reutilizar la misma configuración de ese modelo.
+
+</Tip>
+
+Puedes hacer que tu modelo devuelva lo que quieras, pero devolver un diccionario como lo hicimos para 
+`ResnetModelForImageClassification`, con el `loss` incluido cuando se pasan las etiquetas, hará que tu modelo se pueda 
+usar directamente dentro de la clase [`Trainer`]. Usar otro formato de salida está bien, siempre y cuando estés planeando usar 
+tu propio bucle de entrenamiento u otra biblioteca para el entrenamiento.
+
+Ahora que tenemos nuestra clase, vamos a crear un modelo:
+
+```py
+resnet50d = ResnetModelForImageClassification(resnet50d_config)
+```
+
+Nuevamente, puedes usar cualquiera de los métodos de [`PreTrainedModel`], como [`~PreTrainedModel.save_pretrained`] o 
+[`~PreTrainedModel.push_to_hub`]. Usaremos el segundo en la siguiente sección y veremos cómo pasar los pesos del modelo 
+con el código de nuestro modelo. Pero primero, carguemos algunos pesos previamente entrenados dentro de nuestro modelo.
+
+En tu caso de uso, probablemente estarás entrenando tu modelo personalizado con tus propios datos. Para ir rápido en este 
+tutorial, usaremos la versión preentrenada de resnet50d. Dado que nuestro modelo es solo un envoltorio alrededor del resnet50d 
+original, será fácil transferir esos pesos:
+
+```py
+import timm
+
+pretrained_model = timm.create_model("resnet50d", pretrained=True)
+resnet50d.model.load_state_dict(pretrained_model.state_dict())
+```
+
+Ahora veamos cómo asegurarnos de que cuando hacemos [`~PreTrainedModel.save_pretrained`] o [`~PreTrainedModel.push_to_hub`], 
+se guarda el código del modelo.
+
+## Enviar el código al _Hub_
+
+<Tip warning={true}>
+
+Esta _API_ es experimental y puede tener algunos cambios leves en las próximas versiones.
+
+</Tip>
+
+Primero, asegúrate de que tu modelo esté completamente definido en un archivo `.py`. Puedes basarte en importaciones 
+relativas a otros archivos, siempre que todos los archivos estén en el mismo directorio (aún no admitimos submódulos 
+para esta característica). Para nuestro ejemplo, definiremos un archivo `modeling_resnet.py` y un archivo 
+`configuration_resnet.py` en una carpeta del directorio de trabajo actual llamado `resnet_model`. El archivo de configuración 
+contiene el código de `ResnetConfig` y el archivo del modelo contiene el código de `ResnetModel` y 
+`ResnetModelForImageClassification`.
+
+```
+.
+└── resnet_model
+    ├── __init__.py
+    ├── configuration_resnet.py
+    └── modeling_resnet.py
+```
+
+El `__init__.py`  puede estar vacío, solo está ahí para que Python detecte que `resnet_model` se puede usar como un módulo.
+
+<Tip warning={true}>
+
+Si copias archivos del modelo desde la biblioteca, deberás reemplazar todas las importaciones relativas en la parte superior 
+del archivo para importarlos desde el paquete `transformers`.
+
+</Tip>
+
+Ten en cuenta que puedes reutilizar (o subclasificar) una configuración o modelo existente.
+
+Para compartir tu modelo con la comunidad, sigue estos pasos: primero importa el modelo y la configuración de ResNet desde 
+los archivos recién creados:
+
+```py
+from resnet_model.configuration_resnet import ResnetConfig
+from resnet_model.modeling_resnet import ResnetModel, ResnetModelForImageClassification
+```
+
+Luego, debes decirle a la biblioteca que deseas copiar el código de esos objetos cuando usas el método `save_pretrained` 
+y registrarlos correctamente con una determinada clase automática (especialmente para modelos), simplemente ejecuta:
+
+```py
+ResnetConfig.register_for_auto_class()
+ResnetModel.register_for_auto_class("AutoModel")
+ResnetModelForImageClassification.register_for_auto_class("AutoModelForImageClassification")
+```
+
+Ten en cuenta que no es necesario especificar una clase automática para la configuración (solo hay una clase automática 
+para ellos, [`AutoConfig`]), pero es diferente para los modelos. Tu modelo personalizado podría ser adecuado para muchas 
+tareas diferentes, por lo que debes especificar cuál de las clases automáticas es la correcta para tu modelo.
+
+A continuación, vamos a crear la configuración y los modelos como lo hicimos antes:
+
+```py
+resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
+resnet50d = ResnetModelForImageClassification(resnet50d_config)
+
+pretrained_model = timm.create_model("resnet50d", pretrained=True)
+resnet50d.model.load_state_dict(pretrained_model.state_dict())
+```
+
+Ahora, para enviar el modelo al Hub, asegúrate de haber iniciado sesión. Ejecuta en tu terminal:
+
+```bash
+huggingface-cli login
+```
+
+o desde un _notebook_:
+
+```py
+from huggingface_hub import notebook_login
+
+notebook_login()
+```
+
+Luego puedes ingresar a tu propio espacio (o una organización de la que seas miembro) de esta manera:
+
+```py
+resnet50d.push_to_hub("custom-resnet50d")
+```
+
+Además de los pesos del modelo y la configuración en formato json, esto también copió los archivos `.py` del modelo y la
+configuración en la carpeta `custom-resnet50d` y subió el resultado al Hub. Puedes verificar el resultado en este 
+[repositorio de modelos](https://huggingface.co/sgugger/custom-resnet50d).
+
+Consulta el tutorial sobre cómo [compartir modelos](model_sharing) para obtener más información sobre el método para subir modelos al Hub.
+
+## Usar un modelo con código personalizado
+
+Puedes usar cualquier configuración, modelo o _tokenizador_ con archivos de código personalizado en tu repositorio con las 
+clases automáticas y el método `from_pretrained`. Todos los archivos y códigos cargados en el Hub se analizan en busca de 
+malware (consulta la documentación de [seguridad del Hub](https://huggingface.co/docs/hub/security#malware-scanning) para 
+obtener más información), pero aún debes revisar el código del modelo y el autor para evitar la ejecución de código malicioso 
+en tu computadora. Configura `trust_remote_code=True` para usar un modelo con código personalizado:
+
+```py
+from transformers import AutoModelForImageClassification
+
+model = AutoModelForImageClassification.from_pretrained("sgugger/custom-resnet50d", trust_remote_code=True)
+```
+
+También se recomienda encarecidamente pasar un _hash_ de confirmación como una "revisión" para asegurarte de que el autor 
+de los modelos no actualizó el código con algunas líneas nuevas maliciosas (a menos que confíes plenamente en los autores 
+de los modelos).
+
+```py
+commit_hash = "ed94a7c6247d8aedce4647f00f20de6875b5b292"
+model = AutoModelForImageClassification.from_pretrained(
+    "sgugger/custom-resnet50d", trust_remote_code=True, revision=commit_hash
+)
+```
+
+Ten en cuenta que al navegar por el historial de confirmaciones del repositorio del modelo en Hub, hay un botón para copiar 
+fácilmente el hash de confirmación de cualquier _commit_.
+
+## Registrar un model con código personalizado a las clases automáticas
+
+Si estás escribiendo una biblioteca que amplía 🤗 Transformers, es posible que quieras ampliar las clases automáticas para 
+incluir tu propio modelo. Esto es diferente de enviar el código al Hub en el sentido de que los usuarios necesitarán importar 
+tu biblioteca para obtener los modelos personalizados (al contrario de descargar automáticamente el código del modelo desde Hub).
+
+Siempre que tu configuración tenga un atributo `model_type` que sea diferente de los tipos de modelos existentes, y que tus 
+clases modelo tengan los atributos `config_class` correctos, puedes agregarlos a las clases automáticas de la siguiente manera:
+
+```py
+from transformers import AutoConfig, AutoModel, AutoModelForImageClassification
+
+AutoConfig.register("resnet", ResnetConfig)
+AutoModel.register(ResnetConfig, ResnetModel)
+AutoModelForImageClassification.register(ResnetConfig, ResnetModelForImageClassification)
+```
+
+Ten en cuenta que el primer argumento utilizado al registrar tu configuración personalizada en [`AutoConfig`] debe coincidir
+con el `model_type` de tu configuración personalizada, y el primer argumento utilizado al registrar tus modelos personalizados
+en cualquier clase del modelo automático debe coincidir con el `config_class ` de esos modelos.
--- a/docs/source/es/tasks/language_modeling.mdx
+++ b/docs/source/es/tasks/language_modeling.mdx
@@ -141,6 +141,7 @@ Ahora necesitas una segunda función de preprocesamiento para capturar el texto
 >>> def group_texts(examples):
 ...     concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
 ...     total_length = len(concatenated_examples[list(examples.keys())[0]])
+...     total_length = (total_length // block_size) * block_size
 ...     result = {
 ...         k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
 ...         for k, t in concatenated_examples.items()
--- a/docs/source/it/_toctree.yml
+++ b/docs/source/it/_toctree.yml
@@ -11,4 +11,28 @@
    title: Pipeline per l'inferenza
  - local: autoclass_tutorial
    title: Carica istanze pre-allenate con AutoClass
+  - local: preprocessing
+    title: Preprocess
+  - local: training
+    title: Mettere a punto un modello pre-addestrato
+  - local: accelerate
+    title: Allenamento distribuito con 🤗 Accelerate
+  - local: model_sharing
+    title: Condividere un modello
  title: Esercitazione
+- sections:
+  - local: create_a_model
+    title: Crea un'architettura personalizzata
+  - local: custom_models
+    title: Condividere modelli personalizzati
+  - local: run_scripts
+    title: Addestramento con script
+  - local: multilingual
+    title: Modelli multilingua per l'inferenza
+  - local: converting_tensorflow_models
+    title: Convertire modelli tensorflow
+  - local: serialization
+    title: Esporta modelli Transformers
+  - local: debugging
+    title: Debugging
+  title: Guide pratiche
--- a/docs/source/it/accelerate.mdx
+++ b/docs/source/it/accelerate.mdx
@@ -0,0 +1,132 @@
+<!--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.
+-->
+
+# Allenamento distribuito con 🤗 Accelerate
+
+La parallelizzazione è emersa come strategia per allenare modelli sempre più grandi su hardware limitato e accelerarne la velocità di allenamento di diversi ordini di magnitudine. In Hugging Face, abbiamo creato la libreria [🤗 Accelerate](https://huggingface.co/docs/accelerate/index.html) per aiutarti ad allenare in modo semplice un modello 🤗 Transformers su qualsiasi tipo di configurazione distribuita, sia che si tratti di più GPU su una sola macchina o di più GPU su più macchine. In questo tutorial, imparerai come personalizzare il training loop nativo di PyTorch per consentire l'addestramento in un ambiente distribuito.
+
+## Configurazione
+
+Inizia installando 🤗 Accelerate:
+
+```bash
+pip install accelerate
+```
+
+Poi importa e crea un oggetto [`Accelerator`](https://huggingface.co/docs/accelerate/accelerator.html#accelerate.Accelerator). `Accelerator` rileverà automaticamente il tuo setup distribuito e inizializzerà tutte le componenti necessarie per l'allenamento. Non dovrai allocare esplicitamente il tuo modello su un device.
+
+```py
+>>> from accelerate import Accelerator
+
+>>> accelerator = Accelerator()
+```
+
+## Preparati ad accelerare
+
+Il prossimo passo è quello di passare tutti gli oggetti rilevanti per l'allenamento al metodo [`prepare`](https://huggingface.co/docs/accelerate/accelerator.html#accelerate.Accelerator.prepare). Questo include i tuoi DataLoaders per l'allenamento e per la valutazione, un modello e un ottimizzatore:
+
+```py
+>>> train_dataloader, eval_dataloader, model, optimizer = accelerator.prepare(
+...     train_dataloader, eval_dataloader, model, optimizer
+... )
+```
+
+## Backward
+
+Infine, sostituisci il tipico metodo `loss.backward()` nel tuo loop di allenamento con il metodo [`backward`](https://huggingface.co/docs/accelerate/accelerator.html#accelerate.Accelerator.backward) di 🤗 Accelerate:
+
+```py
+>>> for epoch in range(num_epochs):
+...     for batch in train_dataloader:
+...         outputs = model(**batch)
+...         loss = outputs.loss
+...         accelerator.backward(loss)
+
+...         optimizer.step()
+...         lr_scheduler.step()
+...         optimizer.zero_grad()
+...         progress_bar.update(1)
+```
+
+Come puoi vedere nel seguente codice, hai solo bisogno di aggiungere quattro righe in più di codice al tuo training loop per abilitare l'allenamento distribuito!
+
+```diff
+ from accelerate import Accelerator
+  from transformers import AdamW, AutoModelForSequenceClassification, get_scheduler
+
+ accelerator = Accelerator()
+
+  model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)
+  optimizer = AdamW(model.parameters(), lr=3e-5)
+
+- device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+- model.to(device)
+
+ train_dataloader, eval_dataloader, model, optimizer = accelerator.prepare(
+     train_dataloader, eval_dataloader, model, optimizer
+ )
+
+  num_epochs = 3
+  num_training_steps = num_epochs * len(train_dataloader)
+  lr_scheduler = get_scheduler(
+      "linear",
+      optimizer=optimizer,
+      num_warmup_steps=0,
+      num_training_steps=num_training_steps
+  )
+
+  progress_bar = tqdm(range(num_training_steps))
+
+  model.train()
+  for epoch in range(num_epochs):
+      for batch in train_dataloader:
+-         batch = {k: v.to(device) for k, v in batch.items()}
+          outputs = model(**batch)
+          loss = outputs.loss
+-         loss.backward()
+         accelerator.backward(loss)
+
+          optimizer.step()
+          lr_scheduler.step()
+          optimizer.zero_grad()
+          progress_bar.update(1)
+```
+
+## Allenamento
+
+Una volta che hai aggiunto le righe di codice rilevanti, lancia il tuo allenamento in uno script o in un notebook come Colaboratory.
+
+### Allenamento con uno script
+
+Se stai eseguendo il tuo allenamento da uno script, esegui il comando seguente per creare e salvare un file di configurazione:
+
+```bash
+accelerate config
+```
+
+Poi lancia il tuo allenamento con:
+
+```bash
+accelerate launch train.py
+```
+
+### Allenamento con un notebook
+
+La libreria 🤗 Accelerate può anche essere utilizzata in un notebook se stai pianificando di utilizzare le TPU di Colaboratory. Inserisci tutto il codice legato all'allenamento in una funzione, e passala al `notebook_launcher`:
+
+```py
+>>> from accelerate import notebook_launcher
+
+>>> notebook_launcher(training_function)
+```
+
+Per maggiori informazioni relative a 🤗 Accelerate e le sue numerose funzionalità, fai riferimento alla [documentazione](https://huggingface.co/docs/accelerate/index.html).
--- a/docs/source/it/converting_tensorflow_models.mdx
+++ b/docs/source/it/converting_tensorflow_models.mdx
@@ -0,0 +1,155 @@
+<!--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.
+-->
+
+# Convertire checkpoint di Tensorflow
+
+È disponibile un'interfaccia a linea di comando per convertire gli originali checkpoint di Bert/GPT/GPT-2/Transformer-XL/XLNet/XLM 
+in modelli che possono essere caricati utilizzando i metodi `from_pretrained` della libreria.
+
+<Tip>
+
+A partire dalla versione 2.3.0 lo script di conversione è parte di transformers CLI (**transformers-cli**), disponibile in ogni installazione 
+di transformers >=2.3.0.
+
+La seguente documentazione riflette il formato dei comandi di **transformers-cli convert**.
+
+</Tip>
+
+## BERT
+
+Puoi convertire qualunque checkpoint Tensorflow di BERT (in particolare 
+[i modeli pre-allenati rilasciati da Google](https://github.com/google-research/bert#pre-trained-models)) 
+in un file di salvataggio Pytorch utilizzando lo script 
+[convert_bert_original_tf_checkpoint_to_pytorch.py](https://github.com/huggingface/transformers/tree/main/src/transformers/models/bert/convert_bert_original_tf_checkpoint_to_pytorch.py).
+
+Questo CLI prende come input un checkpoint di Tensorflow (tre files che iniziano con `bert_model.ckpt`) ed il relativo 
+file di configurazione (`bert_config.json`), crea un modello Pytorch per questa configurazione, carica i pesi dal
+checkpoint di Tensorflow nel modello di Pytorch e salva il modello che ne risulta in un file di salvataggio standard di Pytorch che 
+può essere importato utilizzando `from_pretrained()` (vedi l'esempio nel
+[quicktour](quicktour) , [run_glue.py](https://github.com/huggingface/transformers/tree/main/examples/pytorch/text-classification/run_glue.py) ).
+
+Devi soltanto lanciare questo script di conversione **una volta** per ottenere un modello Pytorch. Dopodichè, potrai tralasciare 
+il checkpoint di Tensorflow (i tre files che iniziano con `bert_model.ckpt`), ma assicurati di tenere il file di configurazione 
+(`bert_config.json`) ed il file di vocabolario (`vocab.txt`) in quanto queste componenti sono necessarie anche per il modello di Pytorch.
+
+Per lanciare questo specifico script di conversione avrai bisogno di un'installazione di Tensorflow e di Pytorch
+(`pip install tensorflow`). Il resto della repository richiede soltanto Pytorch.
+
+Questo è un esempio del processo di conversione per un modello `BERT-Base Uncased` pre-allenato:
+
+```bash
+export BERT_BASE_DIR=/path/to/bert/uncased_L-12_H-768_A-12
+transformers-cli convert --model_type bert \
+  --tf_checkpoint $BERT_BASE_DIR/bert_model.ckpt \
+  --config $BERT_BASE_DIR/bert_config.json \
+  --pytorch_dump_output $BERT_BASE_DIR/pytorch_model.bin
+```
+
+Puoi scaricare i modelli pre-allenati di Google per la conversione [qua](https://github.com/google-research/bert#pre-trained-models).
+
+## ALBERT
+
+Per il modello ALBERT, converti checkpoint di Tensoflow in Pytorch utilizzando lo script 
+[convert_albert_original_tf_checkpoint_to_pytorch.py](https://github.com/huggingface/transformers/tree/main/src/transformers/models/albert/convert_albert_original_tf_checkpoint_to_pytorch.py).
+
+Il CLI prende come input un checkpoint di Tensorflow (tre files che iniziano con `model.ckpt-best`) e i relativi file di 
+configurazione (`albert_config.json`), dopodichè crea e salva un modello Pytorch. Per lanciare questa conversione 
+avrai bisogno di un'installazione di Tensorflow e di Pytorch.
+
+Ecco un esempio del procedimento di conversione di un modello `ALBERT Base` pre-allenato:
+
+```bash
+export ALBERT_BASE_DIR=/path/to/albert/albert_base
+transformers-cli convert --model_type albert \
+  --tf_checkpoint $ALBERT_BASE_DIR/model.ckpt-best \
+  --config $ALBERT_BASE_DIR/albert_config.json \
+  --pytorch_dump_output $ALBERT_BASE_DIR/pytorch_model.bin
+```
+
+Puoi scaricare i modelli pre-allenati di Google per la conversione [qui](https://github.com/google-research/albert#pre-trained-models).
+
+## OpenAI GPT
+
+Ecco un esempio del processo di conversione di un modello OpenAI GPT pre-allenato, assumendo che il tuo checkpoint di NumPy
+sia salvato nello stesso formato dei modelli pre-allenati OpenAI (vedi [qui](https://github.com/openai/finetune-transformer-lm)):
+```bash
+export OPENAI_GPT_CHECKPOINT_FOLDER_PATH=/path/to/openai/pretrained/numpy/weights
+transformers-cli convert --model_type gpt \
+  --tf_checkpoint $OPENAI_GPT_CHECKPOINT_FOLDER_PATH \
+  --pytorch_dump_output $PYTORCH_DUMP_OUTPUT \
+  [--config OPENAI_GPT_CONFIG] \
+  [--finetuning_task_name OPENAI_GPT_FINETUNED_TASK] \
+```
+
+## OpenAI GPT-2
+
+Ecco un esempio del processo di conversione di un modello OpenAI GPT-2 pre-allenato (vedi [qui](https://github.com/openai/gpt-2)):
+
+```bash
+export OPENAI_GPT2_CHECKPOINT_PATH=/path/to/gpt2/pretrained/weights
+transformers-cli convert --model_type gpt2 \
+  --tf_checkpoint $OPENAI_GPT2_CHECKPOINT_PATH \
+  --pytorch_dump_output $PYTORCH_DUMP_OUTPUT \
+  [--config OPENAI_GPT2_CONFIG] \
+  [--finetuning_task_name OPENAI_GPT2_FINETUNED_TASK]
+```
+
+## Transformer-XL
+
+
+Ecco un esempio del processo di conversione di un modello Transformer-XL pre-allenato 
+(vedi [qui](https://github.com/kimiyoung/transformer-xl/tree/master/tf#obtain-and-evaluate-pretrained-sota-models)):
+
+```bash
+export TRANSFO_XL_CHECKPOINT_FOLDER_PATH=/path/to/transfo/xl/checkpoint
+transformers-cli convert --model_type transfo_xl \
+  --tf_checkpoint $TRANSFO_XL_CHECKPOINT_FOLDER_PATH \
+  --pytorch_dump_output $PYTORCH_DUMP_OUTPUT \
+  [--config TRANSFO_XL_CONFIG] \
+  [--finetuning_task_name TRANSFO_XL_FINETUNED_TASK]
+```
+
+## XLNet
+
+Ecco un esempio del processo di conversione di un modello XLNet pre-allenato:
+
+```bash
+export TRANSFO_XL_CHECKPOINT_PATH=/path/to/xlnet/checkpoint
+export TRANSFO_XL_CONFIG_PATH=/path/to/xlnet/config
+transformers-cli convert --model_type xlnet \
+  --tf_checkpoint $TRANSFO_XL_CHECKPOINT_PATH \
+  --config $TRANSFO_XL_CONFIG_PATH \
+  --pytorch_dump_output $PYTORCH_DUMP_OUTPUT \
+  [--finetuning_task_name XLNET_FINETUNED_TASK] \
+```
+
+## XLM
+
+Ecco un esempio del processo di conversione di un modello XLM pre-allenato:
+
+```bash
+export XLM_CHECKPOINT_PATH=/path/to/xlm/checkpoint
+transformers-cli convert --model_type xlm \
+  --tf_checkpoint $XLM_CHECKPOINT_PATH \
+  --pytorch_dump_output $PYTORCH_DUMP_OUTPUT
+ [--config XML_CONFIG] \
+ [--finetuning_task_name XML_FINETUNED_TASK]
+```
+
+## T5
+
+Ecco un esempio del processo di conversione di un modello T5 pre-allenato:
+
+```bash
+export T5=/path/to/t5/uncased_L-12_H-768_A-12
+transformers-cli convert --model_type t5 \
+  --tf_checkpoint $T5/t5_model.ckpt \
+  --config $T5/t5_config.json \
+  --pytorch_dump_output $T5/pytorch_model.bin
+```
--- a/docs/source/it/create_a_model.mdx
+++ b/docs/source/it/create_a_model.mdx
@@ -0,0 +1,357 @@
+<!--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.
+-->
+
+# Crea un'architettura personalizzata 
+
+Una [`AutoClass`](model_doc/auto) deduce automaticamente il modello dell'architettura e scarica la configurazione e i pesi pre-allenati. Generalmente, noi consigliamo di usare un `AutoClass` per produrre un codice indipendente dal checkpoint. Ma gli utenti che desiderano un controllo maggiore su parametri specifici del modello possono creare un modello 🤗 Transformers personalizzato da poche classi base. Questo potrebbe essere particolarmente utile per qualunque persona sia interessata nel studiare, allenare o sperimentare con un modello 🤗 Transformers. In questa guida, approfondisci la creazione di un modello personalizzato senza `AutoClass`. Impara come:
+
+- Caricare e personalizzare una configurazione del modello.
+- Creare un'architettura modello.
+- Creare un tokenizer lento e veloce per il testo.
+- Creare un estrattore di caratteristiche per attività riguardanti audio o immagini.
+- Creare un processore per attività multimodali.
+
+## Configurazione
+
+Una [configurazione](main_classes/configuration) si riferisce agli attributi specifici di un modello. Ogni configurazione del modello ha attributi diversi; per esempio, tutti i modelli npl hanno questi attributi in comune `hidden_size`, `num_attention_heads`, `num_hidden_layers` e `vocab_size`. Questi attributi specificano il numero di attention heads o strati nascosti con cui costruire un modello.
+
+Dai un'occhiata più da vicino a [DistilBERT](model_doc/distilbert) accedendo a [`DistilBertConfig`] per ispezionare i suoi attributi:
+
+```py
+>>> from transformers import DistilBertConfig
+
+>>> config = DistilBertConfig()
+>>> print(config)
+DistilBertConfig {
+  "activation": "gelu",
+  "attention_dropout": 0.1,
+  "dim": 768,
+  "dropout": 0.1,
+  "hidden_dim": 3072,
+  "initializer_range": 0.02,
+  "max_position_embeddings": 512,
+  "model_type": "distilbert",
+  "n_heads": 12,
+  "n_layers": 6,
+  "pad_token_id": 0,
+  "qa_dropout": 0.1,
+  "seq_classif_dropout": 0.2,
+  "sinusoidal_pos_embds": false,
+  "transformers_version": "4.16.2",
+  "vocab_size": 30522
+}
+```
+
+[`DistilBertConfig`] mostra tutti gli attributi predefiniti usati per costruire una base [`DistilBertModel`]. Tutti gli attributi sono personalizzabili, creando uno spazio per sperimentare. Per esempio, puoi configurare un modello predefinito per:
+
+- Provare un funzione di attivazione diversa con il parametro `activation`.
+- Utilizzare tasso di drop out più elevato per le probalità di attention con il parametro `attention_dropout`.
+
+```py
+>>> my_config = DistilBertConfig(activation="relu", attention_dropout=0.4)
+>>> print(my_config)
+DistilBertConfig {
+  "activation": "relu",
+  "attention_dropout": 0.4,
+  "dim": 768,
+  "dropout": 0.1,
+  "hidden_dim": 3072,
+  "initializer_range": 0.02,
+  "max_position_embeddings": 512,
+  "model_type": "distilbert",
+  "n_heads": 12,
+  "n_layers": 6,
+  "pad_token_id": 0,
+  "qa_dropout": 0.1,
+  "seq_classif_dropout": 0.2,
+  "sinusoidal_pos_embds": false,
+  "transformers_version": "4.16.2",
+  "vocab_size": 30522
+}
+```
+
+Nella funzione [`~PretrainedConfig.from_pretrained`] possono essere modificati gli attributi del modello pre-allenato:
+
+```py
+>>> my_config = DistilBertConfig.from_pretrained("distilbert-base-uncased", activation="relu", attention_dropout=0.4)
+```
+
+Quando la configurazione del modello ti soddisfa, la puoi salvare con [`~PretrainedConfig.save_pretrained`]. Il file della tua configurazione è memorizzato come file JSON nella save directory specificata:
+
+```py
+>>> my_config.save_pretrained(save_directory="./your_model_save_path")
+```
+
+Per riutilizzare la configurazione del file, caricalo con [`~PretrainedConfig.from_pretrained`]:
+
+```py
+>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/my_config.json")
+```
+
+<Tip>
+
+Puoi anche salvare il file di configurazione come dizionario oppure come la differenza tra gli attributi della tua configurazione personalizzata e gli attributi della configurazione predefinita! Guarda la documentazione [configuration](main_classes/configuration) per più dettagli.
+
+</Tip>
+
+## Modello
+
+Il prossimo passo e di creare [modello](main_classes/models). Il modello - vagamente riferito anche come architettura - definisce cosa ogni strato deve fare e quali operazioni stanno succedendo. Attributi come `num_hidden_layers` provenienti dalla configurazione sono usati per definire l'architettura. Ogni modello condivide la classe base [`PreTrainedModel`] e alcuni metodi comuni come il ridimensionamento degli input embeddings e la soppressione delle self-attention heads . Inoltre, tutti i modelli sono la sottoclasse di [`torch.nn.Module`](https://pytorch.org/docs/stable/generated/torch.nn.Module.html), [`tf.keras.Model`](https://www.tensorflow.org/api_docs/python/tf/keras/Model) o [`flax.linen.Module`](https://flax.readthedocs.io/en/latest/flax.linen.html#module). Cio significa che i modelli sono compatibili con l'uso di ciascun di framework.
+
+<frameworkcontent>
+<pt>
+Carica gli attributi della tua configurazione personalizzata nel modello:
+
+```py
+>>> from transformers import DistilBertModel
+
+>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/my_config.json")
+>>> model = DistilBertModel(my_config)
+```
+
+Questo crea modelli con valori casuali invece di pesi pre-allenati. Non sarai in grado di usare questo modello per niente di utile finché non lo alleni. L'allenamento è un processo costoso e che richiede tempo . Generalmente è meglio usare un modello pre-allenato per ottenere risultati migliori velocemente, utilizzando solo una frazione delle risorse neccesarie per l'allenamento.
+
+Crea un modello pre-allenato con [`~PreTrainedModel.from_pretrained`]:
+
+```py
+>>> model = DistilBertModel.from_pretrained("distilbert-base-uncased")
+```
+
+Quando carichi pesi pre-allenati, la configurazione del modello predefinito è automaticamente caricata se il modello è fornito da 🤗 Transformers. Tuttavia, puoi ancora sostituire gli attributi - alcuni o tutti - di configurazione del modello predefinito con i tuoi se lo desideri:
+
+```py
+>>> model = DistilBertModel.from_pretrained("distilbert-base-uncased", config=my_config)
+```
+</pt>
+<tf>
+Carica gli attributi di configurazione personalizzati nel modello:
+
+```py
+>>> from transformers import TFDistilBertModel
+
+>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/my_config.json")
+>>> tf_model = TFDistilBertModel(my_config)
+```
+
+
+Questo crea modelli con valori casuali invece di pesi pre-allenati. Non sarai in grado di usare questo modello per niente di utile finché non lo alleni. L'allenamento è un processo costoso e che richiede tempo . Generalmente è meglio usare un modello pre-allenato per ottenere risultati migliori velocemente, utilizzando solo una frazione delle risorse neccesarie per l'allenamento.
+
+Crea un modello pre-allenoto con [`~TFPreTrainedModel.from_pretrained`]:
+
+```py
+>>> tf_model = TFDistilBertModel.from_pretrained("distilbert-base-uncased")
+```
+
+Quando carichi pesi pre-allenati, la configurazione del modello predefinito è automaticamente caricato se il modello è fornito da 🤗 Transformers. Tuttavia, puoi ancora sostituire gli attributi - alcuni o tutti - di configurazione del modello predefinito con i tuoi se lo desideri:
+
+```py
+>>> tf_model = TFDistilBertModel.from_pretrained("distilbert-base-uncased", config=my_config)
+```
+
+</tf>
+</frameworkcontent>
+
+### Model head
+
+A questo punto, hai un modello DistilBERT base i cui output sono gli *hidden states* (in italiano stati nascosti). Gli stati nascosti sono passati come input a un model head per produrre l'output finale. 🤗 Transformers fornisce un model head diverso per ogni attività fintanto che il modello supporta l'attività  (i.e., non puoi usare DistilBERT per un attività sequence-to-sequence come la traduzione).
+
+<frameworkcontent>
+<pt>
+Per esempio, [`DistilBertForSequenceClassification`] è un modello DistilBERT base con una testa di classificazione per sequenze. La sequenza di classificazione head è uno strato lineare sopra gli output ragruppati.
+
+```py
+>>> from transformers import DistilBertForSequenceClassification
+
+>>> model = DistilBertForSequenceClassification.from_pretrained("distilbert-base-uncased")
+```
+
+Riutilizza facilmente questo checkpoint per un'altra attività passando ad un model head differente. Per un attività di risposta alle domande, utilizzerai il model head [`DistilBertForQuestionAnswering`]. La head per compiti di question answering è simile alla classificazione di sequenza head tranne per il fatto che è uno strato lineare sopra l'output degli stati nascosti (hidden states in inglese) 
+
+```py
+>>> from transformers import DistilBertForQuestionAnswering
+
+>>> model = DistilBertForQuestionAnswering.from_pretrained("distilbert-base-uncased")
+```
+</pt>
+<tf>
+Per esempio, [`TFDistilBertForSequenceClassification`] è un modello DistilBERT base con classificazione di sequenza head. La classificazione di sequenza head è uno strato lineare sopra gli output raggruppati.
+
+```py
+>>> from transformers import TFDistilBertForSequenceClassification
+
+>>> tf_model = TFDistilBertForSequenceClassification.from_pretrained("distilbert-base-uncased")
+```
+
+Riutilizza facilmente questo checkpoint per un altra attività passando ad un modello head diverso. Per un attività di risposta alle domande, utilizzerai il model head [`TFDistilBertForQuestionAnswering`]. Il head di risposta alle domande è simile alla sequenza di classificazione head tranne per il fatto che è uno strato lineare sopra l'output degli stati nascosti (hidden states in inglese)
+
+```py
+>>> from transformers import TFDistilBertForQuestionAnswering
+
+>>> tf_model = TFDistilBertForQuestionAnswering.from_pretrained("distilbert-base-uncased")
+```
+</tf>
+</frameworkcontent>
+
+## Tokenizer
+
+L'ultima classe base di cui hai bisogno prima di utilizzare un modello per i dati testuali è un [tokenizer](main_classes/tokenizer) per convertire il testo grezzo in tensori. Ci sono due tipi di tokenizer che puoi usare con 🤗 Transformers:
+
+- [`PreTrainedTokenizer`]: un'implementazione Python di un tokenizer.
+- [`PreTrainedTokenizerFast`]: un tokenizer dalla nostra libreria [🤗 Tokenizer](https://huggingface.co/docs/tokenizers/python/latest/) basata su Rust. Questo tipo di tokenizer è significativamente più veloce, specialmente durante la batch tokenization, grazie alla sua implementazione Rust. Il tokenizer veloce offre anche metodi aggiuntivi come *offset mapping* che associa i token alle loro parole o caratteri originali.
+
+Entrambi i tokenizer supportano metodi comuni come la codifica e la decodifica, l'aggiunta di nuovi token e la gestione di token speciali.
+
+<Tip warning={true}>
+
+Non tutti i modelli supportano un tokenizer veloce. Dai un'occhiata a questo [tabella](index#supported-frameworks) per verificare se un modello ha il supporto per tokenizer veloce. 
+
+</Tip>
+
+Se hai addestrato il tuo tokenizer, puoi crearne uno dal tuo file *vocabolario*: 
+
+```py
+>>> from transformers import DistilBertTokenizer
+
+>>> my_tokenizer = DistilBertTokenizer(vocab_file="my_vocab_file.txt", do_lower_case=False, padding_side="left")
+```
+
+È importante ricordare che il vocabolario di un tokenizer personalizzato sarà diverso dal vocabolario generato dal tokenizer di un modello preallenato. È necessario utilizzare il vocabolario di un modello preallenato se si utilizza un modello preallenato, altrimenti gli input non avranno senso. Crea un tokenizer con il vocabolario di un modello preallenato con la classe [`DistilBertTokenizer`]:
+
+```py
+>>> from transformers import DistilBertTokenizer
+
+>>> slow_tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
+```
+
+Crea un tokenizer veloce con la classe [`DistilBertTokenizerFast`]:
+
+```py
+>>> from transformers import DistilBertTokenizerFast
+
+>>> fast_tokenizer = DistilBertTokenizerFast.from_pretrained("distilbert-base-uncased")
+```
+
+<Tip>
+
+Per l'impostazione predefinita, [`AutoTokenizer`] proverà a caricare un tokenizer veloce. Puoi disabilitare questo comportamento impostando `use_fast=False` in `from_pretrained`.
+
+</Tip>
+
+## Estrattore Di Feature
+
+Un estrattore di caratteristiche (feature in inglese) elabora input audio o immagini. Eredita dalla classe [`~feature_extraction_utils.FeatureExtractionMixin`] base e può anche ereditare dalla classe [`ImageFeatureExtractionMixin`] per l'elaborazione delle caratteristiche dell'immagine o dalla classe [`SequenceFeatureExtractor`] per l'elaborazione degli input audio.
+
+A seconda che tu stia lavorando a un'attività audio o visiva, crea un estrattore di caratteristiche associato al modello che stai utilizzando. Ad esempio, crea un [`ViTFeatureExtractor`] predefinito se stai usando [ViT](model_doc/vit) per la classificazione delle immagini:
+
+```py
+>>> from transformers import ViTFeatureExtractor
+
+>>> vit_extractor = ViTFeatureExtractor()
+>>> print(vit_extractor)
+ViTFeatureExtractor {
+  "do_normalize": true,
+  "do_resize": true,
+  "feature_extractor_type": "ViTFeatureExtractor",
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "resample": 2,
+  "size": 224
+}
+```
+
+<Tip>
+
+Se non stai cercando alcuna personalizzazione, usa il metodo `from_pretrained` per caricare i parametri di default dell'estrattore di caratteristiche di un modello.
+
+</Tip>
+
+Modifica uno qualsiasi dei parametri [`ViTFeatureExtractor`] per creare il tuo estrattore di caratteristiche personalizzato:
+
+```py
+>>> from transformers import ViTFeatureExtractor
+
+>>> my_vit_extractor = ViTFeatureExtractor(resample="PIL.Image.BOX", do_normalize=False, image_mean=[0.3, 0.3, 0.3])
+>>> print(my_vit_extractor)
+ViTFeatureExtractor {
+  "do_normalize": false,
+  "do_resize": true,
+  "feature_extractor_type": "ViTFeatureExtractor",
+  "image_mean": [
+    0.3,
+    0.3,
+    0.3
+  ],
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "resample": "PIL.Image.BOX",
+  "size": 224
+}
+```
+
+Per gli input audio, puoi creare un [`Wav2Vec2FeatureExtractor`] e personalizzare i parametri in modo simile:
+
+```py
+>>> from transformers import Wav2Vec2FeatureExtractor
+
+>>> w2v2_extractor = Wav2Vec2FeatureExtractor()
+>>> print(w2v2_extractor)
+Wav2Vec2FeatureExtractor {
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "return_attention_mask": false,
+  "sampling_rate": 16000
+}
+```
+
+## Processore
+
+Per modelli che supportano attività multimodali, 🤗 Transformers offre una classe di processore che racchiude comodamente un estrattore di caratteristiche e un tokenizer in un unico oggetto. Ad esempio, utilizziamo [`Wav2Vec2Processor`] per un'attività di riconoscimento vocale automatico (ASR). ASR trascrive l'audio in testo, quindi avrai bisogno di un estrattore di caratteristiche e di un tokenizer.
+
+Crea un estrattore di feature per gestire gli input audio:
+
+```py
+>>> from transformers import Wav2Vec2FeatureExtractor
+
+>>> feature_extractor = Wav2Vec2FeatureExtractor(padding_value=1.0, do_normalize=True)
+```
+
+Crea un tokenizer per gestire gli input di testo:
+
+```py
+>>> from transformers import Wav2Vec2CTCTokenizer
+
+>>> tokenizer = Wav2Vec2CTCTokenizer(vocab_file="my_vocab_file.txt")
+```
+
+Combinare l'estrattore di caratteristiche e il tokenizer in [`Wav2Vec2Processor`]:
+
+```py
+>>> from transformers import Wav2Vec2Processor
+
+>>> processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer)
+```
+
+Con due classi di base - configurazione e modello - e una classe di preelaborazione aggiuntiva (tokenizer, estrattore di caratteristiche o processore), puoi creare qualsiasi modello supportato da 🤗 Transformers. Ognuna di queste classi base è configurabile, consentendoti di utilizzare gli attributi specifici che desideri. È possibile impostare facilmente un modello per l'addestramento o modificare un modello preallenato esistente per la messa a punto.
--- a/docs/source/it/custom_models.mdx
+++ b/docs/source/it/custom_models.mdx
@@ -0,0 +1,352 @@
+<!--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.
+-->
+
+# Condividere modelli personalizzati
+La libreria 🤗 Transformers è studiata per essere facilmente estendibile. Il codice di ogni modello è interamente 
+situato in una sottocartella del repository senza alcuna astrazione, perciò puoi facilmente copiare il file di un 
+modello e modificarlo in base ai tuoi bisogni.
+
+Se stai scrivendo un nuovo modello, potrebbe essere più semplice iniziare da zero. In questo tutorial, ti mostreremo
+come scrivere un modello personalizzato e la sua configurazione in modo che possa essere utilizzato all’interno di
+Transformers, e come condividerlo con la community (assieme al relativo codice) così che tutte le persone possano usarlo, anche
+se non presente nella libreria 🤗 Transformers.
+
+Illustriamo tutto questo su un modello ResNet, avvolgendo la classe ResNet della 
+[libreria timm](https://github.com/rwightman/pytorch-image-models/tree/master/timm) in un [`PreTrainedModel`].
+
+## Scrivere una configurazione personalizzata
+Prima di iniziare a lavorare al modello, scriviamone la configurazione. La configurazione di un modello è un oggetto
+che contiene tutte le informazioni necessarie per la build del modello. Come vedremo nella prossima sezione, il 
+modello può soltanto essere inizializzato tramite `config`, per cui dovremo rendere tale oggetto più completo possibile.
+
+Nel nostro esempio, prenderemo un paio di argomenti della classe ResNet che potremmo voler modificare. 
+Configurazioni differenti ci daranno quindi i differenti possibili tipi di ResNet. Salveremo poi questi argomenti, 
+dopo averne controllato la validità.
+
+```python
+from transformers import PretrainedConfig
+from typing import List
+
+
+class ResnetConfig(PretrainedConfig):
+    model_type = "resnet"
+
+    def __init__(
+        self,
+        block_type="bottleneck",
+        layers: List[int] = [3, 4, 6, 3],
+        num_classes: int = 1000,
+        input_channels: int = 3,
+        cardinality: int = 1,
+        base_width: int = 64,
+        stem_width: int = 64,
+        stem_type: str = "",
+        avg_down: bool = False,
+        **kwargs,
+    ):
+        if block_type not in ["basic", "bottleneck"]:
+            raise ValueError(f"`block` must be 'basic' or bottleneck', got {block}.")
+        if stem_type not in ["", "deep", "deep-tiered"]:
+            raise ValueError(f"`stem_type` must be '', 'deep' or 'deep-tiered', got {block}.")
+
+        self.block_type = block_type
+        self.layers = layers
+        self.num_classes = num_classes
+        self.input_channels = input_channels
+        self.cardinality = cardinality
+        self.base_width = base_width
+        self.stem_width = stem_width
+        self.stem_type = stem_type
+        self.avg_down = avg_down
+        super().__init__(**kwargs)
+```
+
+Le tre cose più importanti da ricordare quando scrivi le tue configurazioni sono le seguenti:
+- Devi ereditare da `Pretrainedconfig`,
+- Il metodo `__init__` del tuo `Pretrainedconfig` deve accettare i kwargs,
+- I `kwargs` devono essere passati alla superclass `__init__`
+
+L’eredità è importante per assicurarsi di ottenere tutte le funzionalità della libreria 🤗 transformers, 
+mentre gli altri due vincoli derivano dal fatto che un `Pretrainedconfig` ha più campi di quelli che stai settando. 
+Quando ricarichi una config da un metodo `from_pretrained`, questi campi devono essere accettati dalla tua config e
+poi inviati alla superclasse.
+
+Definire un `model_type` per la tua configurazione (qua `model_type = “resnet”`) non è obbligatorio, a meno che tu
+non voglia registrare il modello con le classi Auto (vedi l'ultima sezione).
+
+Una volta completato, puoi facilmente creare e salvare la tua configurazione come faresti con ogni altra configurazione
+di modelli della libreria. Ecco come possiamo creare la config di un resnet50d e salvarlo:
+
+```py
+resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
+resnet50d_config.save_pretrained("custom-resnet")
+```
+
+Questo salverà un file chiamato `config.json` all'interno della cartella `custom-resnet`. Potrai poi ricaricare la tua
+config con il metodo `from_pretrained`.
+
+```py
+resnet50d_config = ResnetConfig.from_pretrained("custom-resnet")
+```
+
+Puoi anche usare qualunque altro metodo della classe [`PretrainedConfig`], come [`~PretrainedConfig.push_to_hub`]
+per caricare direttamente la tua configurazione nell'hub.
+
+## Scrivere un modello personalizzato
+
+Ora che abbiamo la nostra configurazione ResNet, possiamo continuare a scrivere il modello. In realtà, ne scriveremo
+due: uno che estrae le features nascoste da una batch di immagini (come [`BertModel`]) e uno che è utilizzabile per 
+la classificazione di immagini (come [`BertModelForSequenceClassification`]).
+
+Come abbiamo menzionato in precedenza, scriveremo soltanto un wrapper del modello, per mantenerlo semplice ai fini di 
+questo esempio. L'unica cosa che dobbiamo fare prima di scrivere questa classe è una mappatura fra i tipi di blocco e 
+le vere classi dei blocchi. Successivamente il modello è definito tramite la configurazione, passando tutto quanto alla
+classe `ResNet`.
+
+```py
+from transformers import PreTrainedModel
+from timm.models.resnet import BasicBlock, Bottleneck, ResNet
+from .configuration_resnet import ResnetConfig
+
+
+BLOCK_MAPPING = {"basic": BasicBlock, "bottleneck": Bottleneck}
+
+
+class ResnetModel(PreTrainedModel):
+    config_class = ResnetConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        block_layer = BLOCK_MAPPING[config.block_type]
+        self.model = ResNet(
+            block_layer,
+            config.layers,
+            num_classes=config.num_classes,
+            in_chans=config.input_channels,
+            cardinality=config.cardinality,
+            base_width=config.base_width,
+            stem_width=config.stem_width,
+            stem_type=config.stem_type,
+            avg_down=config.avg_down,
+        )
+
+    def forward(self, tensor):
+        return self.model.forward_features(tensor)
+```
+
+Per il modello che classificherà le immagini, cambiamo soltanto il metodo forward:
+
+```py
+class ResnetModelForImageClassification(PreTrainedModel):
+    config_class = ResnetConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        block_layer = BLOCK_MAPPING[config.block_type]
+        self.model = ResNet(
+            block_layer,
+            config.layers,
+            num_classes=config.num_classes,
+            in_chans=config.input_channels,
+            cardinality=config.cardinality,
+            base_width=config.base_width,
+            stem_width=config.stem_width,
+            stem_type=config.stem_type,
+            avg_down=config.avg_down,
+        )
+
+    def forward(self, tensor, labels=None):
+        logits = self.model(tensor)
+        if labels is not None:
+            loss = torch.nn.cross_entropy(logits, labels)
+            return {"loss": loss, "logits": logits}
+        return {"logits": logits}
+```
+
+Nota come, in entrambi i casi, ereditiamo da `PreTrainedModel` e chiamiamo l'inizializzazione della superclasse 
+con il metodo `config` (un po' come quando scrivi un normale `torch.nn.Module`). La riga che imposta la  `config_class`
+non è obbligatoria, a meno che tu non voglia registrare il modello con le classi Auto (vedi l'ultima sezione).
+
+<Tip>
+
+Se il tuo modello è molto simile a un modello all'interno della libreria, puoi ri-usare la stessa configurazione di quel modello.
+
+</Tip>
+
+Puoi fare in modo che il tuo modello restituisca in output qualunque cosa tu voglia, ma far restituire un dizionario 
+come abbiamo fatto per `ResnetModelForImageClassification`, con la funzione di perdita inclusa quando vengono passate le labels,
+renderà il tuo modello direttamente utilizzabile all'interno della classe [`Trainer`]. Utilizzare altri formati di output va bene
+se hai in progetto di utilizzare un tuo loop di allenamento, o se utilizzerai un'altra libreria per l'addestramento.
+
+Ora che abbiamo la classe del nostro modello, creiamone uno:
+
+```py
+resnet50d = ResnetModelForImageClassification(resnet50d_config)
+```
+
+Ribadiamo, puoi usare qualunque metodo dei [`PreTrainedModel`], come [`~PreTrainedModel.save_pretrained`] o
+[`~PreTrainedModel.push_to_hub`]. Utilizzeremo quest'ultimo nella prossima sezione, e vedremo come caricare i pesi del
+modello assieme al codice del modello stesso. Ma prima, carichiamo alcuni pesi pre-allenati all'interno del nostro modello.
+
+Nel tuo caso specifico, probabilmente allenerai il tuo modello sui tuoi dati. Per velocizzare in questo tutorial, 
+utilizzeremo la versione pre-allenata del resnet50d. Dato che il nostro modello è soltanto un wrapper attorno a quel modello,
+sarà facile trasferirne i pesi:
+
+```py
+import timm
+
+pretrained_model = timm.create_model("resnet50d", pretrained=True)
+resnet50d.model.load_state_dict(pretrained_model.state_dict())
+```
+
+Vediamo adesso come assicurarci che quando facciamo [`~PreTrainedModel.save_pretrained`] o [`~PreTrainedModel.push_to_hub`], 
+il codice del modello venga salvato.
+
+## Inviare il codice all'Hub
+
+<Tip warning={true}>
+
+Questa API è sperimentale e potrebbe avere alcuni cambiamenti nei prossimi rilasci.
+
+</Tip>
+
+Innanzitutto, assicurati che il tuo modello sia completamente definito in un file `.py`. Può sfruttare import relativi
+ad altri file, purchè questi siano nella stessa directory (non supportiamo ancora sotto-moduli per questa funzionalità).
+Per questo esempio, definiremo un file `modeling_resnet.py` e un file `configuration_resnet.py` in una cartella dell'attuale
+working directory chiamata `resnet_model`. Il file configuration contiene il codice per `ResnetConfig` e il file modeling 
+contiene il codice di `ResnetModel` e `ResnetModelForImageClassification`.
+
+```
+.
+└── resnet_model
+    ├── __init__.py
+    ├── configuration_resnet.py
+    └── modeling_resnet.py
+```
+
+Il file `__init__.py` può essere vuoto, serve solo perchè Python capisca che `resnet_model` può essere utilizzato come un modulo.
+
+<Tip warning={true}>
+
+Se stai copiando i file relativi alla modellazione della libreria, dovrai sostituire tutti gli import relativi in cima al file con import del 
+    pacchetto `transformers`.
+
+</Tip>
+
+Nota che puoi ri-utilizzare (o usare come sottoclassi) un modello/configurazione esistente.
+
+Per condividere il tuo modello con la community, segui questi passi: prima importa il modello ResNet e la sua configurazione 
+dai nuovi file creati:
+
+```py
+from resnet_model.configuration_resnet import ResnetConfig
+from resnet_model.modeling_resnet import ResnetModel, ResnetModelForImageClassification
+```
+
+Dopodichè dovrai dire alla libreria che vuoi copiare i file con il codice di quegli oggetti quando utilizzi il metodo
+`save_pretrained` e registrarli in modo corretto con una Auto classe (specialmente per i modelli). Utilizza semplicemente:
+
+```py
+ResnetConfig.register_for_auto_class()
+ResnetModel.register_for_auto_class("AutoModel")
+ResnetModelForImageClassification.register_for_auto_class("AutoModelForImageClassification")
+```
+
+Nota che non c'è bisogno di specificare una Auto classe per la configurazione (c'è solo una Auto classe per le configurazioni,
+[`AutoConfig`], ma è diversa per i modelli). Il tuo modello personalizato potrebbe essere utilizzato per diverse tasks, 
+per cui devi specificare quale delle classi Auto è quella corretta per il tuo modello.
+
+Successivamente, creiamo i modelli e la config come abbiamo fatto in precedenza:
+
+```py
+resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
+resnet50d = ResnetModelForImageClassification(resnet50d_config)
+
+pretrained_model = timm.create_model("resnet50d", pretrained=True)
+resnet50d.model.load_state_dict(pretrained_model.state_dict())
+```
+
+Adesso, per inviare il modello all'Hub, assicurati di aver effettuato l'accesso. Lancia dal tuo terminale:
+
+```bash
+huggingface-cli login
+```
+
+O da un notebook:
+
+```py
+from huggingface_hub import notebook_login
+
+notebook_login()
+```
+
+Potrai poi inviare il tutto sul tuo profilo (o di un'organizzazione di cui fai parte) in questo modo:
+
+```py
+resnet50d.push_to_hub("custom-resnet50d")
+```
+
+Oltre ai pesi del modello e alla configurazione in formato json, questo ha anche copiato i file `.py` modeling e
+configuration all'interno della cartella `custom-resnet50d` e ha caricato i risultati sull'Hub. Puoi controllare
+i risultati in questa [model repo](https://huggingface.co/sgugger/custom-resnet50d).
+
+Puoi controllare il tutorial di condivisione [tutorial di condivisione](model_sharing) per più informazioni sul 
+metodo con cui inviare all'Hub.
+
+## Usare un modello con codice personalizzato
+
+Puoi usare ogni configurazione, modello o tokenizer con file di codice personalizzati nella sua repository 
+con le classi Auto e il metodo `from_pretrained`. Tutti i files e il codice caricati sull'Hub sono scansionati da malware
+(fai riferimento alla documentazione [Hub security](https://huggingface.co/docs/hub/security#malware-scanning) per più informazioni),
+ma dovresti comunque assicurarti dell'affidabilità del codice e dell'autore per evitare di eseguire codice dannoso sulla tua macchina. 
+Imposta `trust_remote_code=True` per usare un modello con codice personalizzato:
+
+```py
+from transformers import AutoModelForImageClassification
+
+model = AutoModelForImageClassification.from_pretrained("sgugger/custom-resnet50d", trust_remote_code=True)
+```
+
+Inoltre, raccomandiamo fortemente di passare un hash del commit come `revision` per assicurarti che le autrici o gli autori del modello 
+non abbiano modificato il codice con alcune nuove righe dannose (a meno che non ti fidi completamente della fonte):
+
+```py
+commit_hash = "ed94a7c6247d8aedce4647f00f20de6875b5b292"
+model = AutoModelForImageClassification.from_pretrained(
+    "sgugger/custom-resnet50d", trust_remote_code=True, revision=commit_hash
+)
+```
+
+Nota che quando cerchi la storia dei commit della repo del modello sull'Hub, c'è un bottone con cui facilmente copiare il 
+commit hash di ciascun commit.
+
+## Registrare un modello con codice personalizzato nelle classi Auto
+
+Se stai scrivendo una libreria che estende 🤗 Transformers, potresti voler estendere le classi Auto per includere il tuo modello.
+Questo è diverso dall'inviare codice nell'Hub: gli utenti dovranno importare la tua libreria per ottenere il modello personalizzato
+(anzichè scaricare automaticamente il modello dall'Hub).
+
+Finchè il tuo file di configurazione ha un attributo `model_type` diverso dai model types esistenti, e finchè le tue 
+classi modello hanno i corretti attributi `config_class`, potrai semplicemente aggiungerli alle classi Auto come segue:
+
+```py
+from transformers import AutoConfig, AutoModel, AutoModelForImageClassification
+
+AutoConfig.register("resnet", ResnetConfig)
+AutoModel.register(ResnetConfig, ResnetModel)
+AutoModelForImageClassification.register(ResnetConfig, ResnetModelForImageClassification)
+```
+
+Nota che il primo argomento utilizzato quando registri la configurazione di un modello personalizzato con [`AutoConfig`] 
+deve corrispondere al `model_type` della tua configurazione personalizzata, ed il primo argomento utilizzato quando 
+registri i tuoi modelli personalizzati in una qualunque classe Auto del modello deve corrispondere alla `config_class`
+di quei modelli.
--- a/docs/source/it/debugging.mdx
+++ b/docs/source/it/debugging.mdx
@@ -0,0 +1,314 @@
+<!--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.
+-->
+
+# Debugging
+
+## Debug dei problemi di rete multi-GPU
+
+Quando addestri o fai inferenza con `DistributedDataParallel` e GPU multiple, se si verificano problemi di intercomunicazione tra processi e/o nodi, puoi utilizzare il seguente script per diagnosticare i problemi della rete.
+
+```bash
+wget https://raw.githubusercontent.com/huggingface/transformers/main/scripts/distributed/torch-distributed-gpu-test.py
+```
+
+Per esempio per testare come 2 GPU interagiscono fai:
+
+```bash
+python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py
+```
+
+Se entrambi i processi sono in grado di comunicare tra loro e di allocare la memoria della GPU, ciascuno di essi stamperà lo stato OK.
+
+Per più GPU o nodi adatta gli argumenti nello script.
+
+All'interno dello script di diagnostica troverai molti altri dettagli e anche una guida per eseguirlo in ambiente SLURM.
+
+Un livello di debug superiore è aggiungere la variabile d'ambiente `NCCL_DEBUG=INFO` come di seguito:
+
+```bash
+NCCL_DEBUG=INFO python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py
+```
+
+In questo modo si scaricano molte informazioni di debug relative a NCCL, che puoi cercare online in caso di problemi. Oppure, se non hai la sicurezza di come interpretare l'output, puoi condividere il file di log in una Issue.
+
+## Rilevamento di Underflow e Overflow
+
+<Tip>
+
+Questa funzionalità al momento è disponibile solo per PyTorch.
+
+</Tip>
+
+<Tip>
+
+Per addestramento multi-GPU richiede DDP (`torch.distributed.launch`).
+
+</Tip>
+
+<Tip>
+
+Questa funzionalità può essere usata con modelli basati su `nn.Module`.
+
+</Tip>
+
+Se inizi a ottenere `loss=NaN` o il modello presenta qualche altro comportamento anomalo a causa di valori `inf` o `nan` in
+attivazioni o nei pesi, è necessario scoprire dove si verifica il primo underflow o overflow e cosa lo ha determinato. Fortunatamente
+è possibile farlo facilmente attivando un modulo speciale che effettuerà il rilevamento automaticamente.
+
+Se stai usando [`Trainer`], hai bisogno di aggiungere solo:
+
+```bash
+--debug underflow_overflow
+```
+
+ai normali argomenti della riga di comando, o passa `debug="underflow_overflow"` quando viene creato l'oggetto
+[`TrainingArguments`].
+
+Se stai usando il tuo ciclo di allenamento o un altro trainer, puoi ottenere lo stesso risultato con:
+
+```python
+from .debug_utils import DebugUnderflowOverflow
+
+debug_overflow = DebugUnderflowOverflow(model)
+```
+
+[`~debug_utils.DebugUnderflowOverflow`] inserisce dei ganci nel modello che dopo ogni chiamata
+testeranno le variabili di ingresso e di uscita e anche i pesi del modulo corrispondente. Non appena viene rilevato `inf` o
+o `nan` in almeno un elemento delle attivazioni o dei pesi, il programma lo notifica e stampa un rapporto come il seguente (questo è stato rilevato con `google/mt5-small` sotto fp16 mixed precision):
+
+```
+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
+```
+
+L'output di esempio è stato tagliato al centro per brevità.
+
+La seconda colonna mostra il valore dell'elemento più grande in assoluto,così se osserviamo da vicino gli ultimi istanti,
+input e output sono nel range di `1e4`. Questo addestramento è stato eseguito con una mixed precision fp16 e l'ultimo passo usciva fuori (sotto `fp16` il valore più grande prima di `inf` è `64e3`). Per evitare overflows sotto `fp16` le attivazionioni devono rimanere molto al di sotto di `1e4`, perché `1e4 * 1e4 = 1e8` quindi qualsiasi moltiplicazione di matrice con grandi attivazioni porterà a una condizione di overflow numerico.
+
+All'inizio della traccia è possibile scoprire a quale lotto si è verificato il problema (questo `Detected inf/nan during batch_number=0` significa che il problema si è verificato nel primo lotto).
+
+Ogni frame segnalato inizia dichiarando la voce completamente qualificata per il modulo corrispondente per il quale il frame è stato segnalato. 
+Se osserviamo il seguente frame:
+
+```
+                  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
+```
+
+Questo, `encoder.block.2.layer.1.layer_norm` indica che si tratta di un layer norm nel primo layer, del secondo blocco dell'encoder. E le chiamata specifica di `forward` è `T5LayerNorm`.
+
+Osserviamo gli ultimi frame del report:
+
+```
+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
+```
+
+L'ultimo frame report per la funzione `Dropout.forward` con la prima voce per l'unico input e la seconda per l'unico output. Si può notare che è stato richiamato da un attibuto `dropout` dentro la classe `DenseReluDense`. Si può notare che ciò è avvenuto durante il primo strato, del 2° blocco, durante il primissimo lotto. Infine, gli elementi di input più grandi in assoluto sono stati `6.27e+04` e l'equivalente per l'output era `inf`.
+
+Puoi vedere qui, che `T5DenseGatedGeluDense.forward` risulta in output activations, il cui valore massimo assoluto era circa 62,7K, che è molto vicino al limite massimo di 64K di fp16. Nel prossimo frame abbiamo `Dropout` che rinormalizza i pesi, dopo aver azzerato alcuni elementi, il che spinge il valore massimo assoluto a più di 64K e si verifica un overflow.(`inf`).
+
+Come puoi notare, è nei frames precedenti che occorre esaminare quando i numeri iniziano a diventare molto grandi per i valori fp16.
+
+Confrontiamo il report al codice `models/t5/modeling_t5.py`:
+
+```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
+```
+
+Ora è facile vedere la chiamata `dropout`, e tutte le chiamate precedenti.
+
+Poiché il rilevamento avviene in un avanzamento (forward hook in eng.), i rapporti vengono creati immeditamente dopo ogni rientro da `forward` (forward returns in eng.).
+
+Tornando al rapporto completo, per agire e risolvere il problema, dobbiamo andare qualche frame più in alto, dove i numeri hanno iniziato a salire, e probabilmente passare alla modalità `fp32`, in modo che i numeri non trabocchino quando vengono moltiplicati o sommati. Naturalmente, potrebbero esserci altre soluzioni. Per esempio, potremmo spegnere temporanemante `amp` se è abilitato, successivamente spostare `forward` in un helper wrapper, come:
+
+```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)
+```
+
+Poiché il rilevatore automatico riporta solo gli ingressi e le uscite di fotogrammi completi, una volta che si sa dove cercare, si può
+analizzare anche le fasi intermedie di una specifica funzione `forward`. In alcuni casi puoi usare la funzione di supporto `detect_overflow` per indirizzare il rilevatore dove preferisci, ad esempio:
+
+```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)
+```
+
+Si può vedere che abbiamo aggiunto 2 di questi e ora teniamo traccia se `inf` o `nan` per `forwarded_states` è stato rilevato
+da qualche parte.
+
+In realtà, il rilevatore li riporta già, perché ciascuna delle chiamate nell'esempio precedente è un `nn.Module`, ma
+diciamo che se avessimo dei calcoli diretti locali, questo è il modo in cui lo faremmo.
+
+Inoltre, se si istanzia il debugger nel proprio codice, è possibile modificare il numero di fotogrammi stampati rispetto a
+predefinito, ad esempio.:
+
+```python
+from .debug_utils import DebugUnderflowOverflow
+
+debug_overflow = DebugUnderflowOverflow(model, max_frames_to_save=100)
+```
+
+### Tracciamento della mistura assoluta del lotto specifico e del valore massimo
+
+La stessa classe di debug può essere utilizzata per il tracciamento per-batch con la funzione di rilevamento di underflow/overflow disattivata.
+
+Supponiamo di voler osservare i valori minimi e massimi assoluti per tutti gli ingredienti di ogni chiamata `forward` di un dato lotto.
+lotto, e che lo si voglia fare solo per i lotti 1 e 3. Si istanzia questa classe come:
+
+```python
+debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1, 3])
+```
+
+Ora i batch completi 1 e 3 saranno tracciati utilizzando lo stesso formato del rilevatore di underflow/overflow.
+
+I batches sono 0-indexed.
+
+Questo è utile se si sa che il programma inizia a comportarsi male dopo un certo numero di batch, in modo da poter avanzare velocemente fino a quell'area.
+direttamente a quell'area. Ecco un esempio di output troncato per questa configurazione:
+
+```
+                  *** 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
+[...]
+```
+
+Qui verrà scaricato un numero enorme di fotogrammi, tanti quanti sono le chiamate in avanti nel modello, quindi può essere o non essere quello che volete, ma a volte può essere più utile usarlo di un classico debugger. Per esempio, se il problema inizia a verificarsi a partire dal lotto numero 150. Quindi è possibile scaricare le tracce dei lotti 149 e 150 e confrontare i punti in cui i numeri hanno iniziato a divergere.
+
+È inoltre possibile specificare il numero di batch dopo il quale interrompere l'addestramento, con:
+
+```python
+debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1, 3], abort_after_batch_num=3)
+```
--- a/docs/source/it/model_sharing.mdx
+++ b/docs/source/it/model_sharing.mdx
@@ -0,0 +1,234 @@
+<!--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.
+-->
+
+# Condividi un modello
+
+Gli ultimi due tutorial ti hanno mostrato come puoi fare fine-tuning di un modello con PyTorch, Keras e 🤗 Accelerate per configurazioni distribuite. Il prossimo passo è quello di condividere il tuo modello con la community! In Hugging Face, crediamo nella condivisione della conoscenza e delle risorse in modo da democratizzare l'intelligenza artificiale per chiunque. Ti incoraggiamo a considerare di condividere il tuo modello con la community per aiutare altre persone a risparmiare tempo e risorse.
+
+In questo tutorial, imparerai due metodi per la condivisione di un modello trained o fine-tuned nel [Model Hub](https://huggingface.co/models):
+
+- Condividi in modo programmatico i tuoi file nell'Hub.
+- Trascina i tuoi file nell'Hub mediante interfaccia grafica.
+
+<iframe width="560" height="315" src="https://www.youtube.com/embed/XvSGPZFEjDY" title="YouTube video player"
+frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+picture-in-picture" allowfullscreen></iframe>
+
+<Tip>
+
+Per condividere un modello con la community, hai bisogno di un account su [huggingface.co](https://huggingface.co/join). Puoi anche unirti ad un'organizzazione esistente o crearne una nuova.
+
+</Tip>
+
+## Caratteristiche dei repository
+
+Ogni repository nel Model Hub si comporta come un tipico repository di GitHub. I nostri repository offrono il versionamento, la cronologia dei commit, e la possibilità di visualizzare le differenze.
+
+Il versionamento all'interno del Model Hub è basato su git e [git-lfs](https://git-lfs.github.com/). In altre parole, puoi trattare un modello come un unico repository, consentendo un maggiore controllo degli accessi e maggiore scalabilità. Il controllo delle versioni consente *revisions*, un metodo per appuntare una versione specifica di un modello con un hash di commit, un tag o un branch.
+
+Come risultato, puoi caricare una specifica versione di un modello con il parametro `revision`:
+
+```py
+>>> model = AutoModel.from_pretrained(
+...     "julien-c/EsperBERTo-small", revision="v2.0.1"  # nome di un tag, di un branch, o commit hash
+... )
+```
+
+Anche i file possono essere modificati facilmente in un repository ed è possibile visualizzare la cronologia dei commit e le differenze:
+
+![vis_diff](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/vis_diff.png)
+
+## Configurazione
+
+Prima di condividere un modello nell'Hub, hai bisogno delle tue credenziali di Hugging Face. Se hai accesso ad un terminale, esegui il seguente comando nell'ambiente virtuale in cui è installata la libreria 🤗 Transformers. Questo memorizzerà il tuo token di accesso nella cartella cache di Hugging Face (di default `~/.cache/`):
+
+```bash
+huggingface-cli login
+```
+
+Se stai usando un notebook come Jupyter o Colaboratory, assicurati di avere la libreria [`huggingface_hub`](https://huggingface.co/docs/hub/adding-a-library) installata. Questa libreria ti permette di interagire in maniera programmatica con l'Hub.
+
+```bash
+pip install huggingface_hub
+```
+
+Utilizza `notebook_login` per accedere all'Hub, e segui il link [qui](https://huggingface.co/settings/token) per generare un token con cui effettuare il login:
+
+```py
+>>> from huggingface_hub import notebook_login
+
+>>> notebook_login()
+```
+
+## Converti un modello per tutti i framework
+
+Per assicurarti che il tuo modello possa essere utilizzato da persone che lavorano con un framework differente, ti raccomandiamo di convertire e caricare il tuo modello sia con i checkpoint di PyTorch che con quelli di TensorFlow. Anche se è possibile caricare il modello da un framework diverso, se si salta questo passaggio, il caricamento sarà più lento perché 🤗 Transformers ha bisogno di convertire i checkpoint al momento.
+
+Convertire un checkpoint per un altro framework è semplice. Assicurati di avere PyTorch e TensorFlow installati (vedi [qui](installation) per le istruzioni d'installazione), e poi trova il modello specifico per il tuo compito nell'altro framework.
+
+<frameworkcontent>
+<pt>
+Specifica `from_tf=True` per convertire un checkpoint da TensorFlow a PyTorch:
+
+```py
+>>> pt_model = DistilBertForSequenceClassification.from_pretrained(
+...     "path/verso/il-nome-magnifico-che-hai-scelto", from_tf=True
+... )
+>>> pt_model.save_pretrained("path/verso/il-nome-magnifico-che-hai-scelto")
+```
+</pt>
+<tf>
+Specifica `from_pt=True` per convertire un checkpoint da PyTorch a TensorFlow:
+
+```py
+>>> tf_model = TFDistilBertForSequenceClassification.from_pretrained(
+...     "path/verso/il-nome-magnifico-che-hai-scelto", from_pt=True
+... )
+```
+
+Poi puoi salvare il tuo nuovo modello in TensorFlow con il suo nuovo checkpoint:
+
+```py
+>>> tf_model.save_pretrained("path/verso/il-nome-magnifico-che-hai-scelto")
+```
+</tf>
+<jax>
+Se un modello è disponibile in Flax, puoi anche convertire un checkpoint da PyTorch a Flax:
+
+```py
+>>> flax_model = FlaxDistilBertForSequenceClassification.from_pretrained(
+...     "path/verso/il-nome-magnifico-che-hai-scelto", from_pt=True
+... )
+```
+</jax>
+</frameworkcontent>
+
+## Condividi un modello durante il training
+
+<frameworkcontent>
+<pt>
+<Youtube id="Z1-XMy-GNLQ"/>
+
+Condividere un modello nell'Hub è tanto semplice quanto aggiungere un parametro extra o un callback. Ricorda dal [tutorial sul fine-tuning](training), la classe [`TrainingArguments`] è dove specifichi gli iperparametri e le opzioni addizionali per l'allenamento. Una di queste opzioni di training include l'abilità di condividere direttamente un modello nell'Hub. Imposta `push_to_hub=True` in [`TrainingArguments`]:
+
+```py
+>>> training_args = TrainingArguments(output_dir="il-mio-bellissimo-modello", push_to_hub=True)
+```
+
+Passa gli argomenti per il training come di consueto al [`Trainer`]:
+
+```py
+>>> trainer = Trainer(
+...     model=model,
+...     args=training_args,
+...     train_dataset=small_train_dataset,
+...     eval_dataset=small_eval_dataset,
+...     compute_metrics=compute_metrics,
+... )
+```
+
+Dopo aver effettuato il fine-tuning del tuo modello, chiama [`~transformers.Trainer.push_to_hub`] sul [`Trainer`] per condividere il modello allenato nell'Hub. 🤗 Transformers aggiungerà in modo automatico persino gli iperparametri, i risultati del training e le versioni del framework alla scheda del tuo modello (model card, in inglese)!
+
+```py
+>>> trainer.push_to_hub()
+```
+</pt>
+<tf>
+Condividi un modello nell'Hub con [`PushToHubCallback`]. Nella funzione [`PushToHubCallback`], aggiungi:
+
+- Una directory di output per il tuo modello.
+- Un tokenizer.
+- L'`hub_model_id`, che è il tuo username sull'Hub e il nome del modello.
+
+```py
+>>> from transformers.keras.callbacks import PushToHubCallback
+
+>>> push_to_hub_callback = PushToHubCallback(
+...     output_dir="./il_path_dove_salvare_il_tuo_modello",
+...     tokenizer=tokenizer,
+...     hub_model_id="il-tuo-username/il-mio-bellissimo-modello",
+... )
+```
+
+Aggiungi il callback a [`fit`](https://keras.io/api/models/model_training_apis/), e 🤗 Transformers caricherà il modello allenato nell'Hub:
+
+```py
+>>> model.fit(tf_train_dataset, validation_data=tf_validation_dataset, epochs=3, callbacks=push_to_hub_callback)
+```
+</tf>
+</frameworkcontent>
+
+## Utilizzare la funzione `push_to_hub`
+
+Puoi anche chiamare `push_to_hub` direttamente sul tuo modello per caricarlo nell'Hub.
+
+Specifica il nome del tuo modello in `push_to_hub`:
+
+```py
+>>> pt_model.push_to_hub("il-mio-bellissimo-modello")
+```
+
+Questo crea un repository sotto il proprio username con il nome del modello `il-mio-bellissimo-modello`. Ora chiunque può caricare il tuo modello con la funzione `from_pretrained`:
+
+```py
+>>> from transformers import AutoModel
+
+>>> model = AutoModel.from_pretrained("il-tuo-username/il-mio-bellissimo-modello")
+```
+
+Se fai parte di un'organizzazione e vuoi invece condividere un modello sotto il nome dell'organizzazione, aggiungi il parametro `organization`:
+
+```py
+>>> pt_model.push_to_hub("il-mio-bellissimo-modello", organization="la-mia-fantastica-org")
+```
+
+La funzione `push_to_hub` può essere anche utilizzata per aggiungere altri file al repository del modello. Per esempio, aggiungi un tokenizer ad un repository di un modello:
+
+```py
+>>> tokenizer.push_to_hub("il-mio-bellissimo-modello")
+```
+
+O magari potresti voler aggiungere la versione di TensorFlow del tuo modello PyTorch a cui hai fatto fine-tuning:
+
+```py
+>>> tf_model.push_to_hub("il-mio-bellissimo-modello")
+```
+
+Ora quando navighi nel tuo profilo Hugging Face, dovresti vedere il tuo repository del modello appena creato. Premendo sulla scheda **Files** vengono visualizzati tutti i file caricati nel repository.
+
+Per maggiori dettagli su come creare e caricare file ad un repository, fai riferimento alla documentazione [qui](https://huggingface.co/docs/hub/how-to-upstream).
+
+## Carica un modello utilizzando l'interfaccia web
+
+Chi preferisce un approccio senza codice può caricare un modello tramite l'interfaccia web dell'hub. Visita [huggingface.co/new](https://huggingface.co/new) per creare un nuovo repository:
+
+![new_model_repo](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/new_model_repo.png)
+
+Da qui, aggiungi alcune informazioni sul tuo modello:
+
+- Seleziona il/la **owner** del repository. Puoi essere te o qualunque organizzazione di cui fai parte.
+- Scegli un nome per il tuo modello, il quale sarà anche il nome del repository.
+- Scegli se il tuo modello è pubblico o privato.
+- Specifica la licenza utilizzata per il tuo modello.
+
+Ora premi sulla scheda **Files** e premi sul pulsante **Add file** per caricare un nuovo file al tuo repository. Trascina poi un file per caricarlo e aggiungere un messaggio di commit.
+
+![upload_file](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/upload_file.png)
+
+## Aggiungi una scheda del modello
+
+Per assicurarti che chiunque possa comprendere le abilità, limitazioni, i potenziali bias e le considerazioni etiche del tuo modello, per favore aggiungi una scheda del modello (model card, in inglese) al tuo repository. La scheda del modello è definita nel file `README.md`. Puoi aggiungere una scheda del modello:
+
+* Creando manualmente e caricando un file `README.md`.
+* Premendo sul pulsante **Edit model card** nel repository del tuo modello.
+
+Dai un'occhiata alla [scheda del modello](https://huggingface.co/distilbert-base-uncased) di DistilBert per avere un buon esempio del tipo di informazioni che una scheda di un modello deve includere. Per maggiori dettagli legati ad altre opzioni che puoi controllare nel file `README.md`, come l'impatto ambientale o widget di esempio, fai riferimento alla documentazione [qui](https://huggingface.co/docs/hub/model-repos).
--- a/docs/source/it/multilingual.mdx
+++ b/docs/source/it/multilingual.mdx
@@ -0,0 +1,174 @@
+<!--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.
+-->
+
+# Modelli multilingue per l'inferenza
+
+[[open-in-colab]]
+
+Ci sono diversi modelli multilingue in 🤗 Transformers, e il loro utilizzo per l'inferenza differisce da quello dei modelli monolingua. Non *tutti* gli utilizzi dei modelli multilingue sono però diversi. Alcuni modelli, come [bert-base-multilingual-uncased](https://huggingface.co/bert-base-multilingual-uncased), possono essere usati come un modello monolingua. Questa guida ti mostrerà come utilizzare modelli multilingue che utilizzano un modo diverso per fare l'inferenza.
+
+## XLM
+
+XLM ha dieci diversi checkpoint, di cui solo uno è monolingua. I nove checkpoint rimanenti possono essere suddivisi in due categorie: i checkpoint che utilizzano i language embeddings e quelli che non li utilizzano.
+
+### XLM con language embeddings
+
+I seguenti modelli XLM utilizzano gli embeddings linguistici per specificare la lingua utilizzata per l'inferenza:
+
+- `xlm-mlm-ende-1024` (Modellazione mascherata del linguaggio (Masked language modeling, in inglese), Inglese-Tedesco)
+- `xlm-mlm-enfr-1024` (Modellazione mascherata del linguaggio, Inglese-Francese)
+- `xlm-mlm-enro-1024` (Modellazione mascherata del linguaggio, Inglese-Rumeno)
+- `xlm-mlm-xnli15-1024` (Modellazione mascherata del linguaggio, lingue XNLI)
+- `xlm-mlm-tlm-xnli15-1024` (Modellazione mascherata del linguaggio + traduzione, lingue XNLI)
+- `xlm-clm-enfr-1024` (Modellazione causale del linguaggio, Inglese-Francese)
+- `xlm-clm-ende-1024` (Modellazione causale del linguaggio, Inglese-Tedesco)
+
+Gli embeddings linguistici sono rappresentati come un tensore delle stesse dimensioni dell' `input_ids` passato al modello. I valori in questi tensori dipendono dal linguaggio usato e sono identificati dagli attributi `lang2id` e `id2lang` del tokenizer.
+
+In questo esempio, carica il checkpoint `xlm-clm-enfr-1024` (Modellazione causale del linguaggio, Inglese-Francese):
+
+```py
+>>> import torch
+>>> from transformers import XLMTokenizer, XLMWithLMHeadModel
+
+>>> tokenizer = XLMTokenizer.from_pretrained("xlm-clm-enfr-1024")
+>>> model = XLMWithLMHeadModel.from_pretrained("xlm-clm-enfr-1024")
+```
+
+L'attributo `lang2id` del tokenizer mostra il linguaggio del modello e il suo ids:
+
+```py
+>>> print(tokenizer.lang2id)
+{'en': 0, 'fr': 1}
+```
+
+Poi, crea un esempio di input:
+
+```py
+>>> input_ids = torch.tensor([tokenizer.encode("Wikipedia was used to")])  # batch size of 1
+```
+
+Imposta l'id del linguaggio a `"en"` e usalo per definire il language embedding. Il language embedding è un tensore riempito con `0` perché questo è il language id per l'inglese. Questo tensore dovrebbe avere la stessa dimensione di `input_ids`.
+
+```py
+>>> language_id = tokenizer.lang2id["en"]  # 0
+>>> langs = torch.tensor([language_id] * input_ids.shape[1])  # torch.tensor([0, 0, 0, ..., 0])
+
+>>> # We reshape it to be of size (batch_size, sequence_length)
+>>> langs = langs.view(1, -1)  # is now of shape [1, sequence_length] (we have a batch size of 1)
+```
+
+Adesso puoi inserire `input_ids` e language embedding nel modello:
+
+```py
+>>> outputs = model(input_ids, langs=langs)
+```
+
+Lo script [run_generation.py](https://github.com/huggingface/transformers/tree/main/examples/pytorch/text-generation/run_generation.py) può generare testo tramite i language embeddings usando i checkpoints `xlm-clm`.
+
+### XLM senza language embeddings
+
+I seguenti modelli XLM non richiedono l'utilizzo dei language embeddings per fare inferenza:
+
+- `xlm-mlm-17-1280` (Modellazione mascherata del linguaggio, 17 lingue)
+- `xlm-mlm-100-1280` (Modellazione mascherata del linguaggio, 100 lingue)
+
+Questi modelli sono utilizzati per rappresentazioni generiche di frasi, a differenza dei precedenti checkpoints XML.
+
+## BERT
+
+Il seguente modello BERT può essere usato per compiti multilingue:
+
+- `bert-base-multilingual-uncased` (Modellazione mascherata del linguaggio + Previsione della prossima frase, 102 lingue)
+- `bert-base-multilingual-cased` (Modellazione mascherata del linguaggio + Previsione della prossima frase, 104 lingue)
+
+Questi modelli non richiedono language embeddings per fare inferenza. Riescono ad identificare il linguaggio dal contesto e inferire di conseguenza.
+
+## XLM-RoBERTa
+
+Il seguente modello XLM-RoBERTa può essere usato per compiti multilingue:
+
+- `xlm-roberta-base` (Modellazione mascherata del linguaggio, 100 lingue)
+- `xlm-roberta-large` (Modellazione mascherata del linguaggio, 100 lingue)
+
+XLM-RoBERTa è stato addestrato su 2.5TB di dati CommonCrawl appena creati e puliti in 100 lingue. Offre notevoli vantaggi rispetto ai modelli multilingue rilasciati in precedenza, come mBERT o XLM, in compiti come la classificazione, l'etichettatura delle sequenze e la risposta alle domande.
+
+## M2M100
+
+Il seguente modello M2M100 può essere usato per compiti multilingue:
+
+- `facebook/m2m100_418M` (Traduzione)
+- `facebook/m2m100_1.2B` (Traduzione)
+
+In questo esempio, carica il checkpoint `facebook/m2m100_418M`  per tradurre dal cinese all'inglese. Puoi impostare la lingua di partenza nel tokenizer:
+
+```py
+>>> from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer
+
+>>> en_text = "Do not meddle in the affairs of wizards, for they are subtle and quick to anger."
+>>> chinese_text = "不要插手巫師的事務, 因為他們是微妙的, 很快就會發怒."
+
+>>> tokenizer = M2M100Tokenizer.from_pretrained("facebook/m2m100_418M", src_lang="zh")
+>>> model = M2M100ForConditionalGeneration.from_pretrained("facebook/m2m100_418M")
+```
+
+Applica il tokenizer al testo:
+
+```py
+>>> encoded_zh = tokenizer(chinese_text, return_tensors="pt")
+```
+
+M2M100 forza l'id della lingua obiettivo come primo token generato per tradurre nella lingua obiettivo. Imposta il parametro `forced_bos_token_id` a `en` nel metodo `generate` per tradurre in inglese:
+
+```py
+>>> generated_tokens = model.generate(**encoded_zh, forced_bos_token_id=tokenizer.get_lang_id("en"))
+>>> tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
+'Do not interfere with the matters of the witches, because they are delicate and will soon be angry.'
+```
+
+## MBart
+
+Il seguente modello MBart può essere usato per compiti multilingue:
+
+- `facebook/mbart-large-50-one-to-many-mmt` (Traduzione automatica multilingue uno-a-molti, 50 lingue)
+- `facebook/mbart-large-50-many-to-many-mmt` (Traduzione automatica multilingue molti-a-molti, 50 lingue)
+- `facebook/mbart-large-50-many-to-one-mmt` (Traduzione automatica multilingue molti-a-uno, 50 lingue)
+- `facebook/mbart-large-50` (Traduzione multilingue, 50 lingue)
+- `facebook/mbart-large-cc25`
+
+In questo esempio, carica il checkpoint `facebook/mbart-large-50-many-to-many-mmt` per tradurre dal finlandese all'inglese. Puoi impostare la lingua di partenza nel tokenizer:
+
+```py
+>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+
+>>> en_text = "Do not meddle in the affairs of wizards, for they are subtle and quick to anger."
+>>> fi_text = "Älä sekaannu velhojen asioihin, sillä ne ovat hienovaraisia ja nopeasti vihaisia."
+
+>>> tokenizer = AutoTokenizer.from_pretrained("facebook/mbart-large-50-many-to-many-mmt", src_lang="fi_FI")
+>>> model = AutoModelForSeq2SeqLM.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
+```
+
+Applica il tokenizer sul testo:
+
+```py
+>>> encoded_en = tokenizer(en_text, return_tensors="pt")
+```
+
+MBart forza l'id della lingua obiettivo come primo token generato per tradurre nella lingua obiettivo. Imposta il parametro `forced_bos_token_id` a `en` nel metodo `generate` per tradurre in inglese:
+
+```py
+>>> generated_tokens = model.generate(**encoded_en, forced_bos_token_id=tokenizer.lang_code_to_id("en_XX"))
+>>> tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
+"Don't interfere with the wizard's affairs, because they are subtle, will soon get angry."
+```
+
+Se stai usando il checkpoint `facebook/mbart-large-50-many-to-one-mmt`, non hai bisogno di forzare l'id della lingua obiettivo come primo token generato altrimenti l'uso è lo stesso.
--- a/docs/source/it/preprocessing.mdx
+++ b/docs/source/it/preprocessing.mdx
@@ -0,0 +1,489 @@
+<!--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.
+-->
+
+# Preprocess
+
+[[open-in-colab]]
+
+Prima di poter usare i dati in un modello, bisogna processarli in un formato accettabile per quest'ultimo. Un modello non comprende il testo grezzo, le immagini o l'audio. Bisogna convertire questi input in numeri e assemblarli all'interno di tensori. In questa esercitazione, tu potrai:
+
+* Preprocessare dati testuali con un tokenizer.
+* Preprocessare immagini o dati audio con un estrattore di caratteristiche.
+* Preprocessare dati per attività multimodali mediante un processore.
+
+## NLP
+
+<Youtube id="Yffk5aydLzg"/>
+
+Lo strumento principale per processare dati testuali è un [tokenizer](main_classes/tokenizer). Un tokenizer inizia separando il testo in *tokens* secondo una serie di regole. I tokens sono convertiti in numeri, questi vengono utilizzati per costruire i tensori di input del modello. Anche altri input addizionali se richiesti dal modello vengono aggiunti dal tokenizer.
+
+<Tip>
+
+Se stai pensando si utilizzare un modello preaddestrato, è importante utilizzare il tokenizer preaddestrato associato. Questo assicura che il testo sia separato allo stesso modo che nel corpus usato per l'addestramento, e venga usata la stessa mappatura tokens-to-index (solitamente indicato come il *vocabolario*) come nel preaddestramento.
+
+</Tip>
+
+Iniziamo subito caricando un tokenizer preaddestrato con la classe [`AutoTokenizer`]. Questo scarica il *vocabolario* usato quando il modello è stato preaddestrato.
+
+### Tokenize
+
+Carica un tokenizer preaddestrato con [`AutoTokenizer.from_pretrained`]:
+
+```py
+>>> from transformers import AutoTokenizer
+
+>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
+```
+
+Poi inserisci le tue frasi nel tokenizer:
+
+```py
+>>> encoded_input = tokenizer("Do not meddle in the affairs of wizards, for they are subtle and quick to anger.")
+>>> print(encoded_input)
+{'input_ids': [101, 2079, 2025, 19960, 10362, 1999, 1996, 3821, 1997, 16657, 1010, 2005, 2027, 2024, 11259, 1998, 4248, 2000, 4963, 1012, 102], 
+ 'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 
+ 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
+```
+
+Il tokenizer restituisce un dizionario contenente tre oggetti importanti:
+
+* [input_ids](glossary#input-ids) sono gli indici che corrispondono ad ogni token nella frase.
+* [attention_mask](glossary#attention-mask) indicata se un token deve essere elaborato o no.
+* [token_type_ids](glossary#token-type-ids) identifica a quale sequenza appartiene un token se è presente più di una sequenza.
+
+Si possono decodificare gli `input_ids` per farsi restituire l'input originale:
+
+```py
+>>> tokenizer.decode(encoded_input["input_ids"])
+'[CLS] Do not meddle in the affairs of wizards, for they are subtle and quick to anger. [SEP]'
+```
+
+Come si può vedere, il tokenizer aggiunge due token speciali - `CLS` e `SEP` (classificatore e separatore) - alla frase. Non tutti i modelli hanno bisogno dei token speciali, ma se servono, il tokenizer li aggiungerà automaticamente.
+
+Se ci sono più frasi che vuoi processare, passale come una lista al tokenizer:
+
+```py
+>>> batch_sentences = [
+...     "But what about second breakfast?",
+...     "Don't think he knows about second breakfast, Pip.",
+...     "What about elevensies?",
+... ]
+>>> encoded_inputs = tokenizer(batch_sentences)
+>>> print(encoded_inputs)
+{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102], 
+               [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102], 
+               [101, 1327, 1164, 5450, 23434, 136, 102]], 
+ 'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0], 
+                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 
+                    [0, 0, 0, 0, 0, 0, 0]], 
+ 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1], 
+                    [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
+                    [1, 1, 1, 1, 1, 1, 1]]}
+```
+
+### Pad
+
+Questo è un argomento importante. Quando processi un insieme di frasi potrebbero non avere tutte la stessa lunghezza. Questo è un problema perchè i tensori, in input del modello, devono avere dimensioni uniformi. Il padding è una strategia per assicurarsi che i tensori siano rettangolari aggiungendo uno speciale *padding token* alle frasi più corte.
+
+Imposta il parametro `padding` a `True` per imbottire le frasi più corte nel gruppo in modo che combacino con la massima lunghezza presente:
+
+```py
+>>> batch_sentences = [
+...     "But what about second breakfast?",
+...     "Don't think he knows about second breakfast, Pip.",
+...     "What about elevensies?",
+... ]
+>>> encoded_input = tokenizer(batch_sentences, padding=True)
+>>> print(encoded_input)
+{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0], 
+               [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102], 
+               [101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]], 
+ 'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 
+                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 
+                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]], 
+ 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], 
+                    [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
+                    [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]]}
+```
+
+Nota che il tokenizer aggiunge alle sequenze degli `0` perchè sono troppo corte!
+
+### Truncation
+
+L'altra faccia della medaglia è che avolte le sequenze possono essere troppo lunghe per essere gestite dal modello. In questo caso, avrai bisogno di troncare la sequenza per avere una lunghezza minore.
+
+Imposta il parametro `truncation` a `True` per troncare una sequenza alla massima lunghezza accettata dal modello:
+
+```py
+>>> batch_sentences = [
+...     "But what about second breakfast?",
+...     "Don't think he knows about second breakfast, Pip.",
+...     "What about elevensies?",
+... ]
+>>> encoded_input = tokenizer(batch_sentences, padding=True, truncation=True)
+>>> print(encoded_input)
+{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0], 
+               [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102], 
+               [101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]], 
+ 'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 
+                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 
+                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]], 
+ 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0], 
+                    [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
+                    [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]]}
+```
+
+### Costruire i tensori
+
+Infine, vuoi che il tokenizer restituisca i tensori prodotti dal modello.
+
+Imposta il parametro `return_tensors` su `pt` per PyTorch, o `tf` per TensorFlow:
+
+```py
+>>> batch_sentences = [
+...     "But what about second breakfast?",
+...     "Don't think he knows about second breakfast, Pip.",
+...     "What about elevensies?",
+... ]
+>>> encoded_input = tokenizer(batch, padding=True, truncation=True, return_tensors="pt")
+>>> print(encoded_input)
+{'input_ids': tensor([[  101,   153,  7719, 21490,  1122,  1114,  9582,  1623,   102],
+                      [  101,  5226,  1122,  9649,  1199,  2610,  1236,   102,     0]]), 
+ 'token_type_ids': tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0],
+                           [0, 0, 0, 0, 0, 0, 0, 0, 0]]), 
+ 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1],
+                           [1, 1, 1, 1, 1, 1, 1, 1, 0]])}
+===PT-TF-SPLIT===
+>>> batch_sentences = [
+...     "But what about second breakfast?",
+...     "Don't think he knows about second breakfast, Pip.",
+...     "What about elevensies?",
+... ]
+>>> encoded_input = tokenizer(batch, padding=True, truncation=True, return_tensors="tf")
+>>> print(encoded_input)
+{'input_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
+array([[  101,   153,  7719, 21490,  1122,  1114,  9582,  1623,   102],
+       [  101,  5226,  1122,  9649,  1199,  2610,  1236,   102,     0]],
+      dtype=int32)>, 
+ 'token_type_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
+array([[0, 0, 0, 0, 0, 0, 0, 0, 0],
+       [0, 0, 0, 0, 0, 0, 0, 0, 0]], dtype=int32)>, 
+ 'attention_mask': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
+array([[1, 1, 1, 1, 1, 1, 1, 1, 1],
+       [1, 1, 1, 1, 1, 1, 1, 1, 0]], dtype=int32)>}
+```
+
+## Audio
+
+Gli input audio sono processati in modo differente rispetto al testo, ma l'obiettivo rimane lo stesso: creare sequenze numeriche che il modello può capire. Un [estrattore di caratteristiche](main_classes/feature_extractor) è progettato con lo scopo preciso di estrarre caratteristiche da immagini o dati audio grezzi e convertirli in tensori. Prima di iniziare, installa 🤗 Datasets per caricare un dataset audio e sperimentare:
+
+```bash
+pip install datasets
+```
+
+Carica il dataset [MInDS-14](https://huggingface.co/datasets/PolyAI/minds14) (vedi il 🤗 [Datasets tutorial](https://huggingface.co/docs/datasets/load_hub.html) per avere maggiori dettagli su come caricare un dataset):
+
+```py
+>>> from datasets import load_dataset, Audio
+
+>>> dataset = load_dataset("PolyAI/minds14", name="en-US", split="train")
+```
+
+Accedi al primo elemento della colonna `audio` per dare uno sguardo all'input. Richiamando la colonna `audio` sarà caricato automaticamente e ricampionato il file audio:
+
+```py
+>>> dataset[0]["audio"]
+{'array': array([ 0.        ,  0.00024414, -0.00024414, ..., -0.00024414,
+         0.        ,  0.        ], dtype=float32),
+ 'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~JOINT_ACCOUNT/602ba55abb1e6d0fbce92065.wav',
+ 'sampling_rate': 8000}
+```
+
+Questo restituisce tre oggetti:
+
+* `array` è il segnale vocale caricato - e potenzialmente ricampionato - come vettore 1D.
+* `path` il percorso del file audio.
+* `sampling_rate` si riferisce al numero di campioni del segnale vocale misurati al secondo.
+
+### Ricampionamento
+
+Per questo tutorial, puoi usare il modello [Wav2Vec2](https://huggingface.co/facebook/wav2vec2-base). Come puoi vedere dalla model card, il modello Wav2Vec2 è preaddestrato su un campionamento vocale a 16kHz.È importante che la frequenza di campionamento dei tuoi dati audio combaci con la frequenza di campionamento del dataset usato per preaddestrare il modello. Se la frequenza di campionamento dei tuoi dati non è uguale dovrai ricampionare i tuoi dati audio.
+
+Per esempio, il dataset [MInDS-14](https://huggingface.co/datasets/PolyAI/minds14) ha una frequenza di campionamento di 8000kHz. Utilizzando il modello Wav2Vec2 su questo dataset, alzala a 16kHz:
+
+```py
+>>> dataset = load_dataset("PolyAI/minds14", name="en-US", split="train")
+>>> dataset[0]["audio"]
+{'array': array([ 0.        ,  0.00024414, -0.00024414, ..., -0.00024414,
+         0.        ,  0.        ], dtype=float32),
+ 'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~JOINT_ACCOUNT/602ba55abb1e6d0fbce92065.wav',
+ 'sampling_rate': 8000}
+```
+
+1. Usa il metodo di 🤗 Datasets' [`cast_column`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.cast_column) per alzare la frequenza di campionamento a 16kHz:
+
+```py
+>>> dataset = dataset.cast_column("audio", Audio(sampling_rate=16_000))
+```
+
+2. Carica il file audio:
+
+```py
+>>> dataset[0]["audio"]
+{'array': array([ 2.3443763e-05,  2.1729663e-04,  2.2145823e-04, ...,
+         3.8356509e-05, -7.3497440e-06, -2.1754686e-05], dtype=float32),
+ 'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~JOINT_ACCOUNT/602ba55abb1e6d0fbce92065.wav',
+ 'sampling_rate': 16000}
+```
+
+Come puoi notare, la `sampling_rate` adesso è 16kHz!
+
+### Feature extractor
+
+Il prossimo passo è caricare un estrattore di caratteristiche per normalizzare e fare padding sull'input. Quando applichiamo il padding sui dati testuali, uno `0` è aggiunto alle sequenze più brevi. La stessa idea si applica ai dati audio, l'estrattore di caratteristiche per gli audio aggiungerà uno `0` - interpretato come silenzio - agli `array`.
+
+Carica l'estrattore delle caratteristiche con [`AutoFeatureExtractor.from_pretrained`]:
+
+```py
+>>> from transformers import AutoFeatureExtractor
+
+>>> feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base")
+```
+
+Inserisci l' `array` audio nell'estrattore delle caratteristiche. Noi raccomandiamo sempre di aggiungere il parametro `sampling_rate` nell'estrattore delle caratteristiche per correggere meglio qualche errore, dovuto ai silenzi, che potrebbe verificarsi.
+
+```py
+>>> audio_input = [dataset[0]["audio"]["array"]]
+>>> feature_extractor(audio_input, sampling_rate=16000)
+{'input_values': [array([ 3.8106556e-04,  2.7506407e-03,  2.8015103e-03, ...,
+        5.6335266e-04,  4.6588284e-06, -1.7142107e-04], dtype=float32)]}
+```
+
+### Pad e truncate
+
+Come per il tokenizer, puoi applicare le operazioni padding o truncation per manipolare sequenze di variabili a lotti. Dai uno sguaro alla lunghezza delle sequenze di questi due campioni audio:
+
+```py
+>>> dataset[0]["audio"]["array"].shape
+(173398,)
+
+>>> dataset[1]["audio"]["array"].shape
+(106496,)
+```
+
+Come puoi vedere, il primo campione ha una sequenza più lunga del secondo. Crea una funzione che preprocesserà il dataset. Specifica una lunghezza massima del campione, e l'estrattore di features si occuperà di riempire o troncare la sequenza per coincidervi:
+
+```py
+>>> def preprocess_function(examples):
+...     audio_arrays = [x["array"] for x in examples["audio"]]
+...     inputs = feature_extractor(
+...         audio_arrays,
+...         sampling_rate=16000,
+...         padding=True,
+...         max_length=100000,
+...         truncation=True,
+...     )
+...     return inputs
+```
+
+Applica la funzione ai primi esempi nel dataset:
+
+```py
+>>> processed_dataset = preprocess_function(dataset[:5])
+```
+
+Adesso guarda la lunghezza dei campioni elaborati:
+
+```py
+>>> processed_dataset["input_values"][0].shape
+(100000,)
+
+>>> processed_dataset["input_values"][1].shape
+(100000,)
+```
+
+La lunghezza dei campioni adesso coincide con la massima lunghezza impostata nelle funzione.
+
+## Vision
+
+Un estrattore di caratteristiche si può usare anche per processare immagini e per compiti di visione. Ancora una volta, l'obiettivo è convertire l'immagine grezza in un lotto di tensori come input.
+
+Carica il dataset [food101](https://huggingface.co/datasets/food101) per questa esercitazione. Usa il parametro `split` di 🤗 Datasets  per caricare solo un piccolo campione dal dataset di addestramento poichè il set di dati è molto grande:
+
+```py
+>>> from datasets import load_dataset
+
+>>> dataset = load_dataset("food101", split="train[:100]")
+```
+
+Secondo passo, dai uno sguardo alle immagini usando la caratteristica [`Image`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=image#datasets.Image) di 🤗 Datasets:
+
+```py
+>>> dataset[0]["image"]
+```
+
+![vision-preprocess-tutorial.png](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/vision-preprocess-tutorial.png)
+
+### Feature extractor
+
+Carica l'estrattore di caratteristiche [`AutoFeatureExtractor.from_pretrained`]:
+
+```py
+>>> from transformers import AutoFeatureExtractor
+
+>>> feature_extractor = AutoFeatureExtractor.from_pretrained("google/vit-base-patch16-224")
+```
+
+### Data augmentation
+
+Per le attività di visione, è usuale aggiungere alcuni tipi di data augmentation alle immagini come parte del preprocessing. Puoi aggiungere augmentations con qualsiasi libreria che preferisci, ma in questa esercitazione, userai il modulo [`transforms`](https://pytorch.org/vision/stable/transforms.html) di torchvision.
+
+1. Normalizza l'immagine e usa [`Compose`](https://pytorch.org/vision/master/generated/torchvision.transforms.Compose.html) per concatenare alcune trasformazioni - [`RandomResizedCrop`](https://pytorch.org/vision/main/generated/torchvision.transforms.RandomResizedCrop.html) e [`ColorJitter`](https://pytorch.org/vision/main/generated/torchvision.transforms.ColorJitter.html) - insieme:
+
+```py
+>>> from torchvision.transforms import Compose, Normalize, RandomResizedCrop, ColorJitter, ToTensor
+
+>>> normalize = Normalize(mean=feature_extractor.image_mean, std=feature_extractor.image_std)
+>>> _transforms = Compose(
+...     [RandomResizedCrop(feature_extractor.size), ColorJitter(brightness=0.5, hue=0.5), ToTensor(), normalize]
+... )
+```
+
+2. Il modello accetta [`pixel_values`](model_doc/visionencoderdecoder#transformers.VisionEncoderDecoderModel.forward.pixel_values) come input. Questo valore è generato dall'estrattore di caratteristiche. Crea una funzione che genera `pixel_values` dai transforms:
+
+```py
+>>> def transforms(examples):
+...     examples["pixel_values"] = [_transforms(image.convert("RGB")) for image in examples["image"]]
+...     return examples
+```
+
+3. Poi utilizza 🤗 Datasets [`set_transform`](https://huggingface.co/docs/datasets/process.html#format-transform)per applicare al volo la trasformazione:
+
+```py
+>>> dataset.set_transform(transforms)
+```
+
+4. Adesso quando accedi all'immagine, puoi notare che l'estrattore di caratteristiche ha aggiunto `pixel_values` allo schema di input:
+
+```py
+>>> dataset[0]["image"]
+{'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=384x512 at 0x7F1A7B0630D0>,
+ 'label': 6,
+ 'pixel_values': tensor([[[ 0.0353,  0.0745,  0.1216,  ..., -0.9922, -0.9922, -0.9922],
+          [-0.0196,  0.0667,  0.1294,  ..., -0.9765, -0.9843, -0.9922],
+          [ 0.0196,  0.0824,  0.1137,  ..., -0.9765, -0.9686, -0.8667],
+          ...,
+          [ 0.0275,  0.0745,  0.0510,  ..., -0.1137, -0.1216, -0.0824],
+          [ 0.0667,  0.0824,  0.0667,  ..., -0.0588, -0.0745, -0.0980],
+          [ 0.0353,  0.0353,  0.0431,  ..., -0.0039, -0.0039, -0.0588]],
+ 
+         [[ 0.2078,  0.2471,  0.2863,  ..., -0.9451, -0.9373, -0.9451],
+          [ 0.1608,  0.2471,  0.3098,  ..., -0.9373, -0.9451, -0.9373],
+          [ 0.2078,  0.2706,  0.3020,  ..., -0.9608, -0.9373, -0.8275],
+          ...,
+          [-0.0353,  0.0118, -0.0039,  ..., -0.2392, -0.2471, -0.2078],
+          [ 0.0196,  0.0353,  0.0196,  ..., -0.1843, -0.2000, -0.2235],
+          [-0.0118, -0.0039, -0.0039,  ..., -0.0980, -0.0980, -0.1529]],
+ 
+         [[ 0.3961,  0.4431,  0.4980,  ..., -0.9216, -0.9137, -0.9216],
+          [ 0.3569,  0.4510,  0.5216,  ..., -0.9059, -0.9137, -0.9137],
+          [ 0.4118,  0.4745,  0.5216,  ..., -0.9137, -0.8902, -0.7804],
+          ...,
+          [-0.2314, -0.1922, -0.2078,  ..., -0.4196, -0.4275, -0.3882],
+          [-0.1843, -0.1686, -0.2000,  ..., -0.3647, -0.3804, -0.4039],
+          [-0.1922, -0.1922, -0.1922,  ..., -0.2941, -0.2863, -0.3412]]])}
+```
+
+Di seguito come si vede l'immagine dopo la fase di preprocessing. Come ci si aspetterebbe dalle trasformazioni applicate, l'immagine è stata ritagliata in modo casuale e le proprietà del colore sono diverse.
+
+```py
+>>> import numpy as np
+>>> import matplotlib.pyplot as plt
+
+>>> img = dataset[0]["pixel_values"]
+>>> plt.imshow(img.permute(1, 2, 0))
+```
+
+![preprocessed_image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/preprocessed_image.png)
+
+## Multimodal
+
+Per attività multimodali userai una combinazione di tutto quello che hai imparato poco fa e applicherai le tue competenze alla comprensione automatica del parlato (Automatic Speech Recognition -  ASR). Questo significa che avrai bisogno di:
+
+* Un estrattore delle caratteristiche per processare i dati audio.
+* Il Tokenizer per processare i testi.
+
+Ritorna sul datasere [LJ Speech](https://huggingface.co/datasets/lj_speech):
+
+```py
+>>> from datasets import load_dataset
+
+>>> lj_speech = load_dataset("lj_speech", split="train")
+```
+
+Visto che sei interessato solo alle colonne `audio` e `text`, elimina tutte le altre:
+
+```py
+>>> lj_speech = lj_speech.map(remove_columns=["file", "id", "normalized_text"])
+```
+
+Adesso guarda le colonne `audio` e `text`:
+
+```py
+>>> lj_speech[0]["audio"]
+{'array': array([-7.3242188e-04, -7.6293945e-04, -6.4086914e-04, ...,
+         7.3242188e-04,  2.1362305e-04,  6.1035156e-05], dtype=float32),
+ 'path': '/root/.cache/huggingface/datasets/downloads/extracted/917ece08c95cf0c4115e45294e3cd0dee724a1165b7fc11798369308a465bd26/LJSpeech-1.1/wavs/LJ001-0001.wav',
+ 'sampling_rate': 22050}
+
+>>> lj_speech[0]["text"]
+'Printing, in the only sense with which we are at present concerned, differs from most if not from all the arts and crafts represented in the Exhibition'
+```
+
+Ricorda dalla sezione precedente sull'elaborazione dei dati audio, tu dovresti sempre [ricampionare](preprocessing#audio) la frequenza di campionamento dei tuoi dati audio per farla coincidere con quella del dataset usato dal modello preaddestrato:
+
+```py
+>>> lj_speech = lj_speech.cast_column("audio", Audio(sampling_rate=16_000))
+```
+
+### Processor
+
+Un processor combina un estrattore di caratteristiche e un tokenizer. Carica un processor con [`AutoProcessor.from_pretrained]:
+
+```py
+>>> from transformers import AutoProcessor
+
+>>> processor = AutoProcessor.from_pretrained("facebook/wav2vec2-base-960h")
+```
+
+1. Crea una funzione che processi i dati audio in `input_values`, e tokenizza il testo in `labels`. Questi sono i tuoi input per il modello:
+
+```py
+>>> def prepare_dataset(example):
+...     audio = example["audio"]
+
+...     example["input_values"] = processor(audio["array"], sampling_rate=16000)
+
+...     with processor.as_target_processor():
+...         example["labels"] = processor(example["text"]).input_ids
+...     return example
+```
+
+2. Applica la funzione `prepare_dataset` ad un campione:
+
+```py
+>>> prepare_dataset(lj_speech[0])
+```
+
+Nota che il processor ha aggiunto `input_values` e `labels`. La frequenza di campionamento è stata corretta riducendola a 16kHz.
+
+Fantastico, ora dovresti essere in grado di preelaborare i dati per qualsiasi modalità e persino di combinare modalità diverse! Nella prossima esercitazione, impareremo a mettere a punto un modello sui dati appena pre-elaborati.
--- a/docs/source/it/run_scripts.mdx
+++ b/docs/source/it/run_scripts.mdx
@@ -0,0 +1,347 @@
+<!--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.
+-->
+
+# Addestramento con script
+
+Insieme ai [notebooks](./noteboks/README) 🤗 Transformers, ci sono anche esempi di script che dimostrano come addestrare un modello per un task con [PyTorch](https://github.com/huggingface/transformers/tree/main/examples/pytorch), [TensorFlow](https://github.com/huggingface/transformers/tree/main/examples/tensorflow), o [JAX/Flax](https://github.com/huggingface/transformers/tree/main/examples/flax).
+
+Troverai anche script che abbiamo usato nei nostri [progetti di ricerca](https://github.com/huggingface/transformers/tree/main/examples/research_projects) e [precedenti esempi](https://github.com/huggingface/transformers/tree/main/examples/legacy) a cui contribuisce per lo più la comunità. Questi script non sono attivamente mantenuti e richiedono una specifica versione di 🤗 Transformers che sarà molto probabilmente incompatibile con l'ultima versione della libreria.
+
+Non è dato per scontato che gli script di esempio funzionino senza apportare modifiche per ogni problema, bensì potrebbe essere necessario adattare lo script al tuo caso specifico. Per aiutarti in ciò, la maggioranza degli script espone le modalità di pre-processamento dei dati, consentendoti di modificare lo script come preferisci.
+
+Per qualsiasi feature che vorresti implementare in uno script d'esempio, per favore discutine nel [forum](https://discuss.huggingface.co/) o in un'[issue](https://github.com/huggingface/transformers/issues) prima di inviare una Pull Request. Mentre accogliamo con piacere la correzione di bug, è più improbabile che faremo la stessa con una PR che aggiunge funzionalità sacrificando la leggibilità. 
+
+Questa guida ti mostrerà come eseguire uno script di esempio relativo al task di summarization in [PyTorch](https://github.com/huggingface/transformers/tree/main/examples/pytorch/summarization) e [TensorFlow](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/summarization). Tutti gli esempi funzioneranno con entrambi i framework a meno che non sia specificato altrimenti. 
+
+## Installazione
+
+Per eseguire con successo l'ultima versione degli script di esempio, devi **installare 🤗 Transformers dalla fonte** in un nuovo ambiente virtuale:
+
+```bash
+git clone https://github.com/huggingface/transformers
+cd transformers
+pip install .
+```
+Per le precedenti versioni degli script di esempio, clicca sul pulsante di seguito:
+
+<details>
+  <summary>Esempi per versioni precedenti di 🤗 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>
+
+Successivamente, cambia la tua attuale copia di 🤗 Transformers specificandone la versione, ad esempio v3.5.1:
+
+```bash
+git checkout tags/v3.5.1
+```
+
+ Dopo aver configurato correttamente la versione della libreria, naviga nella cartella degli esempi di tua scelta e installa i requisiti:
+
+```bash
+pip install -r requirements.txt
+```
+
+## Esegui uno script
+
+<frameworkcontent>
+<pt>
+
+Lo script di esempio scarica e pre-processa un dataset dalla libreria 🤗 [Datasets](https://huggingface.co/docs/datasets/). Successivamente, lo script esegue il fine-tuning su un dataset usando il [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) su un'architettura che supporta la summarization. Il seguente esempio mostra come eseguire il fine-tuning di [T5-small](https://huggingface.co/t5-small) sul dataset [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail). Il modello T5 richiede un parametro addizionale `source_prefix` a causa del modo in cui è stato addestrato. Questo prefisso permette a T5 di sapere che si tratta di un task di summarization.
+
+```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>
+Lo script di esempio scarica e pre-processa un dataset dalla libreria 🤗 [Datasets](https://huggingface.co/docs/datasets/). Successivamente, lo script esegue il fine-tuning su un dataset usando Keras su un'architettura che supporta la summarization. Il seguente esempio mostra come eseguire il fine-tuning di [T5-small](https://huggingface.co/t5-small) sul dataset [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail). Il modello T5 richiede un parametro addizionale `source_prefix` a causa del modo in cui è stato addestrato. Questo prefisso permette a T5 di sapere che si tratta di un task di summarization.
+
+```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>
+
+## Addestramento distribuito e precisione mista
+
+Il [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) supporta l'addestramento distribuito e la precisione mista, che significa che puoi anche usarla in uno script. Per abilitare entrambe le funzionalità:
+
+- Aggiunto l'argomento `fp16` per abilitare la precisione mista.
+- Imposta un numero di GPU da usare con l'argomento `nproc_per_node`.
+
+```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
+```
+
+Gli script TensorFlow utilizzano una [`MirroredStrategy`](https://www.tensorflow.org/guide/distributed_training#mirroredstrategy) per il training distribuito e non devi aggiungere alcun argomento addizionale allo script di training. Lo script TensorFlow userà multiple GPU in modo predefinito se quest'ultime sono disponibili:
+
+## Esegui uno script su TPU
+
+<frameworkcontent>
+<pt>
+Le Tensor Processing Units (TPU) sono state progettate per migliorare le prestazioni. PyTorch supporta le TPU con il compilatore per deep learning [XLA](https://www.tensorflow.org/xla) (guarda [questo link](https://github.com/pytorch/xla/blob/master/README.md) per maggiori dettagli). Per usare una TPU, avvia lo script `xla_spawn.py` e usa l'argomento `num_cores` per impostare il numero di core TPU che intendi usare.
+
+```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>
+Le Tensor Processing Units (TPU) sono state progettate per migliorare le prestazioni. Gli script TensorFlow utilizzano una [`TPUStrategy`](https://www.tensorflow.org/guide/distributed_training#tpustrategy) per eseguire l'addestramento su TPU. Per usare una TPU, passa il nome della risorsa TPU all'argomento `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>
+
+## Esegui uno script con 🤗 Accelerate
+
+🤗 [Accelerate](https://huggingface.co/docs/accelerate/index.html) è una libreria compatibile solo con PyTorch che offre un metodo unificato per addestrare modelli su diverse tipologie di configurazioni (CPU, multiple GPU, TPU) mantenendo una completa visibilità rispetto al ciclo di training di PyTorch. Assicurati di aver effettuato l'installazione di 🤗 Accelerate, nel caso non lo avessi fatto:
+
+> Nota: dato che Accelerate è in rapido sviluppo, è necessario installare la versione proveniente da git per eseguire gli script:
+```bash
+pip install git+https://github.com/huggingface/accelerate
+```
+
+Invece che usare lo script `run_summarization.py`, devi usare lo script `run_summarization_no_trainer.py`. Gli script supportati in 🤗 Accelerate avranno un file chiamato `task_no_trainer.py` nella rispettiva cartella. Per iniziare, esegui il seguente comando per creare e salvare un file di configurazione: 
+
+```bash
+accelerate config
+```
+
+Testa la tua configurazione per assicurarti della sua correttezza:
+
+```bash
+accelerate test
+```
+
+Ora sei pronto per avviare l'addestramento:
+
+```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
+```
+
+## Uso di un dataset personalizzato
+
+Lo script di summarization supporta dataset personalizzati purché siano file CSV o JSON Line. Quando usi il tuo dataset, devi specificare diversi argomenti aggiuntivi:
+
+- `train_file` e `validation_file` specificano dove si trovano i file di addestramento e validazione.
+- `text_column` è il file di input da riassumere.
+- `summary_column` è il file di destinazione per l'output.
+
+Uno script di summarization usando un dataset personalizzato sarebbe simile a questo:
+
+```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
+```
+
+## Testare uno script
+
+È spesso una buona idea avviare il tuo script su un numero inferiore di esempi tratti dal dataset, per assicurarti che tutto funzioni come previsto prima di eseguire lo script sull'intero dataset, che potrebbe necessitare di ore. Usa i seguenti argomenti per limitare il dataset ad un massimo numero di esempi:
+
+- `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
+```
+
+Non tutti gli esempi di script supportano l'argomento `max_predict_samples`. Se non sei sicuro circa il supporto di questo argomento da parte del tuo script, aggiungi l'argomento `-h` per controllare:
+
+```bash
+examples/pytorch/summarization/run_summarization.py -h
+```
+
+## Riavviare addestramento da un checkpoint
+
+Un'altra utile opzione è riavviare un addestramento da un checkpoint precedente. Questo garantirà che tu possa riprendere da dove hai interrotto senza ricominciare se l'addestramento viene interrotto. Ci sono due metodi per riavviare l'addestramento da un checkpoint: 
+
+Il primo metodo usa l'argomento `output_dir previous_output_dir` per riavviare l'addestramento dall'ultima versione del checkpoint contenuto in `output_dir`. In questo caso, dovresti rimuovere `overwrite_output_dir`:
+
+```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
+```
+
+Il secondo metodo usa l'argomento `resume_from_checkpoint path_to_specific_checkpoint` per riavviare un addestramento da una specifica cartella di checkpoint.
+
+```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
+```
+
+## Condividi il tuo modello
+
+Tutti gli script possono caricare il tuo modello finale al [Model Hub](https://huggingface.co/models). Prima di iniziare, assicurati di aver effettuato l'accesso su Hugging Face:
+
+```bash
+huggingface-cli login
+```
+
+Poi, aggiungi l'argomento `push_to_hub` allo script. Questo argomento consentirà di creare un repository con il tuo username Hugging Face e la cartella specificata in `output_dir`.
+
+Per dare uno specifico nome al repository, usa l'argomento `push_to_hub_model_id`. Il repository verrà automaticamente elencata sotto al tuo namespace.
+
+Il seguente esempio mostra come caricare un modello specificando il nome del repository:
+
+```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
+```
--- a/docs/source/it/serialization.mdx
+++ b/docs/source/it/serialization.mdx
@@ -0,0 +1,673 @@
+<!--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.
+-->
+
+# Esporta modelli 🤗 Transformers 
+
+Se devi implementare 🤗 modelli Transformers in ambienti di produzione, noi
+consigliamo di esportarli in un formato serializzato che può essere caricato ed eseguito
+su runtime e hardware specializzati. In questa guida ti mostreremo come farlo
+esporta 🤗 Modelli Transformers in due formati ampiamente utilizzati: ONNX e TorchScript.
+
+Una volta esportato, un modello può essere ottimizato per l'inferenza tramite tecniche come 
+la quantizzazione e soppressione. Se sei interessato a ottimizzare i tuoi modelli per l'esecuzione
+con la massima efficienza, dai un'occhiata a [🤗 Optimum
+library](https://github.com/huggingface/optimum).
+
+## ONNX
+
+Il progetto [ONNX (Open Neural Network eXchange)](http://onnx.ai) Il progetto onnx è un open
+standard che definisce un insieme comune di operatori e un formato di file comune a
+rappresentano modelli di deep learning in un'ampia varietà di framework, tra cui
+PyTorch e TensorFlow. Quando un modello viene esportato nel formato ONNX, questi
+operatori sono usati per costruire un grafico computazionale (often called an
+_intermediate representation_) che rappresenta il flusso di dati attraverso la
+rete neurale.
+
+Esponendo un grafico con operatori e tipi di dati standardizzati, ONNX rende
+più facile passare da un framework all'altro. Ad esempio, un modello allenato in PyTorch può
+essere esportato in formato ONNX e quindi importato in TensorFlow (e viceversa).
+
+🤗 Transformers fornisce un pacchetto `transformers.onnx` che ti consente di
+convertire i checkpoint del modello in un grafico ONNX sfruttando gli oggetti di configurazione.
+Questi oggetti di configurazione sono già pronti per una serie di architetture di modelli,
+e sono progettati per essere facilmente estensibili ad altre architetture.
+
+Le configurazioni pronte includono le seguenti architetture:
+
+<!--This table is automatically generated by `make fix-copies`, do not fill manually!-->
+
+- ALBERT
+- BART
+- BEiT
+- BERT
+- BigBird
+- BigBird-Pegasus
+- Blenderbot
+- BlenderbotSmall
+- CamemBERT
+- ConvBERT
+- Data2VecText
+- Data2VecVision
+- DeiT
+- DistilBERT
+- ELECTRA
+- FlauBERT
+- GPT Neo
+- GPT-J
+- I-BERT
+- LayoutLM
+- M2M100
+- Marian
+- mBART
+- MobileBERT
+- OpenAI GPT-2
+- Perceiver
+- PLBart
+- RoBERTa
+- RoFormer
+- SqueezeBERT
+- T5
+- ViT
+- XLM
+- XLM-RoBERTa
+- XLM-RoBERTa-XL
+
+Nelle prossime due sezioni, ti mostreremo come:
+
+* Esporta un modello supportato usando il pacchetto `transformers.onnx`.
+* Esporta un modello personalizzato per un'architettura non supportata.
+
+### Esportazione di un modello in ONNX
+
+Per esportare un modello 🤗 Transformers in ONNX, dovrai prima installarne alcune
+dipendenze extra:
+
+```bash
+pip install transformers[onnx]
+```
+
+Il pacchetto `transformers.onnx` può essere usato come modulo Python:
+
+```bash
+python -m transformers.onnx --help
+
+usage: Hugging Face Transformers ONNX exporter [-h] -m MODEL [--feature {causal-lm, ...}] [--opset OPSET] [--atol ATOL] output
+
+positional arguments:
+  output                Path indicating where to store generated ONNX model.
+
+optional arguments:
+  -h, --help            show this help message and exit
+  -m MODEL, --model MODEL
+                        Model ID on huggingface.co or path on disk to load model from.
+  --feature {causal-lm, ...}
+                        The type of features to export the model with.
+  --opset OPSET         ONNX opset version to export the model with.
+  --atol ATOL           Absolute difference tolerence when validating the model.
+```
+
+L'esportazione di un checkpoint utilizzando una configurazione già pronta può essere eseguita come segue:
+
+```bash
+python -m transformers.onnx --model=distilbert-base-uncased onnx/
+```
+
+che dovrebbe mostrare i seguenti log:
+
+```bash
+Validating ONNX model...
+        -[✓] ONNX model output names match reference model ({'last_hidden_state'})
+        - Validating ONNX Model output "last_hidden_state":
+                -[✓] (2, 8, 768) matches (2, 8, 768)
+                -[✓] all values close (atol: 1e-05)
+All good, model saved at: onnx/model.onnx
+```
+
+Questo esporta un grafico ONNX del checkpoint definito dall'argomento `--model`.
+In questo esempio è `distilbert-base-uncased`, ma può essere qualsiasi checkpoint
+Hugging Face Hub o uno memorizzato localmente.
+
+Il file risultante `model.onnx` può quindi essere eseguito su uno dei [tanti
+acceleratori](https://onnx.ai/supported-tools.html#deployModel) che supportano il
+lo standard ONNX. Ad esempio, possiamo caricare ed eseguire il modello con [ONNX
+Runtime](https://onnxruntime.ai/) come segue:
+
+```python
+>>> from transformers import AutoTokenizer
+>>> from onnxruntime import InferenceSession
+
+>>> tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
+>>> session = InferenceSession("onnx/model.onnx")
+>>> # ONNX Runtime expects NumPy arrays as input
+>>> inputs = tokenizer("Using DistilBERT with ONNX Runtime!", return_tensors="np")
+>>> outputs = session.run(output_names=["last_hidden_state"], input_feed=dict(inputs))
+```
+
+I nomi di output richiesti (cioè `["last_hidden_state"]`) possono essere ottenuti
+dando un'occhiata alla configurazione ONNX di ogni modello. Ad esempio, per
+DistilBERT abbiamo:
+
+```python
+>>> from transformers.models.distilbert import DistilBertConfig, DistilBertOnnxConfig
+
+>>> config = DistilBertConfig()
+>>> onnx_config = DistilBertOnnxConfig(config)
+>>> print(list(onnx_config.outputs.keys()))
+["last_hidden_state"]
+```
+
+Il processo è identico per i checkpoint TensorFlow sull'hub. Ad esempio, noi
+possiamo esportare un checkpoint TensorFlow puro da [Keras
+organizzazione](https://huggingface.co/keras-io) come segue:
+
+```bash
+python -m transformers.onnx --model=keras-io/transformers-qa onnx/
+```
+
+Per esportare un modello memorizzato localmente, devi disporre dei pesi del modello
+e file tokenizer memorizzati in una directory. Ad esempio, possiamo caricare e salvare un
+checkpoint come segue:
+
+<frameworkcontent>
+<pt>
+```python
+>>> from transformers import AutoTokenizer, AutoModelForSequenceClassification
+
+>>> # Load tokenizer and PyTorch weights form the Hub
+>>> tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
+>>> pt_model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased")
+>>> # Save to disk
+>>> tokenizer.save_pretrained("local-pt-checkpoint")
+>>> pt_model.save_pretrained("local-pt-checkpoint")
+```
+
+Una volta salvato il checkpoint, possiamo esportarlo su ONNX puntando l'argomento `--model`
+del pacchetto `transformers.onnx` nella directory desiderata:
+
+```bash
+python -m transformers.onnx --model=local-pt-checkpoint onnx/
+```
+</pt>
+<tf>
+```python
+>>> from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
+
+>>> # Load tokenizer and TensorFlow weights from the Hub
+>>> tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
+>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased")
+>>> # Save to disk
+>>> tokenizer.save_pretrained("local-tf-checkpoint")
+>>> tf_model.save_pretrained("local-tf-checkpoint")
+```
+
+Once the checkpoint is saved, we can export it to ONNX by pointing the `--model`
+argument of the `transformers.onnx` package to the desired directory:
+
+```bash
+python -m transformers.onnx --model=local-tf-checkpoint onnx/
+```
+</tf>
+</frameworkcontent>
+
+### Selezione delle caratteristiche per diverse topologie di modello
+
+Ogni configurazione già pronta viene fornita con una serie di _caratteristiche_ che ti consentono di
+esportare modelli per diversi tipi di topologie o attività. Come mostrato nella tabella
+di seguito, ogni caratteristica è associata a una diversa Auto Class:
+
+| Caratteristica                              | Auto Class                           |
+| ------------------------------------ | ------------------------------------ |
+| `causal-lm`, `causal-lm-with-past`   | `AutoModelForCausalLM`               |
+| `default`, `default-with-past`       | `AutoModel`                          |
+| `masked-lm`                          | `AutoModelForMaskedLM`               |
+| `question-answering`                 | `AutoModelForQuestionAnswering`      |
+| `seq2seq-lm`, `seq2seq-lm-with-past` | `AutoModelForSeq2SeqLM`              |
+| `sequence-classification`            | `AutoModelForSequenceClassification` |
+| `token-classification`               | `AutoModelForTokenClassification`    |
+
+Per ciascuna configurazione, puoi trovare l'elenco delle funzionalità supportate tramite il
+`FeaturesManager`. Ad esempio, per DistilBERT abbiamo:
+
+```python
+>>> from transformers.onnx.features import FeaturesManager
+
+>>> distilbert_features = list(FeaturesManager.get_supported_features_for_model_type("distilbert").keys())
+>>> print(distilbert_features)
+["default", "masked-lm", "causal-lm", "sequence-classification", "token-classification", "question-answering"]
+```
+
+Puoi quindi passare una di queste funzionalità all'argomento `--feature` nel
+pacchetto `transformers.onnx`. Ad esempio, per esportare un modello di classificazione del testo
+possiamo scegliere un modello ottimizzato dall'Hub ed eseguire:
+
+```bash
+python -m transformers.onnx --model=distilbert-base-uncased-finetuned-sst-2-english \
+                            --feature=sequence-classification onnx/
+```
+
+che visualizzerà i seguenti registri:
+
+```bash
+Validating ONNX model...
+        -[✓] ONNX model output names match reference model ({'logits'})
+        - Validating ONNX Model output "logits":
+                -[✓] (2, 2) matches (2, 2)
+                -[✓] all values close (atol: 1e-05)
+All good, model saved at: onnx/model.onnx
+```
+
+Puoi notare che in questo caso, i nomi di output del modello ottimizzato sono
+`logits` invece di `last_hidden_state` che abbiamo visto con il
+checkpoint `distilbert-base-uncased` precedente. Questo è previsto dal
+modello ottimizato visto che ha una testa di e.
+
+<Tip>
+
+Le caratteristiche che hanno un suffisso `wtih-past` (ad es. `causal-lm-with-past`)
+corrispondono a topologie di modello con stati nascosti precalcolati (chiave e valori
+nei blocchi di attenzione) che possono essere utilizzati per la decodifica autoregressiva veloce.
+
+</Tip>
+
+
+### Esportazione di un modello per un'architettura non supportata
+
+Se desideri esportare un modello la cui architettura non è nativamente supportata dalla
+libreria, ci sono tre passaggi principali da seguire:
+
+1. Implementare una configurazione ONNX personalizzata.
+2. Esportare il modello in ONNX.
+3. Convalidare gli output di PyTorch e dei modelli esportati.
+
+In questa sezione, vedremo come DistilBERT è stato implementato per mostrare cosa è
+coinvolto in ogni passaggio.
+
+#### Implementazione di una configurazione ONNX personalizzata
+
+Iniziamo con l'oggetto di configurazione ONNX. Forniamo tre classi
+astratte da cui ereditare, a seconda del tipo di archittettura
+del modello che desideri esportare:
+
+* I modelli basati su encoder ereditano da [`~onnx.config.OnnxConfig`]
+* I modelli basati su decoder ereditano da [`~onnx.config.OnnxConfigWithPast`]
+* I modelli encoder-decoder ereditano da[`~onnx.config.OnnxSeq2SeqConfigWithPast`]
+
+<Tip>
+
+Un buon modo per implementare una configurazione ONNX personalizzata è guardare l'implementazione
+esistente nel file `configuration_<model_name>.py` di un'architettura simile.
+
+</Tip>
+
+Poiché DistilBERT è un modello basato su encoder, la sua configurazione eredita da
+`OnnxConfig`:
+
+```python
+>>> from typing import Mapping, OrderedDict
+>>> from transformers.onnx import OnnxConfig
+
+
+>>> class DistilBertOnnxConfig(OnnxConfig):
+...     @property
+...     def inputs(self) -> Mapping[str, Mapping[int, str]]:
+...         return OrderedDict(
+...             [
+...                 ("input_ids", {0: "batch", 1: "sequence"}),
+...                 ("attention_mask", {0: "batch", 1: "sequence"}),
+...             ]
+...         )
+```
+
+Ogni oggetto di configurazione deve implementare la proprietà `inputs` e restituire una
+mappatura, dove ogni chiave corrisponde a un input previsto e ogni valore
+indica l'asse di quell'input. Per DistilBERT, possiamo vedere che sono richiesti
+due input: `input_ids` e `attention_mask`. Questi inputs hanno la stessa forma di
+`(batch_size, sequence_length)` per questo motivo vediamo gli stessi assi usati nella
+configurazione.
+
+<Tip>
+
+Puoi notare che la proprietà `inputs` per `DistilBertOnnxConfig` restituisce un
+`OrdinatoDict`. Ciò garantisce che gli input corrispondano alla loro posizione
+relativa all'interno del metodo `PreTrainedModel.forward()` durante il tracciamento del grafico.
+Raccomandiamo di usare un `OrderedDict` per le proprietà `inputs` e `outputs`
+quando si implementano configurazioni ONNX personalizzate.
+
+</Tip>
+
+Dopo aver implementato una configurazione ONNX, è possibile istanziarla
+fornendo alla configurazione del modello base come segue:
+
+```python
+>>> from transformers import AutoConfig
+
+>>> config = AutoConfig.from_pretrained("distilbert-base-uncased")
+>>> onnx_config = DistilBertOnnxConfig(config)
+```
+
+L'oggetto risultante ha diverse proprietà utili. Ad esempio è possibile visualizzare il
+Set operatore ONNX che verrà utilizzato durante l'esportazione:
+
+```python
+>>> print(onnx_config.default_onnx_opset)
+11
+```
+
+È inoltre possibile visualizzare gli output associati al modello come segue:
+
+```python
+>>> print(onnx_config.outputs)
+OrderedDict([("last_hidden_state", {0: "batch", 1: "sequence"})])
+```
+
+Puoi notare che la proprietà degli output segue la stessa struttura degli input; esso
+restituisce un `OrderedDict` di output con nome e le loro forme. La struttura di output
+è legato alla scelta della funzione con cui viene inizializzata la configurazione.
+Per impostazione predefinita, la configurazione ONNX viene inizializzata con la funzione 'predefinita'
+che corrisponde all'esportazione di un modello caricato con la classe `AutoModel`. Se tu
+desideri esportare una topologia di modello diversa, è sufficiente fornire una funzionalità diversa a
+l'argomento `task` quando inizializzi la configurazione ONNX. Ad esempio, se
+volevamo esportare DistilBERT con una testa di classificazione per sequenze, potremmo
+usare:
+
+```python
+>>> from transformers import AutoConfig
+
+>>> config = AutoConfig.from_pretrained("distilbert-base-uncased")
+>>> onnx_config_for_seq_clf = DistilBertOnnxConfig(config, task="sequence-classification")
+>>> print(onnx_config_for_seq_clf.outputs)
+OrderedDict([('logits', {0: 'batch'})])
+```
+
+<Tip>
+
+Tutte le proprietà e i metodi di base associati a [`~onnx.config.OnnxConfig`] e le
+altre classi di configurazione possono essere sovrascritte se necessario. Guarda
+[`BartOnnxConfig`] per un esempio avanzato.
+
+</Tip>
+
+#### Esportazione del modello
+
+Una volta implementata la configurazione ONNX, il passaggio successivo consiste nell'esportare il
+modello. Qui possiamo usare la funzione `export()` fornita dal
+pacchetto `transformers.onnx`. Questa funzione prevede la configurazione ONNX, insieme
+con il modello base e il tokenizer e il percorso per salvare il file esportato:
+
+```python
+>>> from pathlib import Path
+>>> from transformers.onnx import export
+>>> from transformers import AutoTokenizer, AutoModel
+
+>>> onnx_path = Path("model.onnx")
+>>> model_ckpt = "distilbert-base-uncased"
+>>> base_model = AutoModel.from_pretrained(model_ckpt)
+>>> tokenizer = AutoTokenizer.from_pretrained(model_ckpt)
+
+>>> onnx_inputs, onnx_outputs = export(tokenizer, base_model, onnx_config, onnx_config.default_onnx_opset, onnx_path)
+```
+
+Gli `onnx_inputs` e `onnx_outputs` restituiti dalla funzione `export()` sono
+liste di chiavi definite nelle proprietà di `input` e `output` della
+configurazione. Una volta esportato il modello, puoi verificare che il modello sia ben
+formato come segue:
+
+```python
+>>> import onnx
+
+>>> onnx_model = onnx.load("model.onnx")
+>>> onnx.checker.check_model(onnx_model)
+```
+
+<Tip>
+
+Se il tuo modello è più largo di 2 GB, vedrai che molti file aggiuntivi sono
+creati durante l'esportazione. Questo è _previsto_ perché ONNX utilizza [Protocol
+Buffer](https://developers.google.com/protocol-buffers/) per memorizzare il modello e
+questi hanno un limite di dimensione 2 GB. Vedi la [Documentazione
+ONNX](https://github.com/onnx/onnx/blob/master/docs/ExternalData.md)
+per istruzioni su come caricare modelli con dati esterni.
+
+</Tip>
+
+#### Convalida degli output del modello
+
+Il passaggio finale consiste nel convalidare gli output dal modello di base e quello esportato
+corrispondere entro una soglia di tolleranza assoluta. Qui possiamo usare la
+Funzione `validate_model_outputs()` fornita dal pacchetto `transformers.onnx`
+come segue:
+
+```python
+>>> from transformers.onnx import validate_model_outputs
+
+>>> validate_model_outputs(
+...     onnx_config, tokenizer, base_model, onnx_path, onnx_outputs, onnx_config.atol_for_validation
+... )
+```
+
+Questa funzione usa il metodo `OnnxConfig.generate_dummy_inputs()` per generare
+input per il modello di base e quello esportato e la tolleranza assoluta può essere
+definita nella configurazione. Generalmente troviamo una corrispondenza numerica nell'intervallo da 1e-6
+a 1e-4, anche se è probabile che qualsiasi cosa inferiore a 1e-3 vada bene.
+
+### Contribuire con una nuova configurazione a 🤗 Transformers
+
+Stiamo cercando di espandere l'insieme di configurazioni già pronte e di accettare
+contributi della community! Se vuoi contribuire con la tua aggiunta
+nella libreria, dovrai:
+
+* Implementare la configurazione ONNX nella corrispondente `configuration file
+_<model_name>.py`
+* Includere l'architettura del modello e le funzioni corrispondenti in [`~onnx.features.FeatureManager`]
+* Aggiungere la tua architettura del modello ai test in `test_onnx_v2.py`
+
+Scopri come stato contribuito la configurazione per [IBERT]
+(https://github.com/huggingface/transformers/pull/14868/files) per
+avere un'idea di cosa è coinvolto.
+
+## TorchScript
+
+<Tip>
+
+Questo è l'inizio dei nostri esperimenti con TorchScript e stiamo ancora esplorando le sue capacità con
+modelli con variable-input-size. È una nostra priorità e approfondiremo le nostre analisi nelle prossime versioni,
+con più esempi di codici, un'implementazione più flessibile e benchmark che confrontano i codici basati su Python con quelli compilati con
+TorchScript.
+
+</Tip>
+
+Secondo la documentazione di Pytorch: "TorchScript è un modo per creare modelli serializzabili e ottimizzabili da codice
+Pytorch". I due moduli di Pytorch [JIT e TRACE](https://pytorch.org/docs/stable/jit.html) consentono allo sviluppatore di esportare
+il loro modello da riutilizzare in altri programmi, come i programmi C++ orientati all'efficienza.
+
+Abbiamo fornito un'interfaccia che consente l'esportazione di modelli 🤗 Transformers in TorchScript in modo che possano essere riutilizzati
+in un ambiente diverso rispetto a un programma Python basato su Pytorch. Qui spieghiamo come esportare e utilizzare i nostri modelli utilizzando
+TorchScript.
+
+Esportare un modello richiede due cose:
+
+- Un passaggio in avanti con input fittizzi.
+- Istanziazione del modello con flag `torchscript`.
+
+Queste necessità implicano diverse cose a cui gli sviluppatori dovrebbero prestare attenzione. Questi dettagli mostrati sotto.
+
+### Flag TorchScript e pesi legati
+
+Questo flag è necessario perché la maggior parte dei modelli linguistici in questo repository hanno pesi legati tra il loro
+strato "Embedding" e lo strato "Decoding". TorchScript non consente l'esportazione di modelli che hanno pesi
+legati, quindi è necessario prima slegare e clonare i pesi.
+
+Ciò implica che i modelli istanziati con il flag `torchscript` hanno il loro strato `Embedding` e strato `Decoding`
+separato, il che significa che non dovrebbero essere addestrati in futuro. L'allenamento de-sincronizza i due
+strati, portando a risultati inaspettati.
+
+Questo non è il caso per i modelli che non hanno una testa del modello linguistico, poiché quelli non hanno pesi legati. Questi modelli
+può essere esportato in sicurezza senza il flag `torchscript`.
+
+### Input fittizi e standard lengths
+
+Gli input fittizzi sono usati per fare un modello passaggio in avanti . Mentre i valori degli input si propagano attraverso i strati,
+Pytorch tiene traccia delle diverse operazioni eseguite su ciascun tensore. Queste operazioni registrate vengono quindi utilizzate per
+creare la "traccia" del modello.
+
+La traccia viene creata relativamente alle dimensioni degli input. È quindi vincolato dalle dimensioni dell'input
+fittizio e non funzionerà per altre lunghezze di sequenza o dimensioni batch. Quando si proverà con una dimensione diversa, ci sarà errore
+come:
+
+`La dimensione espansa del tensore (3) deve corrispondere alla dimensione esistente (7) nella dimensione non singleton 2`
+
+will be raised. Si consiglia pertanto di tracciare il modello con una dimensione di input fittizia grande almeno quanto il più grande
+input che verrà fornito al modello durante l'inferenza. È possibile eseguire il padding per riempire i valori mancanti. Il modello
+sarà tracciato con una grande dimensione di input, tuttavia, anche le dimensioni della diverse matrici saranno grandi,
+risultando in più calcoli.
+
+Si raccomanda di prestare attenzione al numero totale di operazioni eseguite su ciascun input e di seguire da vicino le prestazioni
+durante l'esportazione di modelli di sequenza-lunghezza variabili.
+
+### Usare TorchSscript in Python
+
+Di seguito è riportato un esempio, che mostra come salvare, caricare modelli e come utilizzare la traccia per l'inferenza.
+
+#### Salvare un modello
+
+Questo frammento di codice mostra come usare TorchScript per esportare un `BertModel`. Qui il `BertModel` è istanziato secondo
+una classe `BertConfig` e quindi salvato su disco con il nome del file `traced_bert.pt`
+
+```python
+from transformers import BertModel, BertTokenizer, BertConfig
+import torch
+
+enc = BertTokenizer.from_pretrained("bert-base-uncased")
+
+# Tokenizing input text
+text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
+tokenized_text = enc.tokenize(text)
+
+# Masking one of the input tokens
+masked_index = 8
+tokenized_text[masked_index] = "[MASK]"
+indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
+segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
+
+# Creating a dummy input
+tokens_tensor = torch.tensor([indexed_tokens])
+segments_tensors = torch.tensor([segments_ids])
+dummy_input = [tokens_tensor, segments_tensors]
+
+# Initializing the model with the torchscript flag
+# Flag set to True even though it is not necessary as this model does not have an LM Head.
+config = BertConfig(
+    vocab_size_or_config_json_file=32000,
+    hidden_size=768,
+    num_hidden_layers=12,
+    num_attention_heads=12,
+    intermediate_size=3072,
+    torchscript=True,
+)
+
+# Instantiating the model
+model = BertModel(config)
+
+# The model needs to be in evaluation mode
+model.eval()
+
+# If you are instantiating the model with *from_pretrained* you can also easily set the TorchScript flag
+model = BertModel.from_pretrained("bert-base-uncased", torchscript=True)
+
+# Creating the trace
+traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
+torch.jit.save(traced_model, "traced_bert.pt")
+```
+
+#### Caricare un modello
+
+Questo frammento di codice mostra come caricare il `BertModel` che era stato precedentemente salvato su disco con il nome `traced_bert.pt`.
+Stiamo riutilizzando il `dummy_input` precedentemente inizializzato.
+
+```python
+loaded_model = torch.jit.load("traced_bert.pt")
+loaded_model.eval()
+
+all_encoder_layers, pooled_output = loaded_model(*dummy_input)
+```
+
+#### Utilizzare un modello tracciato per l'inferenza
+
+Usare il modello tracciato per l'inferenza è semplice come usare il suo metodo dunder `__call__`:
+
+```python
+traced_model(tokens_tensor, segments_tensors)
+```
+
+###Implementare modelli HuggingFace TorchScript su AWS utilizzando Neuron SDK
+
+AWS ha introdotto [Amazon EC2 Inf1](https://aws.amazon.com/ec2/instance-types/inf1/)
+famiglia di istanze per l'inferenza di machine learning a basso costo e ad alte prestazioni nel cloud.
+Le istanze Inf1 sono alimentate dal chip AWS Inferentia, un acceleratore hardware personalizzato,
+specializzato in carichi di lavoro di inferenza di deep learning.
+[AWS Neuron](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/#)
+è l'SDK per Inferentia che supporta il tracciamento e l'ottimizzazione dei modelli transformers per
+distribuzione su Inf1. L'SDK Neuron fornisce:
+
+
+1. API di facile utilizzo con una riga di modifica del codice per tracciare e ottimizzare un modello TorchScript per l'inferenza nel cloud.
+2. Ottimizzazioni delle prestazioni pronte all'uso per [miglioramento dei costi-prestazioni](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/benchmark/>)
+3. Supporto per i modelli di trasformatori HuggingFace costruiti con [PyTorch](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/bert_tutorial/tutorial_pretrained_bert.html)
+    o [TensorFlow](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/tensorflow/huggingface_bert/huggingface_bert.html).
+
+#### Implicazioni
+
+Modelli Transformers basati su architettura [BERT (Bidirectional Encoder Representations from Transformers)](https://huggingface.co/docs/transformers/main/model_doc/bert),
+o sue varianti come [distilBERT](https://huggingface.co/docs/transformers/main/model_doc/distilbert)
+e [roBERTa](https://huggingface.co/docs/transformers/main/model_doc/roberta)
+funzioneranno meglio su Inf1 per attività non generative come la question answering estrattive,
+Classificazione della sequenza, Classificazione dei token. In alternativa, generazione di testo
+le attività possono essere adattate per essere eseguite su Inf1, secondo questo [tutorial AWS Neuron MarianMT](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/transformers-marianmt.html).
+Ulteriori informazioni sui modelli che possono essere convertiti fuori dagli schemi su Inferentia possono essere
+trovati nella [sezione Model Architecture Fit della documentazione Neuron](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/models/models-inferentia.html#models-inferentia).
+
+#### Dipendenze
+
+L'utilizzo di AWS Neuron per convertire i modelli richiede le seguenti dipendenze e l'ambiente:
+
+* A [Neuron SDK environment](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/neuron-frameworks/pytorch-neuron/index.html#installation-guide),
+  which comes pre-configured on [AWS Deep Learning AMI](https://docs.aws.amazon.com/dlami/latest/devguide/tutorial-inferentia-launching.html).
+
+#### Convertire un modello per AWS Neuron
+
+Usando lo stesso script come in [Usando TorchScipt in Python](https://huggingface.co/docs/transformers/main/en/serialization#using-torchscript-in-python)
+per tracciare un "BertModel", importi l'estensione del framework `torch.neuron` per accedere
+i componenti di Neuron SDK tramite un'API Python.
+
+```python
+from transformers import BertModel, BertTokenizer, BertConfig
+import torch
+import torch.neuron
+```
+E modificare solo la riga di codice di traccia
+
+Da:
+
+```python
+torch.jit.trace(model, [tokens_tensor, segments_tensors])
+```
+
+A:
+
+```python
+torch.neuron.trace(model, [token_tensor, segments_tensors])
+```
+
+Questa modifica consente a Neuron SDK di tracciare il modello e ottimizzarlo per l'esecuzione nelle istanze Inf1.
+
+Per ulteriori informazioni sulle funzionalità, gli strumenti, i tutorial di esempi e gli ultimi aggiornamenti di AWS Neuron SDK,
+consultare la [documentazione AWS NeuronSDK](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/index.html).
--- a/docs/source/it/training.mdx
+++ b/docs/source/it/training.mdx
@@ -0,0 +1,372 @@
+<!--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.
+-->
+
+# Mettere a punto un modello pre-addestrato
+
+[[open-in-colab]]
+
+Ci sono benefici significativi nell'usare un modello pre-allenato. Riduce i costi computazionali, la tua impronta di carbonio, e ti consente di usare lo stato dell'arte dei modelli senza doverli addestrare da zero. 🤗 Transformers consente l'accesso a migliaia di modelli pre-addestrati per un'ampia gamma di compiti. Quando usi un modello pre-addestrato, lo alleni su un dataset specifico per il tuo compito. Questo è conosciuto come messa a punto (in inglese *fine-tuning*),una tecnica di addestramento incredibilmente potente. In questa esercitazione, potrai mettere a punto un modello pre-addestrato con un framework di deep learning a tua scelta:
+
+* Messa a punto di un modello pre-addestrato con 🤗 Transformers [`Trainer`].
+* Messa a punto di un modello pre-addestrato in TensorFlow con Keras.
+* Messa a punto di un modello pre-addestrato con PyTorch.
+
+<a id='data-processing'></a>
+
+## Preparare un dataset
+
+<Youtube id="_BZearw7f0w"/>
+
+Prima di poter mettere a punto un modello pre-addestrato, scarica un dataset e preparalo per l'addestramento. La precedente esercitazione ti mostra come processare i dati per l'addestramento, e adesso hai l'opportunità di provare queste capacità sul test!
+
+Inizia caricando il dataset [Yelp Reviews](https://huggingface.co/datasets/yelp_review_full):
+
+```py
+>>> from datasets import load_dataset
+
+>>> dataset = load_dataset("yelp_review_full")
+>>> dataset["train"][100]
+{'label': 0,
+ 'text': 'My expectations for McDonalds are t rarely high. But for one to still fail so spectacularly...that takes something special!\\nThe cashier took my friends\'s order, then promptly ignored me. I had to force myself in front of a cashier who opened his register to wait on the person BEHIND me. I waited over five minutes for a gigantic order that included precisely one kid\'s meal. After watching two people who ordered after me be handed their food, I asked where mine was. The manager started yelling at the cashiers for \\"serving off their orders\\" when they didn\'t have their food. But neither cashier was anywhere near those controls, and the manager was the one serving food to customers and clearing the boards.\\nThe manager was rude when giving me my order. She didn\'t make sure that I had everything ON MY RECEIPT, and never even had the decency to apologize that I felt I was getting poor service.\\nI\'ve eaten at various McDonalds restaurants for over 30 years. I\'ve worked at more than one location. I expect bad days, bad moods, and the occasional mistake. But I have yet to have a decent experience at this store. It will remain a place I avoid unless someone in my party needs to avoid illness from low blood sugar. Perhaps I should go back to the racially biased service of Steak n Shake instead!'}
+```
+
+Come già sai, hai bisogno di un tokenizer per processare il testo e includere una strategia di padding e truncation per gestire sequenze di lunghezza variabile. Per processare il dataset in un unico passo, usa il metodo [`map`](https://huggingface.co/docs/datasets/process.html#map) di 🤗 Datasets che applica la funzione di preprocessamento all'intero dataset:
+
+```py
+>>> from transformers import AutoTokenizer
+
+>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
+
+
+>>> def tokenize_function(examples):
+...     return tokenizer(examples["text"], padding="max_length", truncation=True)
+
+
+>>> tokenized_datasets = dataset.map(tokenize_function, batched=True)
+```
+
+If you like, you can create a smaller subset of the full dataset to fine-tune on to reduce the time it takes:
+
+```py
+>>> small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000))
+>>> small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000))
+```
+
+<a id='trainer'></a>
+
+## Addestramento
+
+<frameworkcontent>
+<pt>
+<Youtube id="nvBXf7s7vTI"/>
+
+🤗 Transformers mette a disposizione la classe [`Trainer`] ottimizzata per addestrare modelli 🤗 Transformers, rendendo semplice iniziare l'addestramento senza scrivere manualmente il tuo ciclo di allenamento. L'API [`Trainer`] supporta un'ampia gamma di opzioni e funzionalità di addestramento come logging, gradient accumulation, e mixed precision.
+
+Inizia caricando il tuo modello e specificando il numero di etichette attese. Nel dataset Yelp Review [dataset card](https://huggingface.co/datasets/yelp_review_full#data-fields), sai che ci sono cinque etichette:
+
+```py
+>>> from transformers import AutoModelForSequenceClassification
+
+>>> model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)
+```
+
+<Tip>
+
+Potresti vedere un'allerta dato che alcuni dei pesi pre-addestrati non sono stati utilizzati e altri pesi sono stati inizializzati casualmente. Non preoccuparti, è completamente normale! La testa pre-allenata del modello BERT viene scartata, e rimpiazzata da una testa per la classificazione inizializzata casualmente. Tu metterai a punto la nuova testa del modello sulla tua sequenza per il compito di classificazione, trasferendogli la conoscenza del modello pre-allenato.
+
+</Tip>
+
+### Addestrare gli iperparametri
+
+Successivamente, crea una classe [`TrainingArguments`] contenente tutti gli iperparametri che si possono calibrare nonché le variabili per attivare le differenti opzioni di addestramento. Per questa esercitazione puoi iniziare con gli [iperparametri](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments) di allenamento predefiniti, ma sentiti libero di sperimentare per trovare la configurazione ottimale per te.
+
+Specifica dove salvare i checkpoints del tuo addestramento:
+
+```py
+>>> from transformers import TrainingArguments
+
+>>> training_args = TrainingArguments(output_dir="test_trainer")
+```
+
+### Metriche
+
+[`Trainer`] non valuta automaticamente le performance del modello durante l'addestramento. Dovrai passare a [`Trainer`] una funzione che calcola e restituisce le metriche. La libreria 🤗 Datasets mette a disposizione una semplice funzione [`accuracy`](https://huggingface.co/metrics/accuracy) che puoi caricare con la funzione `load_metric` (guarda questa [esercitazione](https://huggingface.co/docs/datasets/metrics.html) per maggiori informazioni):
+
+```py
+>>> import numpy as np
+>>> from datasets import load_metric
+
+>>> metric = load_metric("accuracy")
+```
+
+Richiama `compute` su `metric` per calcolare l'accuratezza delle tue previsioni. Prima di passare le tue previsioni a `compute`, hai bisogno di convertirle in logits (ricorda che tutti i modelli 🤗 Transformers restituiscono logits):
+
+```py
+>>> def compute_metrics(eval_pred):
+...     logits, labels = eval_pred
+...     predictions = np.argmax(logits, axis=-1)
+...     return metric.compute(predictions=predictions, references=labels)
+```
+
+Se preferisci monitorare le tue metriche di valutazione durante il fine-tuning, specifica il parametro `evaluation_strategy` nei tuoi training arguments per restituire le metriche di valutazione ad ogni epoca di addestramento:
+
+```py
+>>> from transformers import TrainingArguments, Trainer
+
+>>> training_args = TrainingArguments(output_dir="test_trainer", evaluation_strategy="epoch")
+```
+
+### Trainer
+
+Crea un oggetto [`Trainer`] col tuo modello, training arguments, dataset di training e test, e funzione di valutazione:
+
+```py
+>>> trainer = Trainer(
+...     model=model,
+...     args=training_args,
+...     train_dataset=small_train_dataset,
+...     eval_dataset=small_eval_dataset,
+...     compute_metrics=compute_metrics,
+... )
+```
+
+Poi metti a punto il modello richiamando [`~transformers.Trainer.train`]:
+
+```py
+>>> trainer.train()
+```
+</pt>
+<tf>
+<a id='keras'></a>
+
+<Youtube id="rnTGBy2ax1c"/>
+
+I modelli 🤗 Transformers supportano anche l'addestramento in TensorFlow usando l'API di Keras.
+
+### Convertire dataset nel formato per TensorFlow
+
+Il [`DefaultDataCollator`] assembla tensori in lotti su cui il modello si addestrerà. Assicurati di specificare di restituire tensori per TensorFlow in `return_tensors`:
+
+```py
+>>> from transformers import DefaultDataCollator
+
+>>> data_collator = DefaultDataCollator(return_tensors="tf")
+```
+
+<Tip>
+
+[`Trainer`] usa [`DataCollatorWithPadding`] in maniera predefinita in modo da non dover specificare esplicitamente un collettore di dati.
+
+</Tip>
+
+Successivamente, converti i datasets tokenizzati in TensorFlow datasets con il metodo [`to_tf_dataset`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.to_tf_dataset). Specifica il tuo input in `columns`, e le tue etichette in `label_cols`:
+
+```py
+>>> tf_train_dataset = small_train_dataset.to_tf_dataset(
+...     columns=["attention_mask", "input_ids", "token_type_ids"],
+...     label_cols=["labels"],
+...     shuffle=True,
+...     collate_fn=data_collator,
+...     batch_size=8,
+... )
+
+>>> tf_validation_dataset = small_eval_dataset.to_tf_dataset(
+...     columns=["attention_mask", "input_ids", "token_type_ids"],
+...     label_cols=["labels"],
+...     shuffle=False,
+...     collate_fn=data_collator,
+...     batch_size=8,
+... )
+```
+
+### Compilazione e addestramento
+
+Carica un modello TensorFlow col numero atteso di etichette:
+
+```py
+>>> import tensorflow as tf
+>>> from transformers import TFAutoModelForSequenceClassification
+
+>>> model = TFAutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)
+```
+
+Poi compila e metti a punto il tuo modello usando [`fit`](https://keras.io/api/models/model_training_apis/) come si farebbe con qualsiasi altro modello di Keras:
+
+```py
+>>> model.compile(
+...     optimizer=tf.keras.optimizers.Adam(learning_rate=5e-5),
+...     loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
+...     metrics=tf.metrics.SparseCategoricalAccuracy(),
+... )
+
+>>> model.fit(tf_train_dataset, validation_data=tf_validation_dataset, epochs=3)
+```
+</tf>
+</frameworkcontent>
+
+<a id='pytorch_native'></a>
+
+## Addestramento in PyTorch nativo
+
+<frameworkcontent>
+<pt>
+<Youtube id="Dh9CL8fyG80"/>
+
+[`Trainer`] si occupa del ciclo di addestramento e ti consente di mettere a punto un modello con una sola riga di codice. Per chi preferisse scrivere un proprio ciclo di addestramento personale, puoi anche mettere a punto un modello 🤗 Transformers in PyTorch nativo.
+
+A questo punto, potresti avere bisogno di riavviare il tuo notebook o eseguire il seguente codice per liberare un po' di memoria:
+
+```py
+del model
+del pytorch_model
+del trainer
+torch.cuda.empty_cache()
+```
+
+Successivamente, postprocessa manualmente il `tokenized_dataset` per prepararlo ad essere allenato.
+
+1. Rimuovi la colonna `text` perché il modello non accetta testo grezzo come input:
+
+    ```py
+    >>> tokenized_datasets = tokenized_datasets.remove_columns(["text"])
+    ```
+
+2. Rinomina la colonna `label` in `labels` perché il modello si aspetta che questo argomento si chiami `labels`:
+
+    ```py
+    >>> tokenized_datasets = tokenized_datasets.rename_column("label", "labels")
+    ```
+
+3. Imposta il formato del dataset per farti restituire tensori di PyTorch all'interno delle liste:
+
+    ```py
+    >>> tokenized_datasets.set_format("torch")
+    ```
+
+Poi crea un piccolo sottocampione del dataset come visto precedentemente per velocizzare il fine-tuning:
+
+```py
+>>> small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000))
+>>> small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000))
+```
+
+### DataLoader
+
+Crea un `DataLoader` per i tuoi datasets di train e test così puoi iterare sui lotti di dati:
+
+```py
+>>> from torch.utils.data import DataLoader
+
+>>> train_dataloader = DataLoader(small_train_dataset, shuffle=True, batch_size=8)
+>>> eval_dataloader = DataLoader(small_eval_dataset, batch_size=8)
+```
+
+Carica il tuo modello con il numero atteso di etichette:
+
+```py
+>>> from transformers import AutoModelForSequenceClassification
+
+>>> model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)
+```
+
+### Ottimizzatore e pianificatore del tasso di apprendimento
+
+Crea un ottimizzatore e pianificatore del tasso di apprendimento per mettere a punto il modello. Usa l'ottimizzatore [`AdamW`](https://pytorch.org/docs/stable/generated/torch.optim.AdamW.html) di PyTorch:
+
+```py
+>>> from torch.optim import AdamW
+
+>>> optimizer = AdamW(model.parameters(), lr=5e-5)
+```
+
+Crea il pianificatore del tasso di apprendimento predefinito da [`Trainer`]:
+
+```py
+>>> from transformers import get_scheduler
+
+>>> num_epochs = 3
+>>> num_training_steps = num_epochs * len(train_dataloader)
+>>> lr_scheduler = get_scheduler(
+...     name="linear", optimizer=optimizer, num_warmup_steps=0, num_training_steps=num_training_steps
+... )
+```
+
+Infine, specifica come `device` da usare una GPU se ne hai una. Altrimenti, l'addestramento su una CPU può richiedere diverse ore invece di un paio di minuti.
+
+```py
+>>> import torch
+
+>>> device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+>>> model.to(device)
+```
+
+<Tip>
+
+Ottieni l'accesso gratuito a una GPU sul cloud se non ne possiedi una usando un notebook sul web come [Colaboratory](https://colab.research.google.com/) o [SageMaker StudioLab](https://studiolab.sagemaker.aws/).
+
+</Tip>
+
+Ottimo, adesso possiamo addestrare! 🥳 
+
+### Training loop
+
+Per tenere traccia dei tuoi progressi durante l'addestramento, usa la libreria [tqdm](https://tqdm.github.io/) per aggiungere una progress bar sopra il numero dei passi di addestramento:
+
+```py
+>>> from tqdm.auto import tqdm
+
+>>> progress_bar = tqdm(range(num_training_steps))
+
+>>> model.train()
+>>> for epoch in range(num_epochs):
+...     for batch in train_dataloader:
+...         batch = {k: v.to(device) for k, v in batch.items()}
+...         outputs = model(**batch)
+...         loss = outputs.loss
+...         loss.backward()
+
+...         optimizer.step()
+...         lr_scheduler.step()
+...         optimizer.zero_grad()
+...         progress_bar.update(1)
+```
+
+### Metriche
+
+Proprio come è necessario aggiungere una funzione di valutazione del [`Trainer`], è necessario fare lo stesso quando si scrive il proprio ciclo di allenamento. Ma invece di calcolare e riportare la metrica alla fine di ogni epoca, questa volta accumulerai tutti i batch con [`add_batch`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=add_batch#datasets.Metric.add_batch) e calcolerai la metrica alla fine.
+
+```py
+>>> metric = load_metric("accuracy")
+>>> model.eval()
+>>> for batch in eval_dataloader:
+...     batch = {k: v.to(device) for k, v in batch.items()}
+...     with torch.no_grad():
+...         outputs = model(**batch)
+
+...     logits = outputs.logits
+...     predictions = torch.argmax(logits, dim=-1)
+...     metric.add_batch(predictions=predictions, references=batch["labels"])
+
+>>> metric.compute()
+```
+</pt>
+</frameworkcontent>
+
+<a id='additional-resources'></a>
+
+## Altre risorse
+
+Per altri esempi sul fine-tuning, fai riferimento a:
+
+- [🤗 Transformers Examples](https://github.com/huggingface/transformers/tree/main/examples) include scripts per addestrare compiti comuni di NLP in PyTorch e TensorFlow.
+
+- [🤗 Transformers Notebooks](notebooks) contiene diversi notebooks su come mettere a punto un modello per compiti specifici in PyTorch e TensorFlow.
--- a/docs/source/pt/_toctree.yml
+++ b/docs/source/pt/_toctree.yml
@@ -1,9 +1,11 @@
 - sections:
+  - local: index
+    title: 🤗 Transformers
  - local: quicktour
    title: Tour rápido
  - local: installation
    title: Instalação
-  title: Iniciar
+  title: Início
 - sections:
  - local: pipeline_tutorial
    title: Pipelines para inferência
@@ -15,6 +17,8 @@
 - sections:
  - local: fast_tokenizers
    title: Usando os Tokenizers do 🤗 Tokenizers
+  - local: create_a_model
+    title: Criando uma arquitetura customizada
  - sections:
    - local: tasks/sequence_classification
      title: Classificação de texto
--- a/docs/source/pt/create_a_model.mdx
+++ b/docs/source/pt/create_a_model.mdx
@@ -0,0 +1,355 @@
+<!--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.
+-->
+
+# Criar uma arquitetura customizada
+
+Uma [`AutoClass`](model_doc/auto) automaticamente infere a arquitetura do modelo e baixa configurações e pesos pré-treinados. Geralmente, nós recomendamos usar uma `AutoClass` para produzir um código independente de checkpoints. Mas usuários que querem mais contole sobre parâmetros específicos do modelo pode criar um modelo customizado 🤗 Transformers a partir de algumas classes bases. Isso pode ser particulamente útil para alguém que está interessado em estudar, treinar ou fazer experimentos com um modelo 🤗 Transformers. Nesse tutorial, será explicado como criar um modelo customizado sem uma `AutoClass`. Aprenda como:
+
+- Carregar e customizar a configuração de um modelo.
+- Criar a arquitetura de um modelo.
+- Criar um tokenizer rápido e devagar para textos.
+- Criar extrator de features para tarefas envolvendo audio e imagem.
+- Criar um processador para tarefas multimodais.
+
+## configuration
+
+A [configuration](main_classes/configuration) refere-se a atributos específicos de um modelo. Cada configuração de modelo tem atributos diferentes; por exemplo, todos modelo de PLN possuem os atributos `hidden_size`, `num_attention_heads`, `num_hidden_layers` e `vocab_size` em comum. Esse atributos especificam o numero de 'attention heads' ou 'hidden layers' para construir um modelo.
+
+Dê uma olhada a mais em [DistilBERT](model_doc/distilbert) acessando [`DistilBertConfig`] para observar esses atributos:
+
+```py
+>>> from transformers import DistilBertConfig
+
+>>> config = DistilBertConfig()
+>>> print(config)
+DistilBertConfig {
+  "activation": "gelu",
+  "attention_dropout": 0.1,
+  "dim": 768,
+  "dropout": 0.1,
+  "hidden_dim": 3072,
+  "initializer_range": 0.02,
+  "max_position_embeddings": 512,
+  "model_type": "distilbert",
+  "n_heads": 12,
+  "n_layers": 6,
+  "pad_token_id": 0,
+  "qa_dropout": 0.1,
+  "seq_classif_dropout": 0.2,
+  "sinusoidal_pos_embds": false,
+  "transformers_version": "4.16.2",
+  "vocab_size": 30522
+}
+```
+
+[`DistilBertConfig`] mostra todos os atributos padrões usados para construir um [`DistilBertModel`] base. Todos atributos são customizáveis, o que cria espaço para experimentos. Por exemplo, você pode customizar um modelo padrão para:
+
+- Tentar uma função de ativação diferente com o parâmetro `activation`.
+- Usar uma taxa de desistência maior para as probabilidades de 'attention' com o parâmetro `attention_dropout`.
+
+```py
+>>> my_config = DistilBertConfig(activation="relu", attention_dropout=0.4)
+>>> print(my_config)
+DistilBertConfig {
+  "activation": "relu",
+  "attention_dropout": 0.4,
+  "dim": 768,
+  "dropout": 0.1,
+  "hidden_dim": 3072,
+  "initializer_range": 0.02,
+  "max_position_embeddings": 512,
+  "model_type": "distilbert",
+  "n_heads": 12,
+  "n_layers": 6,
+  "pad_token_id": 0,
+  "qa_dropout": 0.1,
+  "seq_classif_dropout": 0.2,
+  "sinusoidal_pos_embds": false,
+  "transformers_version": "4.16.2",
+  "vocab_size": 30522
+}
+```
+
+Atributos de um modelo pré-treinado podem ser modificados na função [`~PretrainedConfig.from_pretrained`]:
+
+```py
+>>> my_config = DistilBertConfig.from_pretrained("distilbert-base-uncased", activation="relu", attention_dropout=0.4)
+```
+
+Uma vez que você está satisfeito com as configurações do seu modelo, você consegue salvar elas com [`~PretrainedConfig.save_pretrained`]. Seu arquivo de configurações está salvo como um arquivo JSON no diretório especificado:
+
+```py
+>>> my_config.save_pretrained(save_directory="./your_model_save_path")
+```
+
+Para reusar o arquivo de configurações, carregue com [`~PretrainedConfig.from_pretrained`]:
+
+```py
+>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/my_config.json")
+```
+
+<Tip>
+
+Você pode também salvar seu arquivo de configurações como um dicionário ou até mesmo com a diferença entre as seus atributos de configuração customizados e os atributos de configuração padrões! Olhe a documentação [configuration](main_classes/configuration) para mais detalhes.
+
+</Tip>
+
+## Modelo
+
+O próximo passo é criar um [model](main_classes/models). O modelo - também vagamente referido como arquitetura - define o que cada camada está fazendo e quais operações estão acontecendo. Atributos como `num_hidden_layers` das configurações são utilizados para definir a arquitetura. Todo modelo compartilha a classe base [`PreTrainedModel`] e alguns métodos em comum como redimensionar o tamanho dos embeddings de entrada e podar as 'self-attention heads'. Além disso, todos os modelos também são subclasses de [`torch.nn.Module`](https://pytorch.org/docs/stable/generated/torch.nn.Module.html), [`tf.keras.Model`](https://www.tensorflow.org/api_docs/python/tf/keras/Model) ou [`flax.linen.Module`](https://flax.readthedocs.io/en/latest/flax.linen.html#module). Isso significa que os modelos são compatíveis com cada respectivo uso de framework.
+
+<frameworkcontent>
+<pt>
+Carregar seus atributos de configuração customizados em um modelo:
+
+```py
+>>> from transformers import DistilBertModel
+
+>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/my_config.json")
+>>> model = DistilBertModel(my_config)
+```
+
+Isso cria um modelo com valores aleatórios ao invés de pré-treinar os pesos. Você não irá conseguir usar usar esse modelo para nada útil ainda, até você treinar ele. Treino é um processo caro e demorado. Geralmente é melhor utilizar um modelo pré-treinado para obter melhores resultados mais rápido, enquanto usa apenas uma fração dos recursos necessários para treinar.
+
+Criar um modelo pré-treinado com [`~PreTrainedModel.from_pretrained`]:
+
+```py
+>>> model = DistilBertModel.from_pretrained("distilbert-base-uncased")
+```
+
+Quando você carregar os pesos pré-treinados, a configuração padrão do modelo é automaticamente carregada se o modelo é provido pelo 🤗 Transformers. No entanto, você ainda consegue mudar - alguns ou todos - os atributos padrões de configuração do modelo com os seus próprio atributos, se você preferir: 
+
+```py
+>>> model = DistilBertModel.from_pretrained("distilbert-base-uncased", config=my_config)
+```
+</pt>
+<tf>
+Carregar os seus próprios atributos padrões de contiguração no modelo:
+
+```py
+>>> from transformers import TFDistilBertModel
+
+>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/my_config.json")
+>>> tf_model = TFDistilBertModel(my_config)
+```
+
+Isso cria um modelo com valores aleatórios ao invés de pré-treinar os pesos. Você não irá conseguir usar usar esse modelo para nada útil ainda, até você treinar ele. Treino é um processo caro e demorado. Geralmente é melhor utilizar um modelo pré-treinado para obter melhores resultados mais rápido, enquanto usa apenas uma fração dos recursos necessários para treinar.
+
+Criar um modelo pré-treinado com [`~TFPreTrainedModel.from_pretrained`]:
+
+```py
+>>> tf_model = TFDistilBertModel.from_pretrained("distilbert-base-uncased")
+```
+
+Quando você carregar os pesos pré-treinados, a configuração padrão do modelo é automaticamente carregada se o modelo é provido pelo 🤗 Transformers. No entanto, você ainda consegue mudar - alguns ou todos - os atributos padrões de configuração do modelo com os seus próprio atributos, se você preferir: 
+
+```py
+>>> tf_model = TFDistilBertModel.from_pretrained("distilbert-base-uncased", config=my_config)
+```
+</tf>
+</frameworkcontent>
+
+### Heads do modelo
+
+Neste ponto, você tem um modelo básico do DistilBERT que gera os *estados ocultos*. Os estados ocultos são passados como entrada para a head do moelo para produzir a saída final. 🤗 Transformers fornece uma head de modelo diferente para cada tarefa desde que o modelo suporte essa tarefa (por exemplo, você não consegue utilizar o modelo DistilBERT para uma tarefa de 'sequence-to-sequence' como tradução).
+
+<frameworkcontent>
+<pt>
+Por exemplo, [`DistilBertForSequenceClassification`] é um modelo DistilBERT base com uma head de classificação de sequência. A head de calssificação de sequência é uma camada linear no topo das saídas agrupadas.
+
+```py
+>>> from transformers import DistilBertForSequenceClassification
+
+>>> model = DistilBertForSequenceClassification.from_pretrained("distilbert-base-uncased")
+```
+
+Reutilize facilmente esse ponto de parada para outra tarefe mudando para uma head de modelo diferente. Para uma tarefe de responder questões, você usaria a head do modelo [`DistilBertForQuestionAnswering`]. A head de responder questões é similar com a de classificação de sequências exceto o fato de que ela é uma camada no topo dos estados das saídas ocultas.
+
+```py
+>>> from transformers import DistilBertForQuestionAnswering
+
+>>> model = DistilBertForQuestionAnswering.from_pretrained("distilbert-base-uncased")
+```
+</pt>
+<tf>
+Por exemplo, [`TFDistilBertForSequenceClassification`] é um modelo DistilBERT base com uma head de classificação de sequência. A head de calssificação de sequência é uma camada linear no topo das saídas agrupadas.
+
+```py
+>>> from transformers import TFDistilBertForSequenceClassification
+
+>>> tf_model = TFDistilBertForSequenceClassification.from_pretrained("distilbert-base-uncased")
+```
+
+Reutilize facilmente esse ponto de parada para outra tarefe mudando para uma head de modelo diferente. Para uma tarefe de responder questões, você usaria a head do modelo [`TFDistilBertForQuestionAnswering`]. A head de responder questões é similar com a de classificação de sequências exceto o fato de que ela é uma camada no topo dos estados das saídas ocultas.
+
+```py
+>>> from transformers import TFDistilBertForQuestionAnswering
+
+>>> tf_model = TFDistilBertForQuestionAnswering.from_pretrained("distilbert-base-uncased")
+```
+</tf>
+</frameworkcontent>
+
+## Tokenizer
+
+A útlima classe base que você precisa antes de usar um modelo para dados textuais é a [tokenizer](main_classes/tokenizer) para converter textos originais para tensores. Existem dois tipos de tokenizers que você pode usar com 🤗 Transformers:
+
+- [`PreTrainedTokenizer`]: uma implementação em Python de um tokenizer.
+- [`PreTrainedTokenizerFast`]: um tokenizer da nossa biblioteca [🤗 Tokenizer](https://huggingface.co/docs/tokenizers/python/latest/) baseada em Rust. Esse tipo de tokenizer é significantemente mais rapido - especialmente durante tokenization de codificação - devido a implementação em Rust. O tokenizer rápido tambem oferece métodos adicionais como *offset mapping* que mapeia tokens para suar palavras ou caracteres originais.
+
+Os dois tokenizers suporta métodos comuns como os de codificar e decodificar, adicionar novos tokens, e gerenciar tokens especiais.
+
+<Tip warning={true}>
+
+Nem todo modelo suporta um 'fast tokenizer'. De uma olhada aqui [table](index#supported-frameworks) pra checar se um modelo suporta 'fast tokenizer'.
+
+</Tip>
+
+Se você treinou seu prórpio tokenizer, você pode criar um a partir do seu arquivo *vocabulary*:
+
+```py
+>>> from transformers import DistilBertTokenizer
+
+>>> my_tokenizer = DistilBertTokenizer(vocab_file="my_vocab_file.txt", do_lower_case=False, padding_side="left")
+```
+
+É importante lembrar que o vocabulário de um tokenizer customizado será diferente de um vocabulário gerado pelo tokenizer de um modelo pré treinado. Você precisa usar o vocabulário de um modelo pré treinado se você estiver usando um modelo pré treinado, caso contrário as entradas não farão sentido. Criando um tokenizer com um vocabulário de um modelo pré treinado com a classe [`DistilBertTokenizer`]:
+
+```py
+>>> from transformers import DistilBertTokenizer
+
+>>> slow_tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
+```
+
+Criando um 'fast tokenizer' com a classe [`DistilBertTokenizerFast`]:
+
+```py
+>>> from transformers import DistilBertTokenizerFast
+
+>>> fast_tokenizer = DistilBertTokenizerFast.from_pretrained("distilbert-base-uncased")
+```
+
+<Tip>
+
+Pos padrão, [`AutoTokenizer`] tentará carregar um 'fast tokenizer'. Você pode disabilitar esse comportamento colocando `use_fast=False` no `from_pretrained`.
+
+</Tip>
+
+## Extrator de features
+
+Um extrator de features processa entradas de imagem ou áudio. Ele herda da classe base [`~feature_extraction_utils.FeatureExtractionMixin`], e pode também herdar da classe [`ImageFeatureExtractionMixin`] para processamento de features de imagem ou da classe [`SequenceFeatureExtractor`] para processamento de entradas de áudio.
+
+Dependendo do que você está trabalhando em um audio ou uma tarefa de visão, crie um estrator de features associado com o modelo que você está usando. Por exemplo, crie um [`ViTFeatureExtractor`] padrão se você estiver usando [ViT](model_doc/vit) para classificação de imagens:
+
+```py
+>>> from transformers import ViTFeatureExtractor
+
+>>> vit_extractor = ViTFeatureExtractor()
+>>> print(vit_extractor)
+ViTFeatureExtractor {
+  "do_normalize": true,
+  "do_resize": true,
+  "feature_extractor_type": "ViTFeatureExtractor",
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "resample": 2,
+  "size": 224
+}
+```
+
+<Tip>
+
+Se você não estiver procurando por nenhuma customização, apenas use o método `from_pretrained` para carregar parâmetros do modelo de extrator de features padrão.
+
+</Tip>
+
+Modifique qualquer parâmetro dentre os [`ViTFeatureExtractor`] para criar seu extrator de features customizado.
+
+```py
+>>> from transformers import ViTFeatureExtractor
+
+>>> my_vit_extractor = ViTFeatureExtractor(resample="PIL.Image.BOX", do_normalize=False, image_mean=[0.3, 0.3, 0.3])
+>>> print(my_vit_extractor)
+ViTFeatureExtractor {
+  "do_normalize": false,
+  "do_resize": true,
+  "feature_extractor_type": "ViTFeatureExtractor",
+  "image_mean": [
+    0.3,
+    0.3,
+    0.3
+  ],
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "resample": "PIL.Image.BOX",
+  "size": 224
+}
+```
+
+Para entradas de áutio, você pode criar um [`Wav2Vec2FeatureExtractor`] e customizar os parâmetros de uma forma similar:
+
+```py
+>>> from transformers import Wav2Vec2FeatureExtractor
+
+>>> w2v2_extractor = Wav2Vec2FeatureExtractor()
+>>> print(w2v2_extractor)
+Wav2Vec2FeatureExtractor {
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "return_attention_mask": false,
+  "sampling_rate": 16000
+}
+```
+
+## Processor
+
+Para modelos que suportam tarefas multimodais, 🤗 Transformers oferece uma classe processadora que convenientemente cobre um extrator de features e tokenizer dentro de um único objeto. Por exemplo, vamos usar o [`Wav2Vec2Processor`] para uma tarefa de reconhecimento de fala automática (ASR). ASR transcreve áudio para texto, então você irá precisar de um extrator de um features e um tokenizer.
+
+Crie um extrator de features para lidar com as entradas de áudio.
+
+```py
+>>> from transformers import Wav2Vec2FeatureExtractor
+
+>>> feature_extractor = Wav2Vec2FeatureExtractor(padding_value=1.0, do_normalize=True)
+```
+
+Crie um tokenizer para lidar com a entrada de textos:
+
+```py
+>>> from transformers import Wav2Vec2CTCTokenizer
+
+>>> tokenizer = Wav2Vec2CTCTokenizer(vocab_file="my_vocab_file.txt")
+```
+
+Combine o extrator de features e o tokenizer no [`Wav2Vec2Processor`]:
+
+```py
+>>> from transformers import Wav2Vec2Processor
+
+>>> processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer)
+```
+
+Com duas classes básicas - configuração e modelo - e um preprocessamento de classe adicional (tokenizer, extrator de features, ou processador), você pode criar qualquer modelo que suportado por 🤗 Transformers. Qualquer uma dessas classes base são configuráveis, te permitindo usar os atributos específicos que você queira. Você pode facilmente preparar um modelo para treinamento ou modificar um modelo pré-treinado com poucas mudanças.
--- a/docs/source/pt/index.mdx
+++ b/docs/source/pt/index.mdx
@@ -0,0 +1,287 @@
+<!--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.
+-->
+
+# 🤗 Transformers
+
+
+Estado da Arte para Aprendizado de Máquina em PyTorch, TensorFlow e JAX.
+O 🤗 Transformers disponibiliza APIs para facilmente baixar e treinar modelos pré-treinados de última geração.
+O uso de modelos pré-treinados pode diminuir os seus custos de computação, a sua pegada de carbono, além de economizar o
+tempo necessário para se treinar um modelo do zero. Os modelos podem ser usados para diversas tarefas:
+
+* 📝 Textos: classificação, extração de informações, perguntas e respostas, resumir, traduzir e gerar textos em mais de 100 idiomas.
+* 🖼 Imagens: classificação, deteção de objetos, e segmentação.
+* 🗣 Audio: reconhecimento de fala e classificação de áudio.
+* 🐙 Multimodal: perguntas tabeladas e respsostas, reconhecimento ótico de charactéres, extração de informação de
+documentos escaneados, classificação de vídeo, perguntas e respostas visuais.
+
+Nossa biblioteca aceita integração contínua entre três das bibliotecas mais populares de aprendizado profundo:
+Our library supports seamless integration between three of the most popular deep learning libraries:
+[PyTorch](https://pytorch.org/), [TensorFlow](https://www.tensorflow.org/) e [JAX](https://jax.readthedocs.io/en/latest/).
+Treine seu modelo em três linhas de código em um framework, e carregue-o para execução em outro.
+
+Cada arquitetura 🤗 Transformers é definida em um módulo individual do Python, para que seja facilmente customizável para pesquisa e experimentos.
+
+## Se você estiver procurando suporte do time da Hugging Face, acesse
+
+<a target="_blank" href="https://huggingface.co/support">
+    <img alt="HuggingFace Expert Acceleration Program" src="https://huggingface.co/front/thumbnails/support.png" style="max-width: 600px; border: 1px solid #eee; border-radius: 4px; box-shadow: 0 1px 2px 0 rgba(0, 0, 0, 0.05);"></img>
+</a><br>
+
+## Conteúdo
+
+A documentação é dividida em cinco partes:
+    - **INÍCIO** contém um tour rápido de instalação e instruções para te dar um empurrão inicial com os 🤗 Transformers.
+    - **TUTORIAIS** são perfeitos para começar a aprender sobre a nossa biblioteca. Essa seção irá te ajudar a desenvolver
+    habilidades básicas necessárias para usar o 🤗 Transformers.
+    - **GUIAS PRÁTICOS** irão te mostrar como alcançar um certo objetivo, como o fine-tuning de um modelo pré-treinado
+    para modelamento de idioma, ou como criar um cabeçalho personalizado para um modelo.
+    - **GUIAS CONCEITUAIS** te darão mais discussões e explicações dos conceitos fundamentais e idéias por trás dos modelos,
+    tarefas e da filosofia de design por trás do 🤗 Transformers.
+    - **API** descreve o funcionamento de cada classe e função, agrupada em:
+
+    - **CLASSES PRINCIPAIS** para as classes que expõe as APIs importantes da biblioteca.
+    - **MODELOS** para as classes e funções relacionadas à cada modelo implementado na biblioteca.
+    - **AUXILIARES INTERNOS** para as classes e funções usadas internamente.
+
+Atualmente a biblioteca contém implementações do PyTorch, TensorFlow e JAX, pesos para modelos pré-treinados e scripts de uso e conversão de utilidades para os seguintes modelos:
+
+### Modelos atuais
+
+<!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->
+
+1. **[ALBERT](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. **[BART](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](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](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](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](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. **[BERTweet](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. **[BERT For Sequence Generation](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. **[BigBird-RoBERTa](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. **[BigBird-Pegasus](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. **[Blenderbot](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](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. **[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. **[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.
+1. **[ConvNeXT](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. **[CLIP](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. **[ConvBERT](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. **[CPM](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. **[CTRL](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. **[Data2Vec](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](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](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](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. **[DiT](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. **[DeiT](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. **[DETR](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](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. **[DistilBERT](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. **[DPR](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](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. **[EncoderDecoder](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. **[ELECTRA](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. **[FlauBERT](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. **[FNet](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. **[Funnel Transformer](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. **[GLPN](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](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-2](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](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 Neo](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. **[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. **[ImageGPT](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. **[LayoutLM](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](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. **[LayoutXLM](model_doc/layoutlmv2)** (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](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. **[Longformer](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. **[LUKE](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. **[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. **[LXMERT](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. **[M2M100](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](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. **[MaskFormer](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. **[MBart](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](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. **[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. **[MPNet](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. **[MT5](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. **[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. **[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. **[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. **[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. **[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.
+1. **[PoolFormer](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. **[ProphetNet](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. **[QDQBert](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. **[REALM](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](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. **[RemBERT](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. **[RegNet](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. **[ResNet](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](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. **[RoFormer](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. **[SegFormer](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. **[SEW](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](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. **[SpeechToTextTransformer](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](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](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](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. **[Swin Transformer](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. **[T5](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](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. **[TAPAS](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](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. **[Transformer-XL](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](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. **[UniSpeech](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](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. **[VAN](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. **[ViLT](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)](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. **[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. **[VisualBERT](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. **[WavLM](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. **[Wav2Vec2](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. **[Wav2Vec2Phoneme](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. **[XGLM](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](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](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](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](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. **[XLNet](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. **[XLSR-Wav2Vec2](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. **[XLS-R](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. **[YOSO](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.
+
+
+### Frameworks aceitos
+
+A tabela abaixo representa a lista de suporte na biblioteca para cada um dos seguintes modelos, caso tenham um tokenizer
+do Python (chamado de "slow"), ou um tokenizer construído em cima da biblioteca 🤗 Tokenizers (chamado de "fast"). Além
+disso, são diferenciados pelo suporte em diferentes frameworks: JAX (por meio do Flax); PyTorch; e/ou Tensorflow.
+
+<!--This table is updated automatically from the auto modules with _make fix-copies_. Do not update manually!-->
+
+|            Model            | Tokenizer slow | Tokenizer fast | PyTorch support | TensorFlow support | Flax Support |
+|:---------------------------:|:--------------:|:--------------:|:---------------:|:------------------:|:------------:|
+|           ALBERT            |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|            BART             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|            BEiT             |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
+|            BERT             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|       Bert Generation       |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           BigBird           |       ✅       |       ✅       |       ✅        |         ❌         |      ✅      |
+|       BigBirdPegasus        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|         Blenderbot          |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|       BlenderbotSmall       |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|          CamemBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|           Canine            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            CLIP             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|          ConvBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|          ConvNext           |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
+|            CTRL             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|        Data2VecAudio        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|        Data2VecText         |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|       Data2VecVision        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           DeBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|         DeBERTa-v2          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|    Decision Transformer     |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            DeiT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            DETR             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|         DistilBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|             DPR             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|             DPT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           ELECTRA           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|       Encoder decoder       |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
+| FairSeq Machine-Translation |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|          FlauBERT           |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|            FNet             |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|     Funnel Transformer      |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|            GLPN             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           GPT Neo           |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
+|            GPT-J            |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
+|           Hubert            |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
+|           I-BERT            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|          ImageGPT           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|          LayoutLM           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|         LayoutLMv2          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|             LED             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|         Longformer          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|            LUKE             |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           LXMERT            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|           M2M100            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           Marian            |       ✅       |       ❌       |       ✅        |         ✅         |      ✅      |
+|         MaskFormer          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            mBART            |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|        MegatronBert         |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|         MobileBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|            MPNet            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|             mT5             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|        Nystromformer        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|         OpenAI GPT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|        OpenAI GPT-2         |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|           Pegasus           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|          Perceiver          |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           PLBart            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|         PoolFormer          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|         ProphetNet          |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           QDQBert           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|             RAG             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|            Realm            |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|          Reformer           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|           RegNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           RemBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|           ResNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|          RetriBERT          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|           RoBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|          RoFormer           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|          SegFormer          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|             SEW             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            SEW-D            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|   Speech Encoder decoder    |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
+|         Speech2Text         |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|        Speech2Text2         |       ✅       |       ❌       |       ❌        |         ❌         |      ❌      |
+|          Splinter           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|         SqueezeBERT         |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|            Swin             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|             T5              |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|            TAPAS            |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|            TAPEX            |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|       Transformer-XL        |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|            TrOCR            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|          UniSpeech          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|        UniSpeechSat         |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|             VAN             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            ViLT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|   Vision Encoder decoder    |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
+|    VisionTextDualEncoder    |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
+|         VisualBert          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|             ViT             |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
+|           ViTMAE            |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
+|          Wav2Vec2           |       ✅       |       ❌       |       ✅        |         ✅         |      ✅      |
+|            WavLM            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            XGLM             |       ✅       |       ✅       |       ✅        |         ❌         |      ✅      |
+|             XLM             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|         XLM-RoBERTa         |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|       XLM-RoBERTa-XL        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|        XLMProphetNet        |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            XLNet            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|            YOSO             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+
+<!-- End table-->
--- a/docs/source/pt/training.mdx
+++ b/docs/source/pt/training.mdx
@@ -19,7 +19,7 @@ permite utilizar modelos de última geração sem ter que treinar um novo desde
 O 🤗 Transformers proporciona acesso a milhares de modelos pré-treinados numa ampla gama de tarefas.
 Quando utilizar um modelo pré-treinado, treine-o com um dataset específico para a sua tarefa.
 Isto é chamado de fine-tuning, uma técnica de treinamento incrivelmente poderosa. Neste tutorial faremos o fine-tuning
-a um modelo pré-treinado com um framework de Deep Learning de sua escolha:
+de um modelo pré-treinado com um framework de Deep Learning da sua escolha:

 * Fine-tuning de um modelo pré-treinado com o 🤗 Transformers [`Trainer`].
 * Fine-tuning de um modelo pré-treinado no TensorFlow com o Keras.
@@ -78,7 +78,7 @@ Se desejar, é possível criar um subconjunto menor do dataset completo para apl
 <Youtube id="nvBXf7s7vTI"/>

 O 🤗 Transformers proporciona uma classe [`Trainer`] otimizada para o treinamento de modelos de 🤗 Transformers,
-facilitando os primeiros passos do treinamento sem a necessidade de escrever manualmente seu próprio ciclo.
+facilitando os primeiros passos do treinamento sem a necessidade de escrever manualmente o seu próprio ciclo.
 A API do [`Trainer`] suporta um grande conjunto de opções de treinamento e funcionalidades, como o logging,
 o gradient accumulation e o mixed precision.

@@ -131,8 +131,8 @@ uma função de [`accuracy`](https://huggingface.co/metrics/accuracy) simples qu
 >>> metric = load_metric("accuracy")
 ```

-Defina a função `compute` dentro de `metric` para calcular a precisão de suas predições.
-Antes de passar suas predições ao `compute`, é necessário converter as predições à logits (lembre-se que
+Defina a função `compute` dentro de `metric` para calcular a precisão das suas predições.
+Antes de passar as suas predições ao `compute`, é necessário converter as predições à logits (lembre-se que
 todos os modelos de 🤗 Transformers retornam logits).

 ```py
@@ -142,8 +142,8 @@ todos os modelos de 🤗 Transformers retornam logits).
 ...     return metric.compute(predictions=predictions, references=labels)
 ```

-Se quiser controlar suas métricas de avaliação durante o fine-tuning, especifique o parâmetro `evaluation_strategy`
-em seus argumentos de treinamento para que o modelo leve em conta a métrica de avaliação ao final de cada época:
+Se quiser controlar as suas métricas de avaliação durante o fine-tuning, especifique o parâmetro `evaluation_strategy`
+nos seus argumentos de treinamento para que o modelo considere a métrica de avaliação ao final de cada época:

 ```py
 >>> from transformers import TrainingArguments
@@ -153,7 +153,7 @@ em seus argumentos de treinamento para que o modelo leve em conta a métrica de

 ### Trainer

-Crie um objeto [`Trainer`] com seu modelo, argumentos de treinamento, conjuntos de dados de treinamento e de teste, e sua função de avaliação:
+Crie um objeto [`Trainer`] com o seu modelo, argumentos de treinamento, conjuntos de dados de treinamento e de teste, e a sua função de avaliação:

 ```py
 >>> trainer = Trainer(
@@ -251,7 +251,7 @@ faria com qualquer outro modelo do Keras:
 <Youtube id="Dh9CL8fyG80"/>

 O [`Trainer`] se encarrega do ciclo de treinamento e permite aplicar o fine-tuning a um modelo em uma linha de código apenas.
-Para os usuários que preferirem escrever seu próprio ciclo de treinamento, também é possível aplicar o fine-tuning a um
+Para os usuários que preferirem escrever o seu próprio ciclo de treinamento, também é possível aplicar o fine-tuning a um
 modelo de 🤗 Transformers em PyTorch nativo.

 Neste momento, talvez ocorra a necessidade de reinicar seu notebook ou executar a seguinte linha de código para liberar
@@ -293,7 +293,7 @@ Em sequência, crie um subconjunto menor do dataset, como foi mostrado anteriorm

 ### DataLoader

-Crie um `DataLoader` para seus datasets de treinamento e de teste para poder iterar sobre batches de dados:
+Crie um `DataLoader` para os seus datasets de treinamento e de teste para poder iterar sobre batches de dados:

 ```py
 >>> from torch.utils.data import DataLoader
@@ -352,7 +352,6 @@ em uma CPU pode acabar levando várias horas em vez de minutos.
 </Tip>

 Perfeito, agora estamos prontos para começar o treinamento! 🥳
-Genial, ¡ahora estamos listos entrenar! 🥳

 ### Ciclo de treinamento

--- a/examples/flax/language-modeling/run_clm_flax.py
+++ b/examples/flax/language-modeling/run_clm_flax.py
@@ -501,7 +501,6 @@ def main():
            config,
            seed=training_args.seed,
            dtype=getattr(jnp, model_args.dtype),
-            use_auth_token=True if model_args.use_auth_token else None,
        )

    # Preprocessing the datasets.
--- a/examples/flax/language-modeling/run_mlm_flax.py
+++ b/examples/flax/language-modeling/run_mlm_flax.py
@@ -326,7 +326,7 @@ class FlaxDataCollatorForLanguageModeling:
        return inputs, labels


-def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
+def generate_batch_splits(samples_idx: np.ndarray, batch_size: int) -> np.ndarray:
    num_samples = len(samples_idx)
    samples_to_remove = num_samples % batch_size

@@ -632,7 +632,6 @@ def main():
            config,
            seed=training_args.seed,
            dtype=getattr(jnp, model_args.dtype),
-            use_auth_token=True if model_args.use_auth_token else None,
        )

    # Store some constant
@@ -755,7 +754,8 @@ def main():

        # Generate an epoch by shuffling sampling indices from the train dataset
        num_train_samples = len(tokenized_datasets["train"])
-        train_samples_idx = jax.random.permutation(input_rng, jnp.arange(num_train_samples))
+        # Avoid using jax.numpy here in case of TPU training
+        train_samples_idx = np.random.permutation(np.arange(num_train_samples))
        train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size)

        # Gather the indexes for creating the batch and do a training step
@@ -787,7 +787,8 @@ def main():
            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
                # ======================== Evaluating ==============================
                num_eval_samples = len(tokenized_datasets["validation"])
-                eval_samples_idx = jnp.arange(num_eval_samples)
+                # Avoid using jax.numpy here in case of TPU training
+                eval_samples_idx = np.arange(num_eval_samples)
                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)

                eval_metrics = []
@@ -825,7 +826,8 @@ def main():
    # Eval after training
    if training_args.do_eval:
        num_eval_samples = len(tokenized_datasets["validation"])
-        eval_samples_idx = jnp.arange(num_eval_samples)
+        # Avoid using jax.numpy here in case of TPU training
+        eval_samples_idx = np.arange(num_eval_samples)
        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)

        eval_metrics = []
--- a/examples/flax/language-modeling/run_t5_mlm_flax.py
+++ b/examples/flax/language-modeling/run_t5_mlm_flax.py
@@ -459,7 +459,7 @@ class FlaxDataCollatorForT5MLM:
        return is_noise[:orig_length]


-def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
+def generate_batch_splits(samples_idx: np.ndarray, batch_size: int) -> np.ndarray:
    num_samples = len(samples_idx)
    samples_to_remove = num_samples % batch_size

@@ -739,7 +739,6 @@ def main():
            config,
            seed=training_args.seed,
            dtype=getattr(jnp, model_args.dtype),
-            use_auth_token=True if model_args.use_auth_token else None,
        )

    # Data collator
@@ -871,6 +870,7 @@ def main():

        # Generate an epoch by shuffling sampling indices from the train dataset
        num_train_samples = len(tokenized_datasets["train"])
+        # Avoid using jax.numpy here in case of TPU training
        train_samples_idx = np.random.permutation(np.arange(num_train_samples))
        train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size)

@@ -908,7 +908,8 @@ def main():
            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
                # ======================== Evaluating ==============================
                num_eval_samples = len(tokenized_datasets["validation"])
-                eval_samples_idx = jnp.arange(num_eval_samples)
+                # Avoid using jax.numpy here in case of TPU training
+                eval_samples_idx = np.arange(num_eval_samples)
                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)

                eval_metrics = []
@@ -944,7 +945,8 @@ def main():
    # Eval after training
    if training_args.do_eval:
        num_eval_samples = len(tokenized_datasets["validation"])
-        eval_samples_idx = jnp.arange(num_eval_samples)
+        # Avoid using jax.numpy here in case of TPU training
+        eval_samples_idx = np.arange(num_eval_samples)
        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)

        eval_metrics = []
--- a/Show More
+++ b/Show More