Patch release: v4.9.2

* Fix tied weights on TPU

* Manually tie weights in no trainer examples

* Fix for test

* One last missing

* Gettning owned by my scripts

* Address review comments

* Fix test

* Fix tests

* Fix reformer tests

.circleci/TROUBLESHOOT.md

@@ -0,0 +1,7 @@
+# Troubleshooting
+
+This is a document explaining how to deal with various issues on Circle-CI. The entries may include actually solutions or pointers to Issues that cover those.
+
+## Circle CI
+
+* pytest worker runs out of resident RAM and gets killed by `cgroups`: https://github.com/huggingface/transformers/issues/11408

.circleci/config.yml

@@ -81,12 +81,18 @@ jobs:
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,tf-cpu,torch,testing,sentencepiece,speech,vision]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.9.0+cpu.html
             - save_cache:
                 key: v0.4-{{ checksum "setup.py" }}
                 paths:
                     - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_torch_and_tf ./tests/ -m is_pt_tf_cross_test --durations=0 | tee tests_output.txt
+            - run: python utils/tests_fetcher.py | tee test_preparation.txt
+            - store_artifacts:
+                  path: ~/transformers/test_preparation.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_torch_and_tf $(cat test_list.txt) -m is_pt_tf_cross_test --durations=0 | tee tests_output.txt
+                  fi
             - store_artifacts:
                   path: ~/transformers/tests_output.txt
             - store_artifacts:
@@ -111,12 +117,18 @@ jobs:
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,flax,torch,testing,sentencepiece,speech,vision]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.9.0+cpu.html
             - save_cache:
                 key: v0.4-{{ checksum "setup.py" }}
                 paths:
                     - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_torch_and_flax ./tests/ -m is_pt_flax_cross_test --durations=0 | tee tests_output.txt
+            - run: python utils/tests_fetcher.py | tee test_preparation.txt
+            - store_artifacts:
+                  path: ~/transformers/test_preparation.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_torch_and_flax $(cat test_list.txt) -m is_pt_flax_cross_test --durations=0 | tee tests_output.txt
+                  fi
             - store_artifacts:
                   path: ~/transformers/tests_output.txt
             - store_artifacts:
@@ -139,13 +151,19 @@ jobs:
                       - v0.4-{{ checksum "setup.py" }}
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
             - run: pip install --upgrade pip
-            - run: pip install .[sklearn,torch,testing,sentencepiece,speech,vision]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
+            - run: pip install .[sklearn,torch,testing,sentencepiece,speech,vision,timm]
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.9.0+cpu.html
             - save_cache:
                   key: v0.4-torch-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -n 3 --dist=loadfile -s --make-reports=tests_torch ./tests/ | tee tests_output.txt
+            - run: python utils/tests_fetcher.py | tee test_preparation.txt
+            - store_artifacts:
+                  path: ~/transformers/test_preparation.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    python -m pytest -n 3 --dist=loadfile -s --make-reports=tests_torch $(cat test_list.txt) | tee tests_output.txt
+                  fi
             - store_artifacts:
                   path: ~/transformers/tests_output.txt
             - store_artifacts:
@@ -172,7 +190,13 @@ jobs:
                   key: v0.4-tf-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_tf ./tests/ | tee tests_output.txt
+            - run: python utils/tests_fetcher.py | tee test_preparation.txt
+            - store_artifacts:
+                  path: ~/transformers/test_preparation.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_tf $(cat test_list.txt) | tee tests_output.txt
+                  fi
             - store_artifacts:
                   path: ~/transformers/tests_output.txt
             - store_artifacts:
@@ -199,7 +223,13 @@ jobs:
                   key: v0.4-flax-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_flax ./tests/ | tee tests_output.txt
+            - run: python utils/tests_fetcher.py | tee test_preparation.txt
+            - store_artifacts:
+                  path: ~/transformers/test_preparation.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_flax $(cat test_list.txt) | tee tests_output.txt
+                  fi
             - store_artifacts:
                   path: ~/transformers/tests_output.txt
             - store_artifacts:
@@ -224,12 +254,18 @@ jobs:
             - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
             - run: pip install --upgrade pip
             - run: pip install .[sklearn,torch,testing,sentencepiece,speech,vision]
-            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
+            - run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.9.0+cpu.html
             - save_cache:
                   key: v0.4-torch-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_pipelines_torch -m is_pipeline_test ./tests/ | tee tests_output.txt
+            - run: python utils/tests_fetcher.py | tee test_preparation.txt
+            - store_artifacts:
+                  path: ~/transformers/test_preparation.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_pipelines_torch -m is_pipeline_test $(cat test_list.txt) | tee tests_output.txt
+                  fi
             - store_artifacts:
                   path: ~/transformers/tests_output.txt
             - store_artifacts:
@@ -257,7 +293,13 @@ jobs:
                   key: v0.4-tf-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_pipelines_tf ./tests/ -m is_pipeline_test | tee tests_output.txt
+            - run: python utils/tests_fetcher.py | tee test_preparation.txt
+            - store_artifacts:
+                  path: ~/transformers/test_preparation.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_pipelines_tf $(cat test_list.txt) -m is_pipeline_test | tee tests_output.txt
+                  fi
             - store_artifacts:
                   path: ~/transformers/tests_output.txt
             - store_artifacts:
@@ -283,7 +325,10 @@ jobs:
                   key: v0.4-custom_tokenizers-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -s --make-reports=tests_custom_tokenizers ./tests/test_tokenization_bert_japanese.py | tee tests_output.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    python -m pytest -s --make-reports=tests_custom_tokenizers ./tests/test_tokenization_bert_japanese.py | tee tests_output.txt
+                  fi
             - store_artifacts:
                   path: ~/transformers/tests_output.txt
             - store_artifacts:
@@ -311,7 +356,13 @@ jobs:
                   key: v0.4-torch_examples-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: TRANSFORMERS_IS_CI=1 python -m pytest -n 8 --dist=loadfile -s --make-reports=examples_torch ./examples/pytorch/ | tee examples_output.txt
+            - run: python utils/tests_fetcher.py | tee test_preparation.txt
+            - store_artifacts:
+                  path: ~/transformers/test_preparation.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    TRANSFORMERS_IS_CI=1 python -m pytest -n 8 --dist=loadfile -s --make-reports=examples_torch ./examples/pytorch/ | tee examples_output.txt
+                  fi
             - store_artifacts:
                   path: ~/transformers/examples_output.txt
             - store_artifacts:
@@ -343,12 +394,51 @@ jobs:
                   key: v0.4-hub-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -sv ./tests/ -m is_staging_test
+            - run: python utils/tests_fetcher.py | tee test_preparation.txt
+            - store_artifacts:
+                  path: ~/transformers/test_preparation.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    python -m pytest -sv $(cat test_list.txt) -m is_staging_test
+                  fi
 
+    run_tests_onnxruntime:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.7
+        environment:
+            OMP_NUM_THREADS: 1
+            TRANSFORMERS_IS_CI: yes
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - restore_cache:
+                  keys:
+                      - v0.4-torch-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
+            - run: pip install --upgrade pip
+            - run: pip install .[torch,testing,sentencepiece,onnxruntime]
+            - save_cache:
+                  key: v0.4-onnx-{{ checksum "setup.py" }}
+                  paths:
+                      - '~/.cache/pip'
+            - run: python utils/tests_fetcher.py | tee test_preparation.txt
+            - store_artifacts:
+                  path: ~/transformers/test_preparation.txt
+            - run: |
+                  if [ -f test_list.txt ]; then
+                    python -m pytest -n 1 --dist=loadfile -s --make-reports=tests_torch $(cat test_list.txt) -k onnx | tee tests_output.txt
+                  fi
+            - store_artifacts:
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
     build_doc:
         working_directory: ~/transformers
         docker:
             - image: circleci/python:3.6
+        resource_class: large
         steps:
             - checkout
             - restore_cache:
@@ -370,6 +460,7 @@ jobs:
         working_directory: ~/transformers
         docker:
             - image: circleci/python:3.6
+        resource_class: large
         steps:
             - add_ssh_keys:
                 fingerprints:
@@ -379,6 +470,8 @@ jobs:
                   keys:
                       - v0.4-deploy_doc-{{ checksum "setup.py" }}
                       - v0.4-{{ checksum "setup.py" }}
+            - run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
+            - run: pip install --upgrade pip
             - run: pip install ."[docs]"
             - save_cache:
                   key: v0.4-deploy_doc-{{ checksum "setup.py" }}
@@ -390,7 +483,7 @@ jobs:
         working_directory: ~/transformers
         docker:
             - image: circleci/python:3.6
-        resource_class: medium
+        resource_class: large
         environment:
             TRANSFORMERS_IS_CI: yes
         parallelism: 1
@@ -401,7 +494,7 @@ jobs:
                       - v0.4-code_quality-{{ checksum "setup.py" }}
                       - v0.4-{{ checksum "setup.py" }}
             - run: pip install --upgrade pip
-            - run: pip install isort
+            - run: pip install isort GitPython
             - run: pip install .[all,quality]
             - save_cache:
                   key: v0.4-code_quality-{{ checksum "setup.py" }}
@@ -417,6 +510,8 @@ jobs:
             - run: python utils/check_dummies.py
             - run: python utils/check_repo.py
             - run: python utils/check_inits.py
+            - run: make deps_table_check_updated
+            - run: python utils/tests_fetcher.py --sanity_check
 
     check_repository_consistency:
         working_directory: ~/transformers
@@ -480,6 +575,7 @@ workflows:
             - run_tests_flax
             - run_tests_pipelines_torch
             - run_tests_pipelines_tf
+            - run_tests_onnxruntime
             - run_tests_hub
             - build_doc
             - deploy_doc: *workflow_filters

.circleci/deploy.sh

@@ -62,4 +62,10 @@ deploy_doc "c988db5" v4.4.0
 deploy_doc "c5d6a28" v4.4.1
 deploy_doc "6bc89ed" v4.4.2
 deploy_doc "4906a29" v4.5.0
-deploy_doc "4bae96e"  # v4.5.1 Latest stable release
+deploy_doc "4bae96e" v4.5.1
+deploy_doc "25dee4a" v4.6.0
+deploy_doc "7a6c9fa" v4.7.0
+deploy_doc "9252a51" v4.8.0
+deploy_doc "1366172" v4.8.1
+deploy_doc "96d1cfb" v4.8.2
+deploy_doc "72aee83" # v4.9.0 Latest stable release

.github/conda/meta.yaml

@@ -26,6 +26,7 @@ requirements:
     - regex !=2019.12.17
     - protobuf
     - tokenizers >=0.10.1,<0.11.0
+    - pyyaml >=5.1
   run:
     - python
     - numpy >=1.17
@@ -40,6 +41,7 @@ requirements:
     - regex !=2019.12.17
     - protobuf
     - tokenizers >=0.10.1,<0.11.0
+    - pyyaml >=5.1
 
 test:
   imports:

.github/workflows/TROUBLESHOOT.md

@@ -0,0 +1,9 @@
+# Troubleshooting
+
+This is a document explaining how to deal with various issues on github-actions self-hosted CI. The entries may include actually solutions or pointers to Issues that cover those.
+
+## GitHub Actions (self-hosted CI)
+
+* Deepspeed
+
+  - if jit build hangs, clear out `rm -rf ~/.cache/torch_extensions/` reference: https://github.com/huggingface/transformers/pull/12723

.github/workflows/github-torch-hub.yml

@@ -37,10 +37,10 @@ jobs:
         # no longer needed
         pip uninstall -y transformers
 
-    - name: Torch hub list
-      run: |
-        python -c "import torch; print(torch.hub.list('huggingface/transformers:$BRANCH'))"
+    #- name: Torch hub list
+    #  run: |
+    #    python -c "import torch; print(torch.hub.list('huggingface/transformers:$BRANCH'))"
 
-    - name: Torch hub help
-      run: |
-        python -c "import torch; print(torch.hub.help('huggingface/transformers:$BRANCH', 'modelForSequenceClassification'))"
+    #- name: Torch hub help
+    #  run: |
+    #    python -c "import torch; print(torch.hub.help('huggingface/transformers:$BRANCH', 'modelForSequenceClassification'))"

.github/workflows/release-conda.yml

@@ -4,6 +4,8 @@ on:
   push:
     tags:
       - v*
+    branches:
+      - conda_*
 
 env:
   ANACONDA_API_TOKEN: ${{ secrets.ANACONDA_API_TOKEN }}

.github/workflows/self-push.yml

@@ -18,12 +18,13 @@ env:
   TRANSFORMERS_IS_CI: yes
   OMP_NUM_THREADS: 8
   MKL_NUM_THREADS: 8
+  PYTEST_TIMEOUT: 60
 
 jobs:
   run_tests_torch_gpu:
     runs-on: [self-hosted, docker-gpu, single-gpu]
     container:
-      image: pytorch/pytorch:1.8.0-cuda11.1-cudnn8-runtime
+      image: pytorch/pytorch:1.9.0-cuda11.1-cudnn8-runtime
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
     steps:
       - name: Launcher docker
@@ -37,7 +38,7 @@ jobs:
         run: |
           apt -y update && apt install -y libsndfile1-dev
           pip install --upgrade pip
-          pip install .[sklearn,testing,onnxruntime,sentencepiece,speech]
+          pip install .[sklearn,testing,onnxruntime,sentencepiece,speech,vision,timm]
 
       - name: Are GPUs recognized by our DL frameworks
         run: |
@@ -48,7 +49,7 @@ jobs:
 
       - name: Run all non-slow tests on GPU
         run: |
-          python -m pytest -n 2 --dist=loadfile --make-reports=tests_torch_gpu tests
+          python -m pytest -n 2 --dist=loadfile -v --make-reports=tests_torch_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -61,53 +62,53 @@ jobs:
           name: run_all_tests_torch_gpu_test_reports
           path: reports
 
-  run_tests_tf_gpu:
-    runs-on: [self-hosted, docker-gpu, single-gpu]
-    timeout-minutes: 120
-    container:
-      image: tensorflow/tensorflow:2.4.1-gpu
-      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: Install dependencies
-        run: |
-          pip install --upgrade pip
-          pip install .[sklearn,testing,onnxruntime,sentencepiece]
-
-      - name: Are GPUs recognized by our DL frameworks
-        run: |
-          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('TF GPUs available:', bool(tf.config.list_physical_devices('GPU')))"
-          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('Number of TF GPUs available:', len(tf.config.list_physical_devices('GPU')))"
-
-      - name: Run all non-slow tests on GPU
-        env:
-          TF_NUM_INTRAOP_THREADS: 8
-          TF_NUM_INTEROP_THREADS: 1
-        run: |
-          python -m pytest -n 1 --dist=loadfile --make-reports=tests_tf_gpu tests
-
-      - name: Failure short reports
-        if: ${{ always() }}
-        run: cat reports/tests_tf_gpu_failures_short.txt
-
-      - name: Test suite reports artifacts
-        if: ${{ always() }}
-        uses: actions/upload-artifact@v2
-        with:
-          name: run_all_tests_tf_gpu_test_reports
-          path: reports
+#  run_tests_tf_gpu:
+#    runs-on: [self-hosted, docker-gpu, single-gpu]
+#    timeout-minutes: 120
+#    container:
+#      image: tensorflow/tensorflow:2.4.1-gpu
+#      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: Install dependencies
+#        run: |
+#          pip install --upgrade pip
+#          pip install .[sklearn,testing,onnxruntime,sentencepiece]
+#
+#      - name: Are GPUs recognized by our DL frameworks
+#        run: |
+#          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('TF GPUs available:', bool(tf.config.list_physical_devices('GPU')))"
+#          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('Number of TF GPUs available:', len(tf.config.list_physical_devices('GPU')))"
+#
+#      - name: Run all non-slow tests on GPU
+#        env:
+#          TF_NUM_INTRAOP_THREADS: 8
+#          TF_NUM_INTEROP_THREADS: 1
+#        run: |
+#          python -m pytest -n 1 --dist=loadfile --make-reports=tests_tf_gpu tests
+#
+#      - name: Failure short reports
+#        if: ${{ always() }}
+#        run: cat reports/tests_tf_gpu_failures_short.txt
+#
+#      - name: Test suite reports artifacts
+#        if: ${{ always() }}
+#        uses: actions/upload-artifact@v2
+#        with:
+#          name: run_all_tests_tf_gpu_test_reports
+#          path: reports
 
 
   run_tests_torch_multi_gpu:
     runs-on: [self-hosted, docker-gpu, multi-gpu]
     container:
-      image: pytorch/pytorch:1.8.0-cuda11.1-cudnn8-runtime
+      image: pytorch/pytorch:1.9.0-cuda11.1-cudnn8-runtime
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
     steps:
       - name: Launcher docker
@@ -121,7 +122,7 @@ jobs:
         run: |
           apt -y update && apt install -y libsndfile1-dev
           pip install --upgrade pip
-          pip install .[sklearn,testing,onnxruntime,sentencepiece,speech]
+          pip install .[sklearn,testing,onnxruntime,sentencepiece,speech,vision,timm]
 
       - name: Are GPUs recognized by our DL frameworks
         run: |
@@ -134,7 +135,7 @@ jobs:
         env:
           MKL_SERVICE_FORCE_INTEL: 1
         run: |
-          python -m pytest -n 2 --dist=loadfile --make-reports=tests_torch_multi_gpu tests
+          python -m pytest -n 2 --dist=loadfile -v --make-reports=tests_torch_multi_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -147,47 +148,47 @@ jobs:
           name: run_all_tests_torch_multi_gpu_test_reports
           path: reports
 
-  run_tests_tf_multi_gpu:
-    runs-on: [self-hosted, docker-gpu, multi-gpu]
-    timeout-minutes: 120
-    container:
-      image: tensorflow/tensorflow:2.4.1-gpu
-      options: --gpus all --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: Install dependencies
-        run: |
-          pip install --upgrade pip
-          pip install .[sklearn,testing,onnxruntime,sentencepiece]
-
-      - name: Are GPUs recognized by our DL frameworks
-        run: |
-          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('TF GPUs available:', bool(tf.config.list_physical_devices('GPU')))"
-          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('Number of TF GPUs available:', len(tf.config.list_physical_devices('GPU')))"
-
-      - name: Run all non-slow tests on GPU
-        env:
-          TF_NUM_INTRAOP_THREADS: 8
-          TF_NUM_INTEROP_THREADS: 1
-        run: |
-          python -m pytest -n 1 --dist=loadfile --make-reports=tests_tf_multi_gpu tests
-
-      - name: Failure short reports
-        if: ${{ always() }}
-        run: cat reports/tests_tf_multi_gpu_failures_short.txt
-
-      - name: Test suite reports artifacts
-        if: ${{ always() }}
-        uses: actions/upload-artifact@v2
-        with:
-          name: run_all_tests_tf_multi_gpu_test_reports
-          path: reports
+#  run_tests_tf_multi_gpu:
+#    runs-on: [self-hosted, docker-gpu, multi-gpu]
+#    timeout-minutes: 120
+#    container:
+#      image: tensorflow/tensorflow:2.4.1-gpu
+#      options: --gpus all --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: Install dependencies
+#        run: |
+#          pip install --upgrade pip
+#          pip install .[sklearn,testing,onnxruntime,sentencepiece]
+#
+#      - name: Are GPUs recognized by our DL frameworks
+#        run: |
+#          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('TF GPUs available:', bool(tf.config.list_physical_devices('GPU')))"
+#          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('Number of TF GPUs available:', len(tf.config.list_physical_devices('GPU')))"
+#
+#      - name: Run all non-slow tests on GPU
+#        env:
+#          TF_NUM_INTRAOP_THREADS: 8
+#          TF_NUM_INTEROP_THREADS: 1
+#        run: |
+#          python -m pytest -n 1 --dist=loadfile --make-reports=tests_tf_multi_gpu tests
+#
+#      - name: Failure short reports
+#        if: ${{ always() }}
+#        run: cat reports/tests_tf_multi_gpu_failures_short.txt
+#
+#      - name: Test suite reports artifacts
+#        if: ${{ always() }}
+#        uses: actions/upload-artifact@v2
+#        with:
+#          name: run_all_tests_tf_multi_gpu_test_reports
+#          path: reports
 
   run_tests_torch_cuda_extensions_gpu:
     runs-on: [self-hosted, docker-gpu, single-gpu]
@@ -217,7 +218,7 @@ jobs:
 
       - name: Run all tests on GPU
         run: |
-          python -m pytest -n 1 --dist=loadfile --make-reports=tests_torch_cuda_extensions_gpu tests/deepspeed tests/extended
+          python -m pytest -n 1 --dist=loadfile -v --make-reports=tests_torch_cuda_extensions_gpu tests/deepspeed tests/extended
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -258,7 +259,7 @@ jobs:
 
       - name: Run all tests on GPU
         run: |
-          python -m pytest -n 1 --dist=loadfile --make-reports=tests_torch_cuda_extensions_multi_gpu tests/deepspeed tests/extended
+          python -m pytest -n 1 --dist=loadfile -v --make-reports=tests_torch_cuda_extensions_multi_gpu tests/deepspeed tests/extended
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -278,9 +279,9 @@ jobs:
     if: always()
     needs: [
         run_tests_torch_gpu,
-        run_tests_tf_gpu,
+#        run_tests_tf_gpu,
         run_tests_torch_multi_gpu,
-        run_tests_tf_multi_gpu,
+#        run_tests_tf_multi_gpu,
         run_tests_torch_cuda_extensions_gpu,
         run_tests_torch_cuda_extensions_multi_gpu
     ]

.github/workflows/self-scheduled.yml

@@ -14,12 +14,13 @@ env:
   RUN_SLOW: yes
   OMP_NUM_THREADS: 16
   MKL_NUM_THREADS: 16
+  PYTEST_TIMEOUT: 600
 
 jobs:
   run_all_tests_torch_gpu:
     runs-on: [self-hosted, docker-gpu, single-gpu]
     container:
-      image: pytorch/pytorch:1.8.0-cuda11.1-cudnn8-runtime
+      image: pytorch/pytorch:1.9.0-cuda11.1-cudnn8-runtime
       options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
     steps:
       - name: Launcher docker
@@ -33,7 +34,7 @@ jobs:
         run: |
           apt -y update && apt install -y libsndfile1-dev
           pip install --upgrade pip
-          pip install .[sklearn,testing,onnxruntime,sentencepiece,speech]
+          pip install .[integrations,sklearn,testing,onnxruntime,sentencepiece,speech,vision,timm]
 
       - name: Are GPUs recognized by our DL frameworks
         run: |
@@ -44,7 +45,7 @@ jobs:
 
       - name: Run all tests on GPU
         run: |
-          python -m pytest -n 1 --dist=loadfile --make-reports=tests_torch_gpu tests
+          python -m pytest -n 1 -v --dist=loadfile --make-reports=tests_torch_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -60,7 +61,7 @@ jobs:
           TRANSFORMERS_IS_CI: yes
         run: |
           pip install -r examples/pytorch/_tests_requirements.txt
-          python -m pytest -n 1 --dist=loadfile --make-reports=examples_torch_gpu examples
+          python -m pytest -n 1 -v --dist=loadfile --make-reports=examples_torch_gpu examples
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -71,7 +72,7 @@ jobs:
         env:
           RUN_PIPELINE_TESTS: yes
         run: |
-          python -m pytest -n 1 --dist=loadfile -m is_pipeline_test --make-reports=tests_torch_pipeline_gpu tests
+          python -m pytest -n 1 -v --dist=loadfile -m is_pipeline_test --make-reports=tests_torch_pipeline_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -112,7 +113,7 @@ jobs:
           TF_NUM_INTEROP_THREADS: 1
           TF_NUM_INTRAOP_THREADS: 16
         run: |
-          python -m pytest -n 1 --dist=loadfile --make-reports=tests_tf_gpu tests
+          python -m pytest -n 1 -v --dist=loadfile --make-reports=tests_tf_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -125,7 +126,7 @@ jobs:
           TF_NUM_INTEROP_THREADS: 1
           TF_NUM_INTRAOP_THREADS: 16
         run: |
-          python -m pytest -n 1 --dist=loadfile -m is_pipeline_test --make-reports=tests_tf_pipeline_gpu tests
+          python -m pytest -n 1 -v --dist=loadfile -m is_pipeline_test --make-reports=tests_tf_pipeline_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -141,7 +142,7 @@ jobs:
   run_all_tests_torch_multi_gpu:
     runs-on: [self-hosted, docker-gpu, multi-gpu]
     container:
-      image: pytorch/pytorch:1.8.0-cuda11.1-cudnn8-runtime
+      image: pytorch/pytorch:1.9.0-cuda11.1-cudnn8-runtime
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
     steps:
       - name: Launcher docker
@@ -155,7 +156,7 @@ jobs:
         run: |
           apt -y update && apt install -y libsndfile1-dev
           pip install --upgrade pip
-          pip install .[sklearn,testing,onnxruntime,sentencepiece,speech]
+          pip install .[integrations,sklearn,testing,onnxruntime,sentencepiece,speech,vision,timm]
 
       - name: Are GPUs recognized by our DL frameworks
         run: |
@@ -168,7 +169,7 @@ jobs:
         env:
           MKL_SERVICE_FORCE_INTEL: 1
         run: |
-          python -m pytest -n 1 --dist=loadfile --make-reports=tests_torch_multi_gpu tests
+          python -m pytest -n 1 -v --dist=loadfile --make-reports=tests_torch_multi_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -179,7 +180,7 @@ jobs:
         env:
           RUN_PIPELINE_TESTS: yes
         run: |
-          python -m pytest -n 1 --dist=loadfile -m is_pipeline_test --make-reports=tests_torch_pipeline_multi_gpu tests
+          python -m pytest -n 1 -v --dist=loadfile -m is_pipeline_test --make-reports=tests_torch_pipeline_multi_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -220,7 +221,7 @@ jobs:
           TF_NUM_INTEROP_THREADS: 1
           TF_NUM_INTRAOP_THREADS: 16
         run: |
-          python -m pytest -n 1 --dist=loadfile --make-reports=tests_tf_multi_gpu tests
+          python -m pytest -n 1 -v --dist=loadfile --make-reports=tests_tf_multi_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -233,7 +234,7 @@ jobs:
           TF_NUM_INTEROP_THREADS: 1
           TF_NUM_INTRAOP_THREADS: 16
         run: |
-          python -m pytest -n 1 --dist=loadfile -m is_pipeline_test --make-reports=tests_tf_pipeline_multi_gpu tests
+          python -m pytest -n 1 -v --dist=loadfile -m is_pipeline_test --make-reports=tests_tf_pipeline_multi_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -274,7 +275,7 @@ jobs:
 
       - name: Run all tests on GPU
         run: |
-          python -m pytest -n 1 --dist=loadfile --make-reports=tests_torch_cuda_extensions_gpu tests/deepspeed tests/extended
+          python -m pytest -n 1 -v --dist=loadfile --make-reports=tests_torch_cuda_extensions_gpu tests/deepspeed tests/extended
 
       - name: Failure short reports
         if: ${{ always() }}
@@ -315,7 +316,7 @@ jobs:
 
       - name: Run all tests on GPU
         run: |
-          python -m pytest -n 1 --dist=loadfile --make-reports=tests_torch_cuda_extensions_multi_gpu tests/deepspeed tests/extended
+          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() }}

CONTRIBUTING.md

@@ -37,7 +37,7 @@ There are 4 ways you can contribute to transformers:
 * Submitting issues related to bugs or desired new features.
 
 In particular there is a special [Good First
-Issue](https://github.com/huggingface/transformers/contribute) listing. Tt will give you a list of
+Issue](https://github.com/huggingface/transformers/contribute) listing. It will give you a list of
 open Issues that are open to anybody to work on. Just comment in the issue that you'd like to work
 on it. In that same listing you will also find some Issues with `Good Second Issue` label. These are
 typically slightly more complicated than the Issues with just `Good First Issue` label. But if you

Makefile

@@ -21,6 +21,12 @@ modified_only_fixup:
 deps_table_update:
 	@python setup.py deps_table_update
 
+deps_table_check_updated:
+	@md5sum src/transformers/dependency_versions_table.py > md5sum.saved
+	@python setup.py deps_table_update
+	@md5sum -c --quiet md5sum.saved || (printf "\nError: the version dependency table is outdated.\nPlease run 'make fixup' or 'make style' and commit the changes.\n\n" && exit 1)
+	@rm md5sum.saved
+
 # autogenerating code
 
 autogenerate_code: deps_table_update
@@ -34,6 +40,7 @@ extra_quality_checks:
 	python utils/check_dummies.py
 	python utils/check_repo.py
 	python utils/check_inits.py
+	python utils/tests_fetcher.py --sanity_check
 
 # this target runs checks on all files
 quality:

README.md

@@ -35,10 +35,23 @@ limitations under the License.
     <a href="https://github.com/huggingface/transformers/blob/master/CODE_OF_CONDUCT.md">
         <img alt="Contributor Covenant" src="https://img.shields.io/badge/Contributor%20Covenant-v2.0%20adopted-ff69b4.svg">
     </a>
+    <a href="https://zenodo.org/badge/latestdoi/155220641"><img src="https://zenodo.org/badge/155220641.svg" alt="DOI"></a>
 </p>
 
+<h4 align="center">
+    <p>
+        <b>English</b> |
+        <a href="https://github.com/huggingface/transformers/blob/master/README_zh-hans.md">简体中文</a> |
+        <a href="https://github.com/huggingface/transformers/blob/master/README_zh-hant.md">繁體中文</a>
+    <p>
+</h4>
+
 <h3 align="center">
-<p>State-of-the-art Natural Language Processing for Jax, PyTorch and TensorFlow
+    <p>State-of-the-art Natural Language Processing for Jax, PyTorch and TensorFlow</p>
+</h3>
+
+<h3 align="center">
+    <a href="https://hf.co/course"><img src="https://raw.githubusercontent.com/huggingface/transformers/master/docs/source/imgs/course_banner.png"></a>
 </h3>
 
 🤗 Transformers provides thousands of pretrained models to perform tasks on texts such as classification, information extraction, question answering, summarization, translation, text generation and more in over 100 languages. Its aim is to make cutting-edge NLP easier to use for everyone.
@@ -62,6 +75,12 @@ Here are a few examples:
 
 **[Write With Transformer](https://transformer.huggingface.co)**, built by the Hugging Face team, is the official demo of this repo’s text generation capabilities.
 
+## If you are looking for custom support from the Hugging Face team
+
+<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);">
+</a><br>
+
 ## Quick tour
 
 To immediately use a model on a given text, we provide the `pipeline` API. Pipelines group together a pretrained model with the preprocessing that was used during that model's training. Here is how to quickly use a pipeline to classify positive versus negative texts:
@@ -199,14 +218,17 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Blenderbot](https://huggingface.co/transformers/model_doc/blenderbot.html)** (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/transformers/model_doc/blenderbot_small.html)** (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](https://huggingface.co/transformers/model_doc/bort.html)** (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/transformers/model_doc/byt5.html)** (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/transformers/model_doc/camembert.html)** (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. **[CLIP](https://huggingface.co/transformers/model_doc/clip.html)** 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. **[CANINE](https://huggingface.co/transformers/model_doc/canine.html)** (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/transformers/model_doc/clip.html)** (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](https://huggingface.co/transformers/model_doc/convbert.html)** (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](https://huggingface.co/transformers/model_doc/cpm.html)** (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/transformers/model_doc/ctrl.html)** (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. **[DeBERTa](https://huggingface.co/transformers/model_doc/deberta.html)** (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/transformers/model_doc/deberta_v2.html)** (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. **[DeiT](https://huggingface.co/transformers/model_doc/deit.html)** (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](https://huggingface.co/transformers/model_doc/detr.html)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
 1. **[DialoGPT](https://huggingface.co/transformers/model_doc/dialogpt.html)** (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](https://huggingface.co/transformers/model_doc/distilbert.html)** (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/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
 1. **[DPR](https://huggingface.co/transformers/model_doc/dpr.html)** (from Facebook) released with the paper [Dense Passage Retrieval
@@ -218,6 +240,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 1. **[GPT](https://huggingface.co/transformers/model_doc/gpt.html)** (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](https://huggingface.co/transformers/model_doc/gpt2.html)** (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 Neo](https://huggingface.co/transformers/model_doc/gpt_neo.html)** (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](https://huggingface.co/transformers/model_doc/hubert.html)** (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/transformers/model_doc/ibert.html)** (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. **[LayoutLM](https://huggingface.co/transformers/model_doc/layoutlm.html)** (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. **[LED](https://huggingface.co/transformers/model_doc/led.html)** (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.
@@ -236,12 +259,14 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 1. **[ProphetNet](https://huggingface.co/transformers/model_doc/prophetnet.html)** (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. **[Reformer](https://huggingface.co/transformers/model_doc/reformer.html)** (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. **[RoBERTa](https://huggingface.co/transformers/model_doc/roberta.html)** (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.
+1. **[RoFormer](https://huggingface.co/transformers/model_doc/roformer.html)** (from ZhuiyiTechnology), released together with the paper a [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
 1. **[SpeechToTextTransformer](https://huggingface.co/transformers/model_doc/speech_to_text.html)** (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. **[SqueezeBert](https://huggingface.co/transformers/model_doc/squeezebert.html)** 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. **[T5](https://huggingface.co/transformers/model_doc/t5.html)** (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. **[TAPAS](https://huggingface.co/transformers/model_doc/tapas.html)** (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. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (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. **[Vision Transformer (ViT)](https://huggingface.co/transformers/model_doc/vit.html)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VisualBERT](https://huggingface.co/transformers/model_doc/visual_bert.html)** (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. **[Wav2Vec2](https://huggingface.co/transformers/model_doc/wav2vec2.html)** (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. **[XLM](https://huggingface.co/transformers/model_doc/xlm.html)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](https://huggingface.co/transformers/model_doc/xlmprophetnet.html)** (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.
@@ -250,7 +275,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 1. **[XLSR-Wav2Vec2](https://huggingface.co/transformers/model_doc/xlsr_wav2vec2.html)** (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. 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.
 
-To check if each model has an implementation in Flax, PyTorch or TensorFlow, or has an associated tokenizer backed by the 🤗 Tokenizers library, refer to [this table](https://huggingface.co/transformers/index.html#bigtable).
+To check if each model has an implementation in Flax, PyTorch or TensorFlow, or has an associated tokenizer backed by the 🤗 Tokenizers library, refer to [this table](https://huggingface.co/transformers/index.html#supported-frameworks).
 
 These implementations have been tested on several datasets (see the example scripts) and should match the performance of the original implementations. You can find more details on performance in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
 

README_zh-hans.md

@@ -0,0 +1,334 @@
+<!---
+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.
+-->
+
+<!---
+A useful guide for English-Chinese translation of Hugging Face documentation
+- Add space around English words and numbers when they appear between Chinese characters. E.g., 共 100 多种语言; 使用 transformers 库。
+- Use square quotes, e.g.,「引用」
+
+Dictionary
+
+Hugging Face: 抱抱脸
+token: 词符（并用括号标注原英文）
+tokenize: 词符化（并用括号标注原英文）
+tokenizer: 词符化器（并用括号标注原英文）
+transformer: transformer（不翻译）
+pipeline: 流水线
+API: API (不翻译）
+inference: 推理
+Trainer: 训练器。当作为类名出现时不翻译。
+pretrained/pretrain: 预训练
+finetune: 微调
+community: 社区
+example: 当特指仓库中 example 目录时翻译为「用例」
+Python data structures (e.g., list, set, dict): 翻译为列表，集合，词典，并用括号标注原英文
+NLP/Natural Language Processing: 以 NLP 出现时不翻译，以 Natural Language Processing 出现时翻译为自然语言处理
+checkpoint: 检查点
+-->
+
+<p align="center">
+    <br>
+    <img src="https://raw.githubusercontent.com/huggingface/transformers/master/docs/source/imgs/transformers_logo_name.png" width="400"/>
+    <br>
+<p>
+<p align="center">
+    <a href="https://circleci.com/gh/huggingface/transformers">
+        <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/master">
+    </a>
+    <a href="https://github.com/huggingface/transformers/blob/master/LICENSE">
+        <img alt="GitHub" src="https://img.shields.io/github/license/huggingface/transformers.svg?color=blue">
+    </a>
+    <a href="https://huggingface.co/transformers/index.html">
+        <img alt="Documentation" src="https://img.shields.io/website/http/huggingface.co/transformers/index.html.svg?down_color=red&down_message=offline&up_message=online">
+    </a>
+    <a href="https://github.com/huggingface/transformers/releases">
+        <img alt="GitHub release" src="https://img.shields.io/github/release/huggingface/transformers.svg">
+    </a>
+    <a href="https://github.com/huggingface/transformers/blob/master/CODE_OF_CONDUCT.md">
+        <img alt="Contributor Covenant" src="https://img.shields.io/badge/Contributor%20Covenant-v2.0%20adopted-ff69b4.svg">
+    </a>
+    <a href="https://zenodo.org/badge/latestdoi/155220641"><img src="https://zenodo.org/badge/155220641.svg" alt="DOI"></a>
+</p>
+
+<h4 align="center">
+    <p>
+        <a href="https://github.com/huggingface/transformers/">English</a> |
+        <b>简体中文</b> |
+        <a href="https://github.com/huggingface/transformers/blob/master/README_zh-hant.md">繁體中文</a>
+    <p>
+</h4>
+
+<h3 align="center">
+    <p>为 Jax、PyTorch 和 TensorFlow 打造的先进的自然语言处理</p>
+</h3>
+
+<h3 align="center">
+    <a href="https://hf.co/course"><img src="https://raw.githubusercontent.com/huggingface/transformers/master/docs/source/imgs/course_banner.png"></a>
+</h3>
+
+🤗 Transformers 提供了数以千计的预训练模型，支持 100 多种语言的文本分类、信息抽取、问答、摘要、翻译、文本生成。它的宗旨让最先进的 NLP 技术人人易用。
+
+🤗 Transformers 提供了便于快速下载和使用的API，让你可以把预训练模型用在给定文本、在你的数据集上微调然后通过 [model hub](https://huggingface.co/models) 与社区共享。同时，每个定义的 Python 模块均完全独立，方便修改和快速研究实验。
+
+🤗 Transformers 支持三个最热门的深度学习库： [Jax](https://jax.readthedocs.io/en/latest/), [PyTorch](https://pytorch.org/) and [TensorFlow](https://www.tensorflow.org/) — 并与之无缝整合。你可以直接使用一个框架训练你的模型然后用另一个加载和推理。
+
+## 在线演示
+
+你可以直接在模型页面上测试大多数 [model hub](https://huggingface.co/models) 上的模型。 我们也提供了 [私有模型托管、模型版本管理以及推理API](https://huggingface.co/pricing)。
+
+这里是一些例子：
+- [用 BERT 做掩码填词](https://huggingface.co/bert-base-uncased?text=Paris+is+the+%5BMASK%5D+of+France)
+- [用 Electra 做命名实体识别](https://huggingface.co/dbmdz/electra-large-discriminator-finetuned-conll03-english?text=My+name+is+Sarah+and+I+live+in+London+city)
+- [用 GPT-2 做文本生成](https://huggingface.co/gpt2?text=A+long+time+ago%2C+)
+- [用 RoBERTa 做自然语言推理](https://huggingface.co/roberta-large-mnli?text=The+dog+was+lost.+Nobody+lost+any+animal)
+- [用 BART 做文本摘要](https://huggingface.co/facebook/bart-large-cnn?text=The+tower+is+324+metres+%281%2C063+ft%29+tall%2C+about+the+same+height+as+an+81-storey+building%2C+and+the+tallest+structure+in+Paris.+Its+base+is+square%2C+measuring+125+metres+%28410+ft%29+on+each+side.+During+its+construction%2C+the+Eiffel+Tower+surpassed+the+Washington+Monument+to+become+the+tallest+man-made+structure+in+the+world%2C+a+title+it+held+for+41+years+until+the+Chrysler+Building+in+New+York+City+was+finished+in+1930.+It+was+the+first+structure+to+reach+a+height+of+300+metres.+Due+to+the+addition+of+a+broadcasting+aerial+at+the+top+of+the+tower+in+1957%2C+it+is+now+taller+than+the+Chrysler+Building+by+5.2+metres+%2817+ft%29.+Excluding+transmitters%2C+the+Eiffel+Tower+is+the+second+tallest+free-standing+structure+in+France+after+the+Millau+Viaduct)
+- [用 DistilBERT 做问答](https://huggingface.co/distilbert-base-uncased-distilled-squad?text=Which+name+is+also+used+to+describe+the+Amazon+rainforest+in+English%3F&context=The+Amazon+rainforest+%28Portuguese%3A+Floresta+Amaz%C3%B4nica+or+Amaz%C3%B4nia%3B+Spanish%3A+Selva+Amaz%C3%B3nica%2C+Amazon%C3%ADa+or+usually+Amazonia%3B+French%3A+For%C3%AAt+amazonienne%3B+Dutch%3A+Amazoneregenwoud%29%2C+also+known+in+English+as+Amazonia+or+the+Amazon+Jungle%2C+is+a+moist+broadleaf+forest+that+covers+most+of+the+Amazon+basin+of+South+America.+This+basin+encompasses+7%2C000%2C000+square+kilometres+%282%2C700%2C000+sq+mi%29%2C+of+which+5%2C500%2C000+square+kilometres+%282%2C100%2C000+sq+mi%29+are+covered+by+the+rainforest.+This+region+includes+territory+belonging+to+nine+nations.+The+majority+of+the+forest+is+contained+within+Brazil%2C+with+60%25+of+the+rainforest%2C+followed+by+Peru+with+13%25%2C+Colombia+with+10%25%2C+and+with+minor+amounts+in+Venezuela%2C+Ecuador%2C+Bolivia%2C+Guyana%2C+Suriname+and+French+Guiana.+States+or+departments+in+four+nations+contain+%22Amazonas%22+in+their+names.+The+Amazon+represents+over+half+of+the+planet%27s+remaining+rainforests%2C+and+comprises+the+largest+and+most+biodiverse+tract+of+tropical+rainforest+in+the+world%2C+with+an+estimated+390+billion+individual+trees+divided+into+16%2C000+species)
+- [用 T5 做翻译](https://huggingface.co/t5-base?text=My+name+is+Wolfgang+and+I+live+in+Berlin)
+
+**[Write With Transformer](https://transformer.huggingface.co)**，由抱抱脸团队打造，是一个文本生成的官方 demo。
+
+## 如果你在寻找由抱抱脸团队提供的定制化支持服务
+
+<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);">
+</a><br>
+
+## 快速上手
+
+我们为快速使用模型提供了 `pipeline` （流水线）API。流水线聚合了预训练模型和对应的文本预处理。下面是一个快速使用流水线去判断正负面情绪的例子：
+
+```python
+>>> from transformers import pipeline
+
+# 使用情绪分析流水线
+>>> classifier = pipeline('sentiment-analysis')
+>>> classifier('We are very happy to introduce pipeline to the transformers repository.')
+[{'label': 'POSITIVE', 'score': 0.9996980428695679}]
+```
+
+第二行代码下载并缓存了流水线使用的预训练模型，而第三行代码则在给定的文本上进行了评估。这里的答案“正面” (positive) 具有 99 的置信度。
+
+许多的 NLP 任务都有开箱即用的预训练流水线。比如说，我们可以轻松的从给定文本中抽取问题答案：
+
+``` python
+>>> from transformers import pipeline
+
+# 使用问答流水线
+>>> 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'}
+
+```
+
+除了给出答案，预训练模型还给出了对应的置信度分数、答案在词符化 (tokenized) 后的文本中开始和结束的位置。你可以从[这个教程](https://huggingface.co/transformers/task_summary.html)了解更多流水线API支持的任务。
+
+要在你的任务上下载和使用任意预训练模型也很简单，只需三行代码。这里是 PyTorch 版的示例：
+```python
+>>> from transformers import AutoTokenizer, AutoModel
+
+>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+>>> model = AutoModel.from_pretrained("bert-base-uncased")
+
+>>> inputs = tokenizer("Hello world!", return_tensors="pt")
+>>> outputs = model(**inputs)
+```
+这里是等效的 TensorFlow 代码：
+```python
+>>> from transformers import AutoTokenizer, TFAutoModel
+
+>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+>>> model = TFAutoModel.from_pretrained("bert-base-uncased")
+
+>>> inputs = tokenizer("Hello world!", return_tensors="tf")
+>>> outputs = model(**inputs)
+```
+
+词符化器 (tokenizer) 为所有的预训练模型提供了预处理，并可以直接对单个字符串进行调用（比如上面的例子）或对列表 (list) 调用。它会输出一个你可以在下游代码里使用或直接通过 `**` 解包表达式传给模型的词典 (dict)。
+
+模型本身是一个常规的 [Pytorch `nn.Module`](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) 或 [TensorFlow `tf.keras.Model`](https://www.tensorflow.org/api_docs/python/tf/keras/Model)（取决于你的后端），可以常规方式使用。 [这个教程](https://huggingface.co/transformers/training.html)解释了如何将这样的模型整合到经典的 PyTorch 或 TensorFlow 训练循环中，或是如何使用我们的 `Trainer` 训练器）API 来在一个新的数据集上快速微调。
+
+## 为什么要用 transformers？
+
+1. 便于使用的先进模型：
+    - NLU 和 NLG 上表现优越
+    - 对教学和实践友好且低门槛
+    - 高级抽象，只需了解三个类
+    - 对所有模型统一的API
+
+1. 更低计算开销，更少的碳排放：
+    - 研究人员可以分享亿训练的模型而非次次从头开始训练
+    - 工程师可以减少计算用时和生产环境开销
+    - 数十种模型架构、两千多个预训练模型、100多种语言支持
+
+1. 对于模型生命周期的每一个部分都面面俱到：
+    - 训练先进的模型，只需 3 行代码
+    - 模型在不同深度学习框架间任意转移，随你心意
+    - 为训练、评估和生产选择最适合的框架，衔接无缝
+
+1. 为你的需求轻松定制专属模型和用例：
+    - 我们为每种模型架构提供了多个用例来复现原论文结果
+    - 模型内部结构保持透明一致
+    - 模型文件可单独使用，方便魔改和快速实验
+
+## 什么情况下我不该用 transformers？
+
+- 本库并不是模块化的神经网络工具箱。模型文件中的代码特意呈若璞玉，未经额外抽象封装，以便研究人员快速迭代魔改而不致溺于抽象和文件跳转之中。
+- `Trainer` API 并非兼容任何模型，只为本库之模型优化。若是在寻找适用于通用机器学习的训练循环实现，请另觅他库。
+- 尽管我们已尽力而为，[examples 目录](https://github.com/huggingface/transformers/tree/master/examples)中的脚本也仅为用例而已。对于你的特定问题，它们并不一定开箱即用，可能需要改几行代码以适之。
+
+## 安装
+
+### 使用 pip
+
+这个仓库已在 Python 3.6+、Flax 0.3.2+、PyTorch 1.3.1+ 和 TensorFlow 2.3+ 下经过测试。
+
+你可以在[虚拟环境](https://docs.python.org/3/library/venv.html)中安装 🤗 Transformers。如果你还不熟悉 Python 的虚拟环境，请阅此[用户说明](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/)。
+
+首先，用你打算使用的版本的 Python 创建一个虚拟环境并激活。
+
+然后，你需要安装 Flax、PyTorch 或 TensorFlow 其中之一。关于在你使用的平台上安装这些框架，请参阅 [TensorFlow 安装页](https://www.tensorflow.org/install/), [PyTorch 安装页](https://pytorch.org/get-started/locally/#start-locally) 或 [Flax 安装页](https://github.com/google/flax#quick-install)。
+
+当这些后端之一安装成功后， 🤗 Transformers 可依此安装：
+
+```bash
+pip install transformers
+```
+
+如果你想要试试用例或者想在正式发布前使用最新的开发中代码，你得[从源代码安装](https://huggingface.co/transformers/installation.html#installing-from-source)。
+
+### 使用 conda
+
+自 Transformers 4.0.0 版始，我们有了一个 conda 频道： `huggingface`。
+
+🤗 Transformers 可以通过 conda 依此安装：
+
+```shell script
+conda install -c huggingface transformers
+```
+
+要通过 conda 安装 Flax、PyTorch 或 TensorFlow 其中之一，请参阅它们各自安装页的说明。
+
+## 模型架构
+
+**🤗 Transformers 支持的[所有的模型检查点](https://huggingface.co/models)** 由[用户](https://huggingface.co/users)和[组织](https://huggingface.co/organizations)上传，均与 huggingface.co [model hub](https://huggingface.co) 无缝整合。
+
+目前的检查点数量： ![](https://img.shields.io/endpoint?url=https://huggingface.co/api/shields/models&color=brightgreen)
+
+🤗 Transformers 目前支持如下的架构（模型概述请阅[这里](https://huggingface.co/transformers/model_summary.html)）：
+
+1. **[ALBERT](https://huggingface.co/transformers/model_doc/albert.html)** (来自 Google Research and the Toyota Technological Institute at Chicago) 伴随论文 [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), 由 Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut 发布。
+1. **[BART](https://huggingface.co/transformers/model_doc/bart.html)** (来自 Facebook) 伴随论文 [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) 由 Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer 发布。
+1. **[BARThez](https://huggingface.co/transformers/model_doc/barthez.html)** (来自 École polytechnique) 伴随论文 [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) 由 Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis 发布。
+1. **[BERT](https://huggingface.co/transformers/model_doc/bert.html)** (来自 Google) 伴随论文 [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) 由 Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova 发布。
+1. **[BERT For Sequence Generation](https://huggingface.co/transformers/model_doc/bertgeneration.html)** (来自 Google) 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. **[BigBird-RoBERTa](https://huggingface.co/transformers/model_doc/bigbird.html)** (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. **[BigBird-Pegasus](https://huggingface.co/transformers/model_doc/bigbird_pegasus.html)** (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/transformers/model_doc/blenderbot.html)** (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/transformers/model_doc/blenderbot_small.html)** (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. **[BORT](https://huggingface.co/transformers/model_doc/bort.html)** (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/transformers/model_doc/byt5.html)** (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/transformers/model_doc/camembert.html)** (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/transformers/model_doc/canine.html)** (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/transformers/model_doc/clip.html)** (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. **[ConvBERT](https://huggingface.co/transformers/model_doc/convbert.html)** (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. **[CPM](https://huggingface.co/transformers/model_doc/cpm.html)** (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/transformers/model_doc/ctrl.html)** (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. **[DeBERTa](https://huggingface.co/transformers/model_doc/deberta.html)** (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/transformers/model_doc/deberta_v2.html)** (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. **[DeiT](https://huggingface.co/transformers/model_doc/deit.html)** (来自 Facebook) 伴随论文 [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) 由 Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou 发布。
+1. **[DETR](https://huggingface.co/transformers/model_doc/detr.html)** (来自 Facebook) 伴随论文 [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) 由 Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko 发布。
+1. **[DialoGPT](https://huggingface.co/transformers/model_doc/dialogpt.html)** (来自 Microsoft Research) 伴随论文 [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) 由 Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan 发布。
+1. **[DistilBERT](https://huggingface.co/transformers/model_doc/distilbert.html)** (来自 HuggingFace), 伴随论文 [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) 由 Victor Sanh, Lysandre Debut and Thomas Wolf 发布。 同样的方法也应用于压缩 GPT-2 到 [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa 到 [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT 到 [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) 和德语版 DistilBERT。
+1. **[DPR](https://huggingface.co/transformers/model_doc/dpr.html)** (来自 Facebook) 伴随论文 [Dense Passage Retrieval
+for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) 由 Vladimir Karpukhin, Barlas Oğuz, Sewon
+Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih 发布。
+1. **[ELECTRA](https://huggingface.co/transformers/model_doc/electra.html)** (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. **[FlauBERT](https://huggingface.co/transformers/model_doc/flaubert.html)** (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. **[Funnel Transformer](https://huggingface.co/transformers/model_doc/funnel.html)** (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. **[GPT](https://huggingface.co/transformers/model_doc/gpt.html)** (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-2](https://huggingface.co/transformers/model_doc/gpt2.html)** (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 Neo](https://huggingface.co/transformers/model_doc/gpt_neo.html)** (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. **[Hubert](https://huggingface.co/transformers/model_doc/hubert.html)** (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/transformers/model_doc/ibert.html)** (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. **[LayoutLM](https://huggingface.co/transformers/model_doc/layoutlm.html)** (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. **[LED](https://huggingface.co/transformers/model_doc/led.html)** (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. **[Longformer](https://huggingface.co/transformers/model_doc/longformer.html)** (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. **[LUKE](https://huggingface.co/transformers/model_doc/luke.html)** (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/transformers/model_doc/lxmert.html)** (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. **[M2M100](https://huggingface.co/transformers/model_doc/m2m_100.html)** (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/transformers/model_doc/marian.html)** 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. **[MBart](https://huggingface.co/transformers/model_doc/mbart.html)** (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/transformers/model_doc/mbart.html)** (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/transformers/model_doc/megatron_bert.html)** (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/transformers/model_doc/megatron_gpt2.html)** (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. **[MPNet](https://huggingface.co/transformers/model_doc/mpnet.html)** (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/transformers/model_doc/mt5.html)** (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. **[Pegasus](https://huggingface.co/transformers/model_doc/pegasus.html)** (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. **[ProphetNet](https://huggingface.co/transformers/model_doc/prophetnet.html)** (来自 Microsoft Research) 伴随论文 [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) 由 Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou 发布。
+1. **[Reformer](https://huggingface.co/transformers/model_doc/reformer.html)** (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. **[RoBERTa](https://huggingface.co/transformers/model_doc/roberta.html)** (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。
+1. **[RoFormer](https://huggingface.co/transformers/model_doc/roformer.html)** (来自 ZhuiyiTechnology), 伴随论文 [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) 由 Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu 发布。
+1. **[SpeechToTextTransformer](https://huggingface.co/transformers/model_doc/speech_to_text.html)** (来自 Facebook), 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, Dmytro Okhonko, Juan Pino 发布。
+1. **[SqueezeBert](https://huggingface.co/transformers/model_doc/squeezebert.html)** 伴随论文 [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://arxiv.org/abs/2006.11316) 由 Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer 发布。
+1. **[T5](https://huggingface.co/transformers/model_doc/t5.html)** (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. **[TAPAS](https://huggingface.co/transformers/model_doc/tapas.html)** (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. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (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. **[Vision Transformer (ViT)](https://huggingface.co/transformers/model_doc/vit.html)** (来自 Google AI) 伴随论文 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 由 Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 发布。
+1. **[VisualBERT](https://huggingface.co/transformers/model_doc/visual_bert.html)** (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. **[Wav2Vec2](https://huggingface.co/transformers/model_doc/wav2vec2.html)** (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. **[XLM](https://huggingface.co/transformers/model_doc/xlm.html)** (来自 Facebook) 伴随论文 [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) 由 Guillaume Lample and Alexis Conneau 发布。
+1. **[XLM-ProphetNet](https://huggingface.co/transformers/model_doc/xlmprophetnet.html)** (来自 Microsoft Research) 伴随论文 [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) 由 Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou 发布。
+1. **[XLM-RoBERTa](https://huggingface.co/transformers/model_doc/xlmroberta.html)** (来自 Facebook AI), 伴随论文 [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) 由 Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov 发布。
+1. **[XLNet](https://huggingface.co/transformers/model_doc/xlnet.html)** (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. **[XLSR-Wav2Vec2](https://huggingface.co/transformers/model_doc/xlsr_wav2vec2.html)** (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. 想要贡献新的模型？我们这里有一份**详细指引和模板**来引导你添加新的模型。你可以在 [`templates`](./templates) 目录中找到他们。记得查看 [贡献指南](./CONTRIBUTING.md) 并在开始写 PR 前联系维护人员或开一个新的 issue 来获得反馈。
+
+要检查某个模型是否已有 Flax、PyTorch 或 TensorFlow 的实现，或其是否在 🤗 Tokenizers 库中有对应词符化器（tokenizer），敬请参阅[此表](https://huggingface.co/transformers/index.html#supported-frameworks)。
+
+这些实现均已于多个数据集测试（请参看用例脚本）并应于原版实现表现相当。你可以在用例文档的[此节](https://huggingface.co/transformers/examples.html)中了解表现的细节。
+
+
+## 了解更多
+
+| 章节 | 描述 |
+|-|-|
+| [文档](https://huggingface.co/transformers/) | 完整的 API 文档和教程 |
+| [任务总结](https://huggingface.co/transformers/task_summary.html) | 🤗 Transformers 支持的任务 |
+| [预处理教程](https://huggingface.co/transformers/preprocessing.html) | 使用 `Tokenizer` 来为模型准备数据 |
+| [训练和微调](https://huggingface.co/transformers/training.html) | 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/master/examples) | 为各种任务提供的用例脚本 |
+| [模型分享和上传](https://huggingface.co/transformers/model_sharing.html) | 和社区上传和分享你微调的模型 |
+| [迁移](https://huggingface.co/transformers/migration.html) | 从 `pytorch-transformers` 或 `pytorch-pretrained-bert` 迁移到 🤗 Transformers |
+
+## 引用
+
+我们已将此库的[论文](https://www.aclweb.org/anthology/2020.emnlp-demos.6/)正式发表，如果你使用了 🤗 Transformers 库，请引用:
+```bibtex
+@inproceedings{wolf-etal-2020-transformers,
+    title = "Transformers: State-of-the-Art Natural Language Processing",
+    author = "Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and Rémi Louf and Morgan Funtowicz and Joe Davison and Sam Shleifer and Patrick von Platen and Clara Ma and Yacine Jernite and Julien Plu and Canwen Xu and Teven Le Scao and Sylvain Gugger and Mariama Drame and Quentin Lhoest and Alexander M. Rush",
+    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
+    month = oct,
+    year = "2020",
+    address = "Online",
+    publisher = "Association for Computational Linguistics",
+    url = "https://www.aclweb.org/anthology/2020.emnlp-demos.6",
+    pages = "38--45"
+}
+```

README_zh-hant.md

@@ -0,0 +1,346 @@
+<!---
+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.
+-->
+
+<!---
+A useful guide for English-Traditional Chinese translation of Hugging Face documentation
+- Add space around English words and numbers when they appear between Chinese characters. E.g., 共 100 多種語言; 使用 transformers 函式庫。
+- Use square quotes, e.g.,「引用」
+- Some of terms in the file can be found at National Academy for Educational Research (https://terms.naer.edu.tw/), an official website providing bilingual translations between English and Traditional Chinese.
+
+Dictionary
+
+API: API (不翻譯）
+add: 加入
+checkpoint: 檢查點
+code: 程式碼
+community: 社群
+confidence: 信賴度
+dataset: 資料集
+documentation: 文件
+example: 基本翻譯為「範例」，或依語意翻為「例子」
+finetune: 微調
+Hugging Face: Hugging Face（不翻譯）
+implementation: 實作
+inference: 推論
+library: 函式庫
+module: 模組
+NLP/Natural Language Processing: 以 NLP 出現時不翻譯，以 Natural Language Processing 出現時翻譯為自然語言處理
+online demos: 線上Demo
+pipeline: pipeline（不翻譯）
+pretrained/pretrain: 預訓練
+Python data structures (e.g., list, set, dict): 翻譯為串列，集合，字典，並用括號標註原英文
+repository: repository（不翻譯）
+summary: 概覽
+token-: token-（不翻譯）
+Trainer: Trainer（不翻譯）
+transformer: transformer（不翻譯）
+tutorial: 教學
+user: 使用者
+-->
+
+<p align="center">
+    <br>
+    <img src="https://raw.githubusercontent.com/huggingface/transformers/master/docs/source/imgs/transformers_logo_name.png" width="400"/>
+    <br>
+<p>
+<p align="center">
+    <a href="https://circleci.com/gh/huggingface/transformers">
+        <img alt="Build" src="https://img.shields.io/circleci/build/github/huggingface/transformers/master">
+    </a>
+    <a href="https://github.com/huggingface/transformers/blob/master/LICENSE">
+        <img alt="GitHub" src="https://img.shields.io/github/license/huggingface/transformers.svg?color=blue">
+    </a>
+    <a href="https://huggingface.co/transformers/index.html">
+        <img alt="Documentation" src="https://img.shields.io/website/http/huggingface.co/transformers/index.html.svg?down_color=red&down_message=offline&up_message=online">
+    </a>
+    <a href="https://github.com/huggingface/transformers/releases">
+        <img alt="GitHub release" src="https://img.shields.io/github/release/huggingface/transformers.svg">
+    </a>
+    <a href="https://github.com/huggingface/transformers/blob/master/CODE_OF_CONDUCT.md">
+        <img alt="Contributor Covenant" src="https://img.shields.io/badge/Contributor%20Covenant-v2.0%20adopted-ff69b4.svg">
+    </a>
+    <a href="https://zenodo.org/badge/latestdoi/155220641"><img src="https://zenodo.org/badge/155220641.svg" alt="DOI"></a>
+</p>
+
+<h4 align="center">
+    <p>
+        <a href="https://github.com/huggingface/transformers/">English</a> |
+        <a href="https://github.com/huggingface/transformers/blob/master/README_zh-hans.md">简体中文</a> |
+        <b>繁體中文</b>
+    <p>
+</h4>
+
+<h3 align="center">
+    <p>為 Jax、PyTorch 以及 TensorFlow 打造的先進自然語言處理函式庫</p>
+</h3>
+
+<h3 align="center">
+    <a href="https://hf.co/course"><img src="https://raw.githubusercontent.com/huggingface/transformers/master/docs/source/imgs/course_banner.png"></a>
+</h3>
+
+🤗 Transformers 提供了數以千計的預訓練模型，支援 100 多種語言的文本分類、資訊擷取、問答、摘要、翻譯、文本生成。它的宗旨是讓最先進的 NLP 技術人人易用。
+
+🤗 Transformers 提供了便於快速下載和使用的API，讓你可以將預訓練模型用在給定文本、在你的資料集上微調然後經由 [model hub](https://huggingface.co/models) 與社群共享。同時，每個定義的 Python 模組架構均完全獨立，方便修改和快速研究實驗。
+
+🤗 Transformers 支援三個最熱門的深度學習函式庫： [Jax](https://jax.readthedocs.io/en/latest/), [PyTorch](https://pytorch.org/) 以及 [TensorFlow](https://www.tensorflow.org/) — 並與之完美整合。你可以直接使用其中一個框架訓練你的模型，然後用另一個載入和推論。
+
+## 線上Demo
+
+你可以直接在 [model hub](https://huggingface.co/models) 上測試大多數的模型。我們也提供了 [私有模型託管、模型版本管理以及推論API](https://huggingface.co/pricing)。
+
+這裡是一些範例：
+- [用 BERT 做遮蓋填詞](https://huggingface.co/bert-base-uncased?text=Paris+is+the+%5BMASK%5D+of+France)
+- [用 Electra 做專有名詞辨識](https://huggingface.co/dbmdz/electra-large-discriminator-finetuned-conll03-english?text=My+name+is+Sarah+and+I+live+in+London+city)
+- [用 GPT-2 做文本生成](https://huggingface.co/gpt2?text=A+long+time+ago%2C+)
+- [用 RoBERTa 做自然語言推論](https://huggingface.co/roberta-large-mnli?text=The+dog+was+lost.+Nobody+lost+any+animal)
+- [用 BART 做文本摘要](https://huggingface.co/facebook/bart-large-cnn?text=The+tower+is+324+metres+%281%2C063+ft%29+tall%2C+about+the+same+height+as+an+81-storey+building%2C+and+the+tallest+structure+in+Paris.+Its+base+is+square%2C+measuring+125+metres+%28410+ft%29+on+each+side.+During+its+construction%2C+the+Eiffel+Tower+surpassed+the+Washington+Monument+to+become+the+tallest+man-made+structure+in+the+world%2C+a+title+it+held+for+41+years+until+the+Chrysler+Building+in+New+York+City+was+finished+in+1930.+It+was+the+first+structure+to+reach+a+height+of+300+metres.+Due+to+the+addition+of+a+broadcasting+aerial+at+the+top+of+the+tower+in+1957%2C+it+is+now+taller+than+the+Chrysler+Building+by+5.2+metres+%2817+ft%29.+Excluding+transmitters%2C+the+Eiffel+Tower+is+the+second+tallest+free-standing+structure+in+France+after+the+Millau+Viaduct)
+- [用 DistilBERT 做問答](https://huggingface.co/distilbert-base-uncased-distilled-squad?text=Which+name+is+also+used+to+describe+the+Amazon+rainforest+in+English%3F&context=The+Amazon+rainforest+%28Portuguese%3A+Floresta+Amaz%C3%B4nica+or+Amaz%C3%B4nia%3B+Spanish%3A+Selva+Amaz%C3%B3nica%2C+Amazon%C3%ADa+or+usually+Amazonia%3B+French%3A+For%C3%AAt+amazonienne%3B+Dutch%3A+Amazoneregenwoud%29%2C+also+known+in+English+as+Amazonia+or+the+Amazon+Jungle%2C+is+a+moist+broadleaf+forest+that+covers+most+of+the+Amazon+basin+of+South+America.+This+basin+encompasses+7%2C000%2C000+square+kilometres+%282%2C700%2C000+sq+mi%29%2C+of+which+5%2C500%2C000+square+kilometres+%282%2C100%2C000+sq+mi%29+are+covered+by+the+rainforest.+This+region+includes+territory+belonging+to+nine+nations.+The+majority+of+the+forest+is+contained+within+Brazil%2C+with+60%25+of+the+rainforest%2C+followed+by+Peru+with+13%25%2C+Colombia+with+10%25%2C+and+with+minor+amounts+in+Venezuela%2C+Ecuador%2C+Bolivia%2C+Guyana%2C+Suriname+and+French+Guiana.+States+or+departments+in+four+nations+contain+%22Amazonas%22+in+their+names.+The+Amazon+represents+over+half+of+the+planet%27s+remaining+rainforests%2C+and+comprises+the+largest+and+most+biodiverse+tract+of+tropical+rainforest+in+the+world%2C+with+an+estimated+390+billion+individual+trees+divided+into+16%2C000+species)
+- [用 T5 做翻譯](https://huggingface.co/t5-base?text=My+name+is+Wolfgang+and+I+live+in+Berlin)
+
+**[Write With Transformer](https://transformer.huggingface.co)**，由 Hugging Face 團隊所打造，是一個文本生成的官方 demo。
+
+## 如果你在尋找由 Hugging Face 團隊所提供的客製化支援服務
+
+<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);">
+</a><br>
+
+## 快速上手
+
+我們為快速使用模型提供了 `pipeline` API。 Pipeline 包含了預訓練模型和對應的文本預處理。下面是一個快速使用 pipeline 去判斷正負面情緒的例子：
+
+```python
+>>> from transformers import pipeline
+
+# 使用情緒分析 pipeline
+>>> classifier = pipeline('sentiment-analysis')
+>>> classifier('We are very happy to introduce pipeline to the transformers repository.')
+[{'label': 'POSITIVE', 'score': 0.9996980428695679}]
+```
+
+第二行程式碼下載並快取 pipeline 使用的預訓練模型，而第三行程式碼則在給定的文本上進行了評估。這裡的答案“正面” (positive) 具有 99.97% 的信賴度。
+
+許多的 NLP 任務都有隨選即用的預訓練 `pipeline`。例如，我們可以輕鬆地從給定文本中擷取問題答案：
+
+``` python
+>>> from transformers import pipeline
+
+# 使用問答 pipeline
+>>> 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'}
+
+```
+
+除了提供問題解答，預訓練模型還提供了對應的信賴度分數以及解答在 tokenized 後的文本中開始和結束的位置。你可以從[這個教學](https://huggingface.co/transformers/task_summary.html)了解更多 `pipeline` API支援的任務。
+
+要在你的任務中下載和使用任何預訓練模型很簡單，只需三行程式碼。這裡是 PyTorch 版的範例：
+```python
+>>> from transformers import AutoTokenizer, AutoModel
+
+>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+>>> model = AutoModel.from_pretrained("bert-base-uncased")
+
+>>> inputs = tokenizer("Hello world!", return_tensors="pt")
+>>> outputs = model(**inputs)
+```
+這裡是對應的 TensorFlow 程式碼：
+```python
+>>> from transformers import AutoTokenizer, TFAutoModel
+
+>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+>>> model = TFAutoModel.from_pretrained("bert-base-uncased")
+
+>>> inputs = tokenizer("Hello world!", return_tensors="tf")
+>>> outputs = model(**inputs)
+```
+
+Tokenizer 為所有的預訓練模型提供了預處理，並可以直接轉換單一字串（比如上面的例子）或串列 (list)。它會輸出一個的字典 (dict) 讓你可以在下游程式碼裡使用或直接藉由 `**` 運算式傳給模型。
+
+模型本身是一個常規的 [Pytorch `nn.Module`](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) 或 [TensorFlow `tf.keras.Model`](https://www.tensorflow.org/api_docs/python/tf/keras/Model)（取決於你的後端），可依常規方式使用。 [這個教學](https://huggingface.co/transformers/training.html)解釋了如何將這樣的模型整合到一般的 PyTorch 或 TensorFlow 訓練迴圈中，或是如何使用我們的 `Trainer` API 在一個新的資料集上快速進行微調。
+
+## 為什麼要用 transformers？
+
+1. 便於使用的先進模型：
+    - NLU 和 NLG 上性能卓越
+    - 對教學和實作友好且低門檻
+    - 高度抽象，使用者只須學習 3 個類別
+    - 對所有模型使用的制式化API
+
+1. 更低的運算成本，更少的碳排放：
+    - 研究人員可以分享預訓練的模型而非從頭開始訓練
+    - 工程師可以減少計算時間以及生產成本
+    - 數十種模型架構、兩千多個預訓練模型、100多種語言支援
+
+1. 對於模型生命週期的每一個部分都面面俱到：
+    - 訓練先進的模型，只需 3 行程式碼
+    - 模型可以在不同深度學習框架之間任意轉換
+    - 為訓練、評估和生產選擇最適合的框架，並完美銜接
+
+1. 為你的需求輕鬆客製化專屬模型和範例：
+    - 我們為每種模型架構提供了多個範例來重現原論文結果
+    - 一致的模型內部架構
+    - 模型檔案可單獨使用，便於修改和快速實驗
+
+## 什麼情況下我不該用 transformers？
+
+- 本函式庫並不是模組化的神經網絡工具箱。模型文件中的程式碼並未做額外的抽象封裝，以便研究人員快速地翻閱及修改程式碼，而不會深陷複雜的類別包裝之中。
+- `Trainer` API 並非相容任何模型，它只為本函式庫中的模型最佳化。對於一般的機器學習用途，請使用其他函式庫。
+- 儘管我們已盡力而為，[examples 目錄](https://github.com/huggingface/transformers/tree/master/examples)中的腳本也僅為範例而已。對於特定問題，它們並不一定隨選即用，可能需要修改幾行程式碼以符合需求。
+
+## 安裝
+
+### 使用 pip
+
+這個 Repository 已在 Python 3.6+、Flax 0.3.2+、PyTorch 1.3.1+ 和 TensorFlow 2.3+ 下經過測試。
+
+你可以在[虛擬環境](https://docs.python.org/3/library/venv.html)中安裝 🤗 Transformers。如果你還不熟悉 Python 的虛擬環境，請閱此[使用者指引](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/)。
+
+首先，用你打算使用的版本的 Python 創建一個虛擬環境並進入。
+
+然後，你需要安裝 Flax、PyTorch 或 TensorFlow 其中之一。對於該如何在你使用的平台上安裝這些框架，請參閱 [TensorFlow 安裝頁面](https://www.tensorflow.org/install/), [PyTorch 安裝頁面](https://pytorch.org/get-started/locally/#start-locally) 或 [Flax 安裝頁面](https://github.com/google/flax#quick-install)。
+
+當其中一個後端安裝成功後，🤗 Transformers 可依此安裝：
+
+```bash
+pip install transformers
+```
+
+如果你想要試試範例或者想在正式發布前使用最新開發中的程式碼，你必須[從原始碼安裝](https://huggingface.co/transformers/installation.html#installing-from-source)。
+
+### 使用 conda
+
+自 Transformers 4.0.0 版始，我們有了一個 conda channel： `huggingface`。
+
+🤗 Transformers 可以藉由 conda 依此安裝：
+
+```shell script
+conda install -c huggingface transformers
+```
+
+要藉由 conda 安裝 Flax、PyTorch 或 TensorFlow 其中之一，請參閱它們各自安裝頁面的說明。
+
+## 模型架構
+
+**🤗 Transformers 支援的[所有的模型檢查點](https://huggingface.co/models)**，由[使用者](https://huggingface.co/users)和[組織](https://huggingface.co/organizations)上傳，均與 huggingface.co [model hub](https://huggingface.co) 完美結合。
+
+目前的檢查點數量： ![](https://img.shields.io/endpoint?url=https://huggingface.co/api/shields/models&color=brightgreen)
+
+🤗 Transformers 目前支援以下的架構（模型概覽請參閱[這裡](https://huggingface.co/transformers/model_summary.html)）：
+
+1. **[ALBERT](https://huggingface.co/transformers/model_doc/albert.html)** (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](https://huggingface.co/transformers/model_doc/bart.html)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
+1. **[BARThez](https://huggingface.co/transformers/model_doc/barthez.html)** (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. **[BERT](https://huggingface.co/transformers/model_doc/bert.html)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
+1. **[BERT For Sequence Generation](https://huggingface.co/transformers/model_doc/bertgeneration.html)** (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](https://huggingface.co/transformers/model_doc/bigbird.html)** (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](https://huggingface.co/transformers/model_doc/bigbird_pegasus.html)** (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/transformers/model_doc/blenderbot.html)** (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/transformers/model_doc/blenderbot_small.html)** (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](https://huggingface.co/transformers/model_doc/bort.html)** (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/transformers/model_doc/byt5.html)** (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/transformers/model_doc/camembert.html)** (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/transformers/model_doc/canine.html)** (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/transformers/model_doc/clip.html)** (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](https://huggingface.co/transformers/model_doc/convbert.html)** (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](https://huggingface.co/transformers/model_doc/cpm.html)** (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/transformers/model_doc/ctrl.html)** (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. **[DeBERTa](https://huggingface.co/transformers/model_doc/deberta.html)** (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/transformers/model_doc/deberta_v2.html)** (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. **[DeiT](https://huggingface.co/transformers/model_doc/deit.html)** (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](https://huggingface.co/transformers/model_doc/detr.html)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
+1. **[DialoGPT](https://huggingface.co/transformers/model_doc/dialogpt.html)** (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](https://huggingface.co/transformers/model_doc/distilbert.html)** (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/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
+1. **[DPR](https://huggingface.co/transformers/model_doc/dpr.html)** (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. **[ELECTRA](https://huggingface.co/transformers/model_doc/electra.html)** (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](https://huggingface.co/transformers/model_doc/flaubert.html)** (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. **[Funnel Transformer](https://huggingface.co/transformers/model_doc/funnel.html)** (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. **[GPT](https://huggingface.co/transformers/model_doc/gpt.html)** (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](https://huggingface.co/transformers/model_doc/gpt2.html)** (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 Neo](https://huggingface.co/transformers/model_doc/gpt_neo.html)** (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](https://huggingface.co/transformers/model_doc/hubert.html)** (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/transformers/model_doc/ibert.html)** (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. **[LayoutLM](https://huggingface.co/transformers/model_doc/layoutlm.html)** (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. **[LED](https://huggingface.co/transformers/model_doc/led.html)** (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](https://huggingface.co/transformers/model_doc/longformer.html)** (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](https://huggingface.co/transformers/model_doc/luke.html)** (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/transformers/model_doc/lxmert.html)** (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](https://huggingface.co/transformers/model_doc/m2m_100.html)** (from Facebook) released with the paper [Beyond English-Centric Multilingual Machine Translation](https://arxiv.org/abs/2010.11125) by 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/transformers/model_doc/marian.html)** 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. **[MBart](https://huggingface.co/transformers/model_doc/mbart.html)** (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/transformers/model_doc/mbart.html)** (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/transformers/model_doc/megatron_bert.html)** (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/transformers/model_doc/megatron_gpt2.html)** (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](https://huggingface.co/transformers/model_doc/mpnet.html)** (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/transformers/model_doc/mt5.html)** (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. **[Pegasus](https://huggingface.co/transformers/model_doc/pegasus.html)** (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. **[ProphetNet](https://huggingface.co/transformers/model_doc/prophetnet.html)** (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. **[Reformer](https://huggingface.co/transformers/model_doc/reformer.html)** (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. **[RoBERTa](https://huggingface.co/transformers/model_doc/roberta.html)** (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.
+1. **[RoFormer](https://huggingface.co/transformers/model_doc/roformer.html)** (from ZhuiyiTechnology), released together with the paper a [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
+1. **[SpeechToTextTransformer](https://huggingface.co/transformers/model_doc/speech_to_text.html)** (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. **[SqueezeBert](https://huggingface.co/transformers/model_doc/squeezebert.html)** 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. **[T5](https://huggingface.co/transformers/model_doc/t5.html)** (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. **[TAPAS](https://huggingface.co/transformers/model_doc/tapas.html)** (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. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (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. **[Vision Transformer (ViT)](https://huggingface.co/transformers/model_doc/vit.html)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VisualBERT](https://huggingface.co/transformers/model_doc/visual_bert.html)** (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. **[Wav2Vec2](https://huggingface.co/transformers/model_doc/wav2vec2.html)** (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. **[XLM](https://huggingface.co/transformers/model_doc/xlm.html)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
+1. **[XLM-ProphetNet](https://huggingface.co/transformers/model_doc/xlmprophetnet.html)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
+1. **[XLM-RoBERTa](https://huggingface.co/transformers/model_doc/xlmroberta.html)** (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. **[XLNet](https://huggingface.co/transformers/model_doc/xlnet.html)** (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](https://huggingface.co/transformers/model_doc/xlsr_wav2vec2.html)** (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. 想要貢獻新的模型？我們這裡有一份**詳細指引和模板**來引導你加入新的模型。你可以在 [`templates`](./templates) 目錄中找到它們。記得查看[貢獻指引](./CONTRIBUTING.md)並在開始寫 PR 前聯繫維護人員或開一個新的 issue 來獲得 feedbacks。
+
+要檢查某個模型是否已有 Flax、PyTorch 或 TensorFlow 的實作，或其是否在🤗 Tokenizers 函式庫中有對應的 tokenizer，敬請參閱[此表](https://huggingface.co/transformers/index.html#supported-frameworks)。
+
+這些實作均已於多個資料集測試（請參閱範例腳本）並應與原版實作表現相當。你可以在範例文件的[此節](https://huggingface.co/transformers/examples.html)中了解實作的細節。
+
+
+## 了解更多
+
+| 章節 | 描述 |
+|-|-|
+| [文件](https://huggingface.co/transformers/) | 完整的 API 文件和教學 |
+| [任務概覽](https://huggingface.co/transformers/task_summary.html) | 🤗 Transformers 支援的任務 |
+| [預處理教學](https://huggingface.co/transformers/preprocessing.html) | 使用 `Tokenizer` 來為模型準備資料 |
+| [訓練和微調](https://huggingface.co/transformers/training.html) | 使用 PyTorch/TensorFlow 的內建的訓練方式或於 `Trainer` API 中使用 🤗 Transformers 提供的模型 |
+| [快速上手：微調和範例腳本](https://github.com/huggingface/transformers/tree/master/examples) | 為各種任務提供的範例腳本 |
+| [模型分享和上傳](https://huggingface.co/transformers/model_sharing.html) | 上傳並與社群分享你微調的模型 |
+| [遷移](https://huggingface.co/transformers/migration.html) | 從 `pytorch-transformers` 或 `pytorch-pretrained-bert` 遷移到 🤗 Transformers |
+
+## 引用
+
+我們已將此函式庫的[論文](https://www.aclweb.org/anthology/2020.emnlp-demos.6/)正式發表。如果你使用了 🤗 Transformers 函式庫，可以引用：
+```bibtex
+@inproceedings{wolf-etal-2020-transformers,
+    title = "Transformers: State-of-the-Art Natural Language Processing",
+    author = "Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and Rémi Louf and Morgan Funtowicz and Joe Davison and Sam Shleifer and Patrick von Platen and Clara Ma and Yacine Jernite and Julien Plu and Canwen Xu and Teven Le Scao and Sylvain Gugger and Mariama Drame and Quentin Lhoest and Alexander M. Rush",
+    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
+    month = oct,
+    year = "2020",
+    address = "Online",
+    publisher = "Association for Computational Linguistics",
+    url = "https://www.aclweb.org/anthology/2020.emnlp-demos.6",
+    pages = "38--45"
+}
+```

docs/source/_static/js/custom.js

@@ -1,10 +1,14 @@
 // These two things need to be updated at each release for the version selector.
 // Last stable version
-const stableVersion = "v4.5.1"
+const stableVersion = "v4.9.0"
 // Dictionary doc folder to label. The last stable version should have an empty key.
 const versionMapping = {
     "master": "master",
-    "": "v4.5.0/v4.5.1 (stable)",
+    "": "v4.9.0 (stable)",
+    "v4.8.2": "v4.8.0/v4.8.1/v4.8.2",
+    "v4.7.0": "v4.7.0",
+    "v4.6.0": "v4.6.0",
+    "v4.5.1": "v4.5.0/v4.5.1",
     "v4.4.2": "v4.4.0/v4.4.1/v4.4.2",
     "v4.3.3": "v4.3.0/v4.3.1/v4.3.2/v4.3.3",
     "v4.2.2": "v4.2.0/v4.2.1/v4.2.2",

docs/source/add_new_model.rst

@@ -518,7 +518,7 @@ PyTorch, called ``SimpleModel`` as follows:
 
 .. code:: python
 
-   import torch.nn as nn
+   from torch import nn
 
    class SimpleModel(nn.Module):
        def __init__(self):

docs/source/benchmarks.rst

@@ -358,4 +358,6 @@ available `here
 <https://docs.google.com/spreadsheets/d/1sryqufw2D0XlUH4sq3e9Wnxu5EAQkaohzrJbd5HdQ_w/edit?usp=sharing>`__.
 
 With the new `benchmark` tools, it is easier than ever to share your benchmark results with the community
-:prefix_link:`here <examples/benchmarking/README.md>`.
+
+- :prefix_link:`PyTorch Benchmarking Results<examples/pytorch/benchmarking/README.md>`.
+- :prefix_link:`TensorFlow Benchmarking Results<examples/tensorflow/benchmarking/README.md>`.

docs/source/community.md

@@ -52,7 +52,12 @@ This page regroups resources around 🤗 Transformers developed by the community
 |[Fine-tune BART for summarization in two languages with Trainer class](https://github.com/elsanns/xai-nlp-notebooks/blob/master/fine_tune_bart_summarization_two_langs.ipynb) | How to fine-tune BART for summarization in two languages with Trainer class | [Eliza Szczechla](https://github.com/elsanns) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/elsanns/xai-nlp-notebooks/blob/master/fine_tune_bart_summarization_two_langs.ipynb)|
 |[Evaluate Big Bird on Trivia QA](https://github.com/patrickvonplaten/notebooks/blob/master/Evaluating_Big_Bird_on_TriviaQA.ipynb) | How to evaluate BigBird on long document question answering on Trivia QA | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/Evaluating_Big_Bird_on_TriviaQA.ipynb)|
 | [Create video captions using Wav2Vec2](https://github.com/Muennighoff/ytclipcc/blob/main/wav2vec_youtube_captions.ipynb) | How to create YouTube captions from any video by transcribing the audio with Wav2Vec | [Niklas Muennighoff](https://github.com/Muennighoff) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/Muennighoff/ytclipcc/blob/main/wav2vec_youtube_captions.ipynb) |
+| [Fine-tune the Vision Transformer on CIFAR-10 using PyTorch Lightning](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_PyTorch_Lightning.ipynb) | How to fine-tune the Vision Transformer (ViT) on CIFAR-10 using HuggingFace Transformers, Datasets and PyTorch Lightning | [Niels Rogge](https://github.com/nielsrogge) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_PyTorch_Lightning.ipynb) |
+| [Fine-tune the Vision Transformer on CIFAR-10 using the 🤗 Trainer](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_the_%F0%9F%A4%97_Trainer.ipynb) | How to fine-tune the Vision Transformer (ViT) on CIFAR-10 using HuggingFace Transformers, Datasets and the 🤗 Trainer | [Niels Rogge](https://github.com/nielsrogge) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_the_%F0%9F%A4%97_Trainer.ipynb) |
 | [Evaluate LUKE on Open Entity, an entity typing dataset](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_open_entity.ipynb) | How to evaluate *LukeForEntityClassification* on the Open Entity dataset | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_open_entity.ipynb) |
 | [Evaluate LUKE on TACRED, a relation extraction dataset](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_tacred.ipynb) | How to evaluate *LukeForEntityPairClassification* on the TACRED dataset | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_tacred.ipynb) |
 | [Evaluate LUKE on CoNLL-2003, an important NER benchmark](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_conll_2003.ipynb) | How to evaluate *LukeForEntitySpanClassification* on the CoNLL-2003 dataset | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_conll_2003.ipynb) |
 | [Evaluate BigBird-Pegasus on PubMed dataset](https://github.com/vasudevgupta7/bigbird/blob/main/notebooks/bigbird_pegasus_evaluation.ipynb) | How to evaluate *BigBirdPegasusForConditionalGeneration* on PubMed dataset | [Vasudev Gupta](https://github.com/vasudevgupta7) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/vasudevgupta7/bigbird/blob/main/notebooks/bigbird_pegasus_evaluation.ipynb) |
+| [Speech Emotion Classification with Wav2Vec2](https://github/m3hrdadfi/soxan/blob/main/notebooks/Emotion_recognition_in_Greek_speech_using_Wav2Vec2.ipynb) | How to leverage a pretrained Wav2Vec2 model for Emotion Classification on the MEGA dataset | [Mehrdad Farahani](https://github.com/m3hrdadfi) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/m3hrdadfi/soxan/blob/main/notebooks/Emotion_recognition_in_Greek_speech_using_Wav2Vec2.ipynb) |
+| [Detect objects in an image with DETR](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/DETR/DETR_minimal_example_(with_DetrFeatureExtractor).ipynb) | How to use a trained *DetrForObjectDetection* model to detect objects in an image and visualize attention | [Niels Rogge](https://github.com/NielsRogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/DETR/DETR_minimal_example_(with_DetrFeatureExtractor).ipynb) |
+| [Fine-tune DETR on a custom object detection dataset](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/DETR/Fine_tuning_DetrForObjectDetection_on_custom_dataset_(balloon).ipynb) | How to fine-tune *DetrForObjectDetection* on a custom object detection dataset | [Niels Rogge](https://github.com/NielsRogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/DETR/Fine_tuning_DetrForObjectDetection_on_custom_dataset_(balloon).ipynb) |

docs/source/conf.py

@@ -27,7 +27,8 @@ author = "huggingface"
 # The short X.Y version
 version = ""
 # The full version, including alpha/beta/rc tags
-release = "4.5.0.dev0"
+release = u'4.7.0'
+
 
 
 # Prefix link to point to master, comment this during version release and uncomment below line

docs/source/debugging.rst

@@ -24,7 +24,11 @@ Underflow and Overflow Detection
 
 .. note::
 
-   This feature can be used with any ``nn.Module``-based model
+   For multi-GPU training it requires DDP (``torch.distributed.launch``).
+
+.. note::
+
+   This feature can be used with any ``nn.Module``-based model.
 
 If you start getting ``loss=NaN`` or the model inhibits some other abnormal behavior due to ``inf`` or ``nan`` in
 activations or weights one needs to discover where the first underflow or overflow happens and what led to it. Luckily

docs/source/glossary.rst

@@ -55,6 +55,12 @@ Input IDs
 The input ids are often the only required parameters to be passed to the model as input. *They are token indices,
 numerical representations of tokens building the sequences that will be used as input by the model*.
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/VFp38yj8h3A" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 Each tokenizer works differently but the underlying mechanism remains the same. Here's an example using the BERT
 tokenizer, which is a `WordPiece <https://arxiv.org/pdf/1609.08144.pdf>`__ tokenizer:
 
@@ -120,8 +126,15 @@ because this is the way a :class:`~transformers.BertModel` is going to expect it
 Attention mask
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-The attention mask is an optional argument used when batching sequences together. This argument indicates to the model
-which tokens should be attended to, and which should not.
+The attention mask is an optional argument used when batching sequences together.
+
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/M6adb1j2jPI" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
+This argument indicates to the model which tokens should be attended to, and which should not.
 
 For example, consider these two sequences:
 
@@ -175,10 +188,17 @@ in the dictionary returned by the tokenizer under the key "attention_mask":
 Token Type IDs
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-Some models' purpose is to do sequence classification or question answering. These require two different sequences to
-be joined in a single "input_ids" entry, which usually is performed with the help of special tokens, such as the
-classifier (``[CLS]``) and separator (``[SEP]``) tokens. For example, the BERT model builds its two sequence input as
-such:
+Some models' purpose is to do classification on pairs of sentences or question answering.
+
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/0u3ioSwev3s" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
+These require two different sequences to be joined in a single "input_ids" entry, which usually is performed with the
+help of special tokens, such as the classifier (``[CLS]``) and separator (``[SEP]``) tokens. For example, the BERT
+model builds its two sequence input as such:
 
 .. code-block::
 

docs/source/index.rst

@@ -8,7 +8,18 @@ architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet...) for Natural Lang
 Language Generation (NLG) with over 32+ pretrained models in 100+ languages and deep interoperability between Jax,
 PyTorch and TensorFlow.
 
-This is the documentation of our repository `transformers <https://github.com/huggingface/transformers>`_.
+This is the documentation of our repository `transformers <https://github.com/huggingface/transformers>`__. You can
+also follow our `online course <https://huggingface.co/course>`__ that teaches how to use this library, as well as the
+other libraries developed by Hugging Face and the Hub.
+
+If you are looking for custom support from the Hugging Face team
+-----------------------------------------------------------------------------------------------------------------------
+
+.. raw:: html
+
+    <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);">
+    </a><br>
 
 Features
 -----------------------------------------------------------------------------------------------------------------------
@@ -75,7 +86,10 @@ The documentation is organized in five parts:
     - **INTERNAL HELPERS** for the classes and functions we use internally.
 
 The library currently contains Jax, PyTorch and Tensorflow implementations, pretrained model weights, usage scripts and
-conversion utilities for the following models:
+conversion utilities for the following models.
+
+Supported models
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 ..
     This list is updated automatically from the README with `make fix-copies`. Do not update manually!
@@ -111,154 +125,173 @@ conversion utilities for the following models:
    Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 10. :doc:`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.
-11. :doc:`CamemBERT <model_doc/camembert>` (from Inria/Facebook/Sorbonne) released with the paper `CamemBERT: a Tasty
+11. :doc:`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.
+12. :doc:`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.
-12. :doc:`CLIP <model_doc/clip>` from (OpenAI) released with the paper `Learning Transferable Visual Models From
+13. :doc:`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.
+14. :doc:`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.
-13. :doc:`ConvBERT <model_doc/convbert>` (from YituTech) released with the paper `ConvBERT: Improving BERT with
+15. :doc:`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.
-14. :doc:`CPM <model_doc/cpm>` (from Tsinghua University) released with the paper `CPM: A Large-scale Generative
+16. :doc:`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.
-15. :doc:`CTRL <model_doc/ctrl>` (from Salesforce) released with the paper `CTRL: A Conditional Transformer Language
+17. :doc:`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.
-16. :doc:`DeBERTa <model_doc/deberta>` (from Microsoft) released with the paper `DeBERTa: Decoding-enhanced BERT with
+18. :doc:`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.
-17. :doc:`DeBERTa-v2 <model_doc/deberta_v2>` (from Microsoft) released with the paper `DeBERTa: Decoding-enhanced BERT
+19. :doc:`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.
-18. :doc:`DeiT <model_doc/deit>` (from Facebook) released with the paper `Training data-efficient image transformers &
+20. :doc:`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.
-19. :doc:`DialoGPT <model_doc/dialogpt>` (from Microsoft Research) released with the paper `DialoGPT: Large-Scale
+21. :doc:`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.
+22. :doc:`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.
-20. :doc:`DistilBERT <model_doc/distilbert>` (from HuggingFace), released together with the paper `DistilBERT, a
+23. :doc:`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/master/examples/distillation>`__, RoBERTa into `DistilRoBERTa
     <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__, Multilingual BERT into
     `DistilmBERT <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__ and a German
     version of DistilBERT.
-21. :doc:`DPR <model_doc/dpr>` (from Facebook) released with the paper `Dense Passage Retrieval for Open-Domain
+24. :doc:`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.
-22. :doc:`ELECTRA <model_doc/electra>` (from Google Research/Stanford University) released with the paper `ELECTRA:
+25. :doc:`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.
-23. :doc:`FlauBERT <model_doc/flaubert>` (from CNRS) released with the paper `FlauBERT: Unsupervised Language Model
+26. :doc:`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.
-24. :doc:`Funnel Transformer <model_doc/funnel>` (from CMU/Google Brain) released with the paper `Funnel-Transformer:
+27. :doc:`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.
-25. :doc:`GPT <model_doc/gpt>` (from OpenAI) released with the paper `Improving Language Understanding by Generative
+28. :doc:`GPT <model_doc/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.
-26. :doc:`GPT-2 <model_doc/gpt2>` (from OpenAI) released with the paper `Language Models are Unsupervised Multitask
+29. :doc:`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**.
-27. :doc:`GPT Neo <model_doc/gpt_neo>` (from EleutherAI) released in the repository `EleutherAI/gpt-neo
+30. :doc:`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.
-28. :doc:`I-BERT <model_doc/ibert>` (from Berkeley) released with the paper `I-BERT: Integer-only BERT Quantization
+31. :doc:`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.
+32. :doc:`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
-29. :doc:`LayoutLM <model_doc/layoutlm>` (from Microsoft Research Asia) released with the paper `LayoutLM: Pre-training
+33. :doc:`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.
-30. :doc:`LED <model_doc/led>` (from AllenAI) released with the paper `Longformer: The Long-Document Transformer
+34. :doc:`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.
-31. :doc:`Longformer <model_doc/longformer>` (from AllenAI) released with the paper `Longformer: The Long-Document
+35. :doc:`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.
-32. :doc:`LUKE <model_doc/luke>` (from Studio Ousia) released with the paper `LUKE: Deep Contextualized Entity
+36. :doc:`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.
-33. :doc:`LXMERT <model_doc/lxmert>` (from UNC Chapel Hill) released with the paper `LXMERT: Learning Cross-Modality
+37. :doc:`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.
-34. :doc:`M2M100 <model_doc/m2m_100>` (from Facebook) released with the paper `Beyond English-Centric Multilingual
+38. :doc:`M2M100 <model_doc/m2m_100>` (from Facebook) released with the paper `Beyond English-Centric Multilingual
     Machine Translation <https://arxiv.org/abs/2010.11125>`__ by 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.
-35. :doc:`MarianMT <model_doc/marian>` Machine translation models trained using `OPUS <http://opus.nlpl.eu/>`__ data by
+39. :doc:`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.
-36. :doc:`MBart <model_doc/mbart>` (from Facebook) released with the paper `Multilingual Denoising Pre-training for
+40. :doc:`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.
-37. :doc:`MBart-50 <model_doc/mbart>` (from Facebook) released with the paper `Multilingual Translation with Extensible
+41. :doc:`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.
-38. :doc:`Megatron-BERT <model_doc/megatron_bert>` (from NVIDIA) released with the paper `Megatron-LM: Training
+42. :doc:`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.
-39. :doc:`Megatron-GPT2 <model_doc/megatron_gpt2>` (from NVIDIA) released with the paper `Megatron-LM: Training
+43. :doc:`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.
-40. :doc:`MPNet <model_doc/mpnet>` (from Microsoft Research) released with the paper `MPNet: Masked and Permuted
+44. :doc:`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.
-41. :doc:`MT5 <model_doc/mt5>` (from Google AI) released with the paper `mT5: A massively multilingual pre-trained
+45. :doc:`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.
-42. :doc:`Pegasus <model_doc/pegasus>` (from Google) released with the paper `PEGASUS: Pre-training with Extracted
+46. :doc:`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.
-43. :doc:`ProphetNet <model_doc/prophetnet>` (from Microsoft Research) released with the paper `ProphetNet: Predicting
+47. :doc:`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.
-44. :doc:`Reformer <model_doc/reformer>` (from Google Research) released with the paper `Reformer: The Efficient
+48. :doc:`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.
-45. :doc:`RoBERTa <model_doc/roberta>` (from Facebook), released together with the paper a `Robustly Optimized BERT
+49. :doc:`RoBERTa <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.
-46. :doc:`SpeechToTextTransformer <model_doc/speech_to_text>` (from Facebook), released together with the paper
+50. :doc:`RoFormer <model_doc/roformer>` (from ZhuiyiTechnology), released together with the paper a `RoFormer:
+    Enhanced Transformer with Rotary Position Embedding <https://arxiv.org/pdf/2104.09864v1.pdf>`__ by Jianlin Su and
+    Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
+51. :doc:`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.
-47. :doc:`SqueezeBert <model_doc/squeezebert>` released with the paper `SqueezeBERT: What can computer vision teach NLP
+52. :doc:`SqueezeBert <model_doc/squeezebert>` 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.
-48. :doc:`T5 <model_doc/t5>` (from Google AI) released with the paper `Exploring the Limits of Transfer Learning with a
+53. :doc:`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.
-49. :doc:`TAPAS <model_doc/tapas>` (from Google AI) released with the paper `TAPAS: Weakly Supervised Table Parsing via
+54. :doc:`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.
-50. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
+55. :doc:`Transformer-XL <model_doc/transformerxl>` (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.
-51. :doc:`Vision Transformer (ViT) <model_doc/vit>` (from Google AI) released with the paper `An Image is Worth 16x16
+56. :doc:`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.
-52. :doc:`Wav2Vec2 <model_doc/wav2vec2>` (from Facebook AI) released with the paper `wav2vec 2.0: A Framework for
+57. :doc:`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.
+58. :doc:`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.
-53. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
+59. :doc:`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.
-54. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
+60. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (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.
-55. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
+61. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (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.
-56. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive
+62. :doc:`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.
-57. :doc:`XLSR-Wav2Vec2 <model_doc/xlsr_wav2vec2>` (from Facebook AI) released with the paper `Unsupervised
+63. :doc:`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.
 
 
-.. _bigtable:
+Supported frameworks
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 The table below represents the current support in the library for each of those models, whether they have a Python
 tokenizer (called "slow"). A "fast" tokenizer backed by the 🤗 Tokenizers library, whether they have support in Jax (via
@@ -274,13 +307,13 @@ Flax), PyTorch, and/or TensorFlow.
 +=============================+================+================+=================+====================+==============+
 |           ALBERT            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
-|            BART             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|            BART             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |            BERT             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |       Bert Generation       |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
-|           BigBird           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|           BigBird           |       ✅       |       ✅       |       ✅        |         ❌         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |       BigBirdPegasus        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
@@ -288,14 +321,18 @@ Flax), PyTorch, and/or TensorFlow.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |       BlenderbotSmall       |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
-|            CLIP             |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|            CLIP             |       ✅       |       ✅       |       ✅        |         ❌         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |            CTRL             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |          CamemBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
+|           Canine            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
++-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |          ConvBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
+|            DETR             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
++-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |             DPR             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |           DeBERTa           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
@@ -316,7 +353,9 @@ Flax), PyTorch, and/or TensorFlow.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |     Funnel Transformer      |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
-|           GPT Neo           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           GPT Neo           |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
++-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
+|           Hubert            |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |           I-BERT            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
@@ -334,7 +373,7 @@ Flax), PyTorch, and/or TensorFlow.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |            MPNet            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
-|           Marian            |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
+|           Marian            |       ✅       |       ❌       |       ✅        |         ✅         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |        MegatronBert         |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
@@ -342,7 +381,7 @@ Flax), PyTorch, and/or TensorFlow.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |         OpenAI GPT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
-|        OpenAI GPT-2         |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|        OpenAI GPT-2         |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |           Pegasus           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
@@ -356,19 +395,23 @@ Flax), PyTorch, and/or TensorFlow.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |           RoBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
+|          RoFormer           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
++-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |         Speech2Text         |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |         SqueezeBERT         |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
-|             T5              |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|             T5              |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |            TAPAS            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |       Transformer-XL        |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
-|             ViT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|             ViT             |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
-|          Wav2Vec2           |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|         VisualBert          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
++-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
+|          Wav2Vec2           |       ✅       |       ❌       |       ✅        |         ✅         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |             XLM             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
@@ -378,7 +421,7 @@ Flax), PyTorch, and/or TensorFlow.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |            XLNet            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
-|            mBART            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+|            mBART            |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |             mT5             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
@@ -420,6 +463,8 @@ Flax), PyTorch, and/or TensorFlow.
     contributing
     add_new_model
     fast_tokenizers
+    performance
+    parallelism
     testing
     debugging
     serialization
@@ -447,6 +492,7 @@ Flax), PyTorch, and/or TensorFlow.
     main_classes/processors
     main_classes/tokenizer
     main_classes/trainer
+    main_classes/deepspeed
     main_classes/feature_extractor
 
 .. toctree::
@@ -466,7 +512,9 @@ Flax), PyTorch, and/or TensorFlow.
     model_doc/blenderbot
     model_doc/blenderbot_small
     model_doc/bort
+    model_doc/byt5
     model_doc/camembert
+    model_doc/canine
     model_doc/clip
     model_doc/convbert
     model_doc/cpm
@@ -474,6 +522,7 @@ Flax), PyTorch, and/or TensorFlow.
     model_doc/deberta
     model_doc/deberta_v2
     model_doc/deit
+    model_doc/detr
     model_doc/dialogpt
     model_doc/distilbert
     model_doc/dpr
@@ -500,6 +549,7 @@ Flax), PyTorch, and/or TensorFlow.
     model_doc/gpt
     model_doc/gpt2
     model_doc/gpt_neo
+    model_doc/hubert
     model_doc/pegasus
     model_doc/phobert
     model_doc/prophetnet
@@ -507,12 +557,14 @@ Flax), PyTorch, and/or TensorFlow.
     model_doc/reformer
     model_doc/retribert
     model_doc/roberta
+    model_doc/roformer
     model_doc/speech_to_text
     model_doc/squeezebert
     model_doc/t5
     model_doc/tapas
     model_doc/transformerxl
     model_doc/vit
+    model_doc/visual_bert
     model_doc/wav2vec2
     model_doc/xlm
     model_doc/xlmprophetnet

docs/source/installation.md

@@ -107,7 +107,7 @@ This command performs a magical link between the folder you cloned the repositor
 ```
 now this editable install will reside where you clone the folder to, e.g. `~/transformers/` and python will search it too.
 
-Do note that you have to keep that `transformers` folder around and not delete it to continue using the  `transfomers` library.
+Do note that you have to keep that `transformers` folder around and not delete it to continue using the  `transformers` library.
 
 Now, let's get to the real benefit of this installation approach. Say, you saw some new feature has been just committed into `master`. If you have already performed all the steps above, to update your transformers to include all the latest commits, all you need to do is to `cd` into that cloned repository folder and update the clone to the latest version:
 
@@ -172,7 +172,19 @@ python examples/pytorch/translation/run_translation.py --model_name_or_path t5-s
 ```
 and it should succeed without any hanging waiting to timeout.
 
+#### Fetching models and tokenizers to use offline
 
+When running a script the first time like mentioned above, the downloaded files will be cached for future reuse. 
+However, it is also possible to download files and point to their local path instead.
+
+Downloading files can be done through the Web Interface by clicking on the "Download" button, but it can also be handled
+programmatically using the `huggingface_hub` library that is a dependency to `transformers`:
+
+- Using `snapshot_download` to download an entire repository
+- Using `hf_hub_download` to download a specific file
+
+See the reference for these methods in the huggingface_hub
+[documentation](https://github.com/huggingface/huggingface_hub/tree/main/src/huggingface_hub).
 
 ## Do you want to run a Transformer model on a mobile device?
 

docs/source/internal/file_utils.rst

@@ -51,4 +51,4 @@ Special Properties
 Other Utilities
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: transformers.file_utils._BaseLazyModule
+.. autoclass:: transformers.file_utils._LazyModule

docs/source/internal/generation_utils.rst

@@ -13,19 +13,21 @@
 Utilities for Generation
 -----------------------------------------------------------------------------------------------------------------------
 
-This page lists all the utility functions used by :meth:`~transformers.PreTrainedModel.generate`,
-:meth:`~transformers.PreTrainedModel.greedy_search`, :meth:`~transformers.PreTrainedModel.sample`,
-:meth:`~transformers.PreTrainedModel.beam_search`, :meth:`~transformers.PreTrainedModel.beam_sample`, and
-:meth:`~transformers.PreTrainedModel.group_beam_search`.
+This page lists all the utility functions used by :meth:`~transformers.generation_utils.GenerationMixin.generate`,
+:meth:`~transformers.generation_utils.GenerationMixin.greedy_search`,
+:meth:`~transformers.generation_utils.GenerationMixin.sample`,
+:meth:`~transformers.generation_utils.GenerationMixin.beam_search`,
+:meth:`~transformers.generation_utils.GenerationMixin.beam_sample`, and
+:meth:`~transformers.generation_utils.GenerationMixin.group_beam_search`.
 
 Most of those are only useful if you are studying the code of the generate methods in the library.
 
 Generate Outputs
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-The output of :meth:`~transformers.PreTrainedModel.generate` is an instance of a subclass of
+The output of :meth:`~transformers.generation_utils.GenerationMixin.generate` is an instance of a subclass of
 :class:`~transformers.file_utils.ModelOutput`. This output is a data structure containing all the information returned
-by :meth:`~transformers.PreTrainedModel.generate`, but that can also be used as tuple or dictionary.
+by :meth:`~transformers.generation_utils.GenerationMixin.generate`, but that can also be used as tuple or dictionary.
 
 Here's an example:
 
@@ -78,6 +80,9 @@ GreedySearchOutput
 .. autoclass:: transformers.generation_utils.GreedySearchEncoderDecoderOutput
     :members:
 
+.. autoclass:: transformers.generation_flax_utils.FlaxGreedySearchOutput
+    :members:
+
 
 SampleOutput
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -88,6 +93,9 @@ SampleOutput
 .. autoclass:: transformers.generation_utils.SampleEncoderDecoderOutput
     :members:
 
+.. autoclass:: transformers.generation_flax_utils.FlaxSampleOutput
+    :members:
+
 
 BeamSearchOutput
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -160,6 +168,33 @@ generation.
 .. autoclass:: transformers.InfNanRemoveLogitsProcessor
     :members: __call__
 
+.. autoclass:: transformers.FlaxLogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.FlaxLogitsProcessorList
+    :members: __call__
+
+.. autoclass:: transformers.FlaxLogitsWarper
+    :members: __call__
+
+.. autoclass:: transformers.FlaxTemperatureLogitsWarper
+    :members: __call__
+
+.. autoclass:: transformers.FlaxTopPLogitsWarper
+    :members: __call__
+
+.. autoclass:: transformers.FlaxTopKLogitsWarper
+    :members: __call__
+
+.. autoclass:: transformers.FlaxForcedBOSTokenLogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.FlaxForcedEOSTokenLogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.FlaxMinLengthLogitsProcessor
+    :members: __call__
+
 
 StoppingCriteria
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

docs/source/main_classes/configuration.rst

@@ -22,4 +22,5 @@ PretrainedConfig
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.PretrainedConfig
+    :special-members: push_to_hub
     :members:

docs/source/main_classes/model.rst

@@ -1,4 +1,4 @@
-.. 
+..
     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
@@ -26,17 +26,50 @@ are common among all the models to:
 
 The other methods that are common to each model are defined in :class:`~transformers.modeling_utils.ModuleUtilsMixin`
 (for the PyTorch models) and :class:`~transformers.modeling_tf_utils.TFModuleUtilsMixin` (for the TensorFlow models) or
-for text generation, :class:`~transformers.generation_utils.GenerationMixin` (for the PyTorch models) and
-:class:`~transformers.generation_tf_utils.TFGenerationMixin` (for the TensorFlow models)
+for text generation, :class:`~transformers.generation_utils.GenerationMixin` (for the PyTorch models),
+:class:`~transformers.generation_tf_utils.TFGenerationMixin` (for the TensorFlow models) and
+:class:`~transformers.generation_flax_utils.FlaxGenerationMixin` (for the Flax/JAX models).
 
 
 PreTrainedModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.PreTrainedModel
+    :special-members: push_to_hub
     :members:
 
 
+.. _from_pretrained-torch-dtype:
+
+Model Instantiation dtype
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Under Pytorch a model normally gets instantiated with ``torch.float32`` format. This can be an issue if one tries to
+load a model whose weights are in fp16, since it'd require twice as much memory. To overcome this limitation, you can
+either explicitly pass the desired ``dtype`` using ``torch_dtype`` argument:
+
+.. code-block:: python
+
+    model = T5ForConditionalGeneration.from_pretrained("t5", torch_dtype=torch.float16)
+
+or, if you want the model to always load in the most optimal memory pattern, you can use the special value ``"auto"``,
+and then ``dtype`` will be automatically derived from the model's weights:
+
+.. code-block:: python
+
+    model = T5ForConditionalGeneration.from_pretrained("t5", torch_dtype="auto")
+
+Models instantiated from scratch can also be told which ``dtype`` to use with:
+
+.. code-block:: python
+
+    config = T5Config.from_pretrained("t5")
+    model = AutoModel.from_config(config)
+
+Due to Pytorch design, this functionality is only available for floating dtypes.
+
+
+
 ModuleUtilsMixin
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -48,6 +81,7 @@ TFPreTrainedModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.TFPreTrainedModel
+    :special-members: push_to_hub
     :members:
 
 
@@ -62,6 +96,7 @@ FlaxPreTrainedModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.FlaxPreTrainedModel
+    :special-members: push_to_hub
     :members:
 
 
@@ -74,6 +109,9 @@ Generation
 .. autoclass:: transformers.generation_tf_utils.TFGenerationMixin
     :members:
 
+.. autoclass:: transformers.generation_flax_utils.FlaxGenerationMixin
+    :members:
+
 
 Pushing to the Hub
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

docs/source/main_classes/pipelines.rst

@@ -27,6 +27,7 @@ There are two categories of pipeline abstractions to be aware about:
     - :class:`~transformers.ConversationalPipeline`
     - :class:`~transformers.FeatureExtractionPipeline`
     - :class:`~transformers.FillMaskPipeline`
+    - :class:`~transformers.ImageClassificationPipeline`
     - :class:`~transformers.QuestionAnsweringPipeline`
     - :class:`~transformers.SummarizationPipeline`
     - :class:`~transformers.TextClassificationPipeline`
@@ -36,7 +37,6 @@ There are two categories of pipeline abstractions to be aware about:
     - :class:`~transformers.ZeroShotClassificationPipeline`
     - :class:`~transformers.Text2TextGenerationPipeline`
     - :class:`~transformers.TableQuestionAnsweringPipeline`
-    - :class:`~transformers.ImageClassificationPipeline`
 
 The pipeline abstraction
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

docs/source/main_classes/tokenizer.rst

@@ -53,10 +53,8 @@ PreTrainedTokenizer
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.PreTrainedTokenizer
-    :special-members: __call__
-    :members: batch_decode, convert_ids_to_tokens, convert_tokens_to_ids, convert_tokens_to_string, decode, encode, 
-        get_added_vocab, get_special_tokens_mask, num_special_tokens_to_add, prepare_for_tokenization, tokenize,
-        vocab_size
+    :special-members: __call__, batch_decode, decode, encode, push_to_hub
+    :members: 
 
 
 PreTrainedTokenizerFast
@@ -68,10 +66,8 @@ loaded very simply into 🤗 transformers. Take a look at the :doc:`Using tokeni
 <../fast_tokenizers>` page to understand how this is done.
 
 .. autoclass:: transformers.PreTrainedTokenizerFast
-    :special-members: __call__
-    :members: batch_decode, convert_ids_to_tokens, convert_tokens_to_ids, convert_tokens_to_string, decode, encode, 
-        get_added_vocab, get_special_tokens_mask, num_special_tokens_to_add,
-        set_truncation_and_padding,tokenize, vocab_size
+    :special-members: __call__, batch_decode, decode, encode, push_to_hub
+    :members:
 
 
 BatchEncoding

docs/source/migration.md

@@ -23,7 +23,7 @@ expected changes:
 
 #### 1. AutoTokenizers and pipelines now use fast (rust) tokenizers by default.
 
-The python and rust tokenizers have roughly the same API, but the rust tokenizers have a more complete feature set. 
+The python and rust tokenizers have roughly the same API, but the rust tokenizers have a more complete feature set.
 
 This introduces two breaking changes:
 - The handling of overflowing tokens between the python and rust tokenizers is different.
@@ -85,7 +85,7 @@ This is a breaking change as importing intermediary layers using a model's modul
 
 ##### How to obtain the same behavior as v3.x in v4.x
 
-In order to obtain the same behavior as version `v3.x`, you should update the path used to access the layers. 
+In order to obtain the same behavior as version `v3.x`, you should update the path used to access the layers.
 
 In version `v3.x`:
 ```bash

docs/source/model_doc/auto.rst

@@ -205,6 +205,13 @@ FlaxAutoModel
     :members:
 
 
+FlaxAutoModelForCausalLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxAutoModelForCausalLM
+    :members:
+
+
 FlaxAutoModelForPreTraining
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -219,6 +226,13 @@ FlaxAutoModelForMaskedLM
     :members:
 
 
+FlaxAutoModelForSeq2SeqLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxAutoModelForSeq2SeqLM
+    :members:
+
+
 FlaxAutoModelForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -252,3 +266,10 @@ FlaxAutoModelForNextSentencePrediction
 
 .. autoclass:: transformers.FlaxAutoModelForNextSentencePrediction
     :members:
+
+
+FlaxAutoModelForImageClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxAutoModelForImageClassification
+    :members:

docs/source/model_doc/bart.rst

@@ -61,7 +61,7 @@ Implementation Notes
 - Model predictions are intended to be identical to the original implementation when
   :obj:`force_bos_token_to_be_generated=True`. This only works, however, if the string you pass to
   :func:`fairseq.encode` starts with a space.
-- :meth:`~transformers.BartForConditionalGeneration.generate` should be used for conditional generation tasks like
+- :meth:`~transformers.generation_utils.GenerationMixin.generate` should be used for conditional generation tasks like
   summarization, see the example in that docstrings.
 - Models that load the `facebook/bart-large-cnn` weights will not have a :obj:`mask_token_id`, or be able to perform
   mask-filling tasks.
@@ -131,6 +131,7 @@ BartForQuestionAnswering
 .. autoclass:: transformers.BartForQuestionAnswering
     :members: forward
 
+
 BartForCausalLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -138,7 +139,6 @@ BartForCausalLM
     :members: forward
 
 
-
 TFBartModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -151,3 +151,32 @@ TFBartForConditionalGeneration
 
 .. autoclass:: transformers.TFBartForConditionalGeneration
     :members: call
+
+
+FlaxBartModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBartModel
+    :members: __call__, encode, decode
+
+
+FlaxBartForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBartForConditionalGeneration
+    :members: __call__, encode, decode
+
+
+FlaxBartForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBartForSequenceClassification
+    :members: __call__, encode, decode
+
+
+FlaxBartForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBartForQuestionAnswering
+    :members: __call__, encode, decode
+

docs/source/model_doc/bigbird.rst

@@ -134,3 +134,52 @@ BigBirdForQuestionAnswering
 
 .. autoclass:: transformers.BigBirdForQuestionAnswering
     :members: forward
+
+
+FlaxBigBirdModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBigBirdModel
+    :members: __call__
+
+
+FlaxBigBirdForPreTraining
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBigBirdForPreTraining
+    :members: __call__
+
+
+FlaxBigBirdForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBigBirdForMaskedLM
+    :members: __call__
+
+
+FlaxBigBirdForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBigBirdForSequenceClassification
+    :members: __call__
+
+
+FlaxBigBirdForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBigBirdForMultipleChoice
+    :members: __call__
+
+
+FlaxBigBirdForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBigBirdForTokenClassification
+    :members: __call__
+
+
+FlaxBigBirdForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBigBirdForQuestionAnswering
+    :members: __call__

docs/source/model_doc/byt5.rst

@@ -0,0 +1,83 @@
+.. 
+    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.
+
+ByT5
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The ByT5 model was presented in `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.
+
+The abstract from the paper is the following:
+
+*Most widely-used pre-trained language models operate on sequences of tokens corresponding to word or subword units.
+Encoding text as a sequence of tokens requires a tokenizer, which is typically created as an independent artifact from
+the model. Token-free models that instead operate directly on raw text (bytes or characters) have many benefits: they
+can process text in any language out of the box, they are more robust to noise, and they minimize technical debt by
+removing complex and error-prone text preprocessing pipelines. Since byte or character sequences are longer than token
+sequences, past work on token-free models has often introduced new model architectures designed to amortize the cost of
+operating directly on raw text. In this paper, we show that a standard Transformer architecture can be used with
+minimal modifications to process byte sequences. We carefully characterize the trade-offs in terms of parameter count,
+training FLOPs, and inference speed, and show that byte-level models are competitive with their token-level
+counterparts. We also demonstrate that byte-level models are significantly more robust to noise and perform better on
+tasks that are sensitive to spelling and pronunciation. As part of our contribution, we release a new set of
+pre-trained byte-level Transformer models based on the T5 architecture, as well as all code and data used in our
+experiments.*
+
+This model was contributed by `patrickvonplaten <https://huggingface.co/patrickvonplaten>`__. The original code can be
+found `here <https://github.com/google-research/byt5>`__.
+
+
+ByT5's architecture is based on the T5 model, so one can refer to :doc:`T5's documentation page <t5>`.
+
+
+Example
+_______________________________________________________________________________________________________________________
+
+ByT5 works on raw UTF-8 bytes, so it can be used without a tokenizer:
+
+.. code-block::
+
+    from transformers import T5ForConditionalGeneration
+    import torch
+
+    model = T5ForConditionalGeneration.from_pretrained('google/byt5-small')
+
+    input_ids = torch.tensor([list("Life is like a box of chocolates.".encode("utf-8"))]) + 3  # add 3 for special tokens
+    labels = torch.tensor([list("La vie est comme une boîte de chocolat.".encode("utf-8"))]) + 3  # add 3 for special tokens
+
+    loss = model(input_ids, labels=labels).loss # forward pass
+
+
+For batched inference and training it is however recommended to make use of the tokenizer:
+
+.. code-block::
+
+    from transformers import T5ForConditionalGeneration, AutoTokenizer
+
+    model = T5ForConditionalGeneration.from_pretrained('google/byt5-small')
+    tokenizer = AutoTokenizer.from_pretrained('google/byt5-small')
+
+    model_inputs = tokenizer(["Life is like a box of chocolates.", "Today is Monday."], padding="longest", return_tensors="pt")
+    labels = tokenizer(["La vie est comme une boîte de chocolat.", "Aujourd'hui c'est lundi."], padding="longest", return_tensors="pt").input_ids
+
+    loss = model(**model_inputs, labels=labels).loss # forward pass
+
+ByT5Tokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ByT5Tokenizer
+
+See :class:`~transformers.ByT5Tokenizer` for all details.

docs/source/model_doc/canine.rst

@@ -0,0 +1,155 @@
+.. 
+    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.
+
+CANINE
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The CANINE model was proposed in `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. It's
+among the first papers that trains a Transformer without using an explicit tokenization step (such as Byte Pair
+Encoding (BPE), WordPiece or SentencePiece). Instead, the model is trained directly at a Unicode character-level.
+Training at a character-level inevitably comes with a longer sequence length, which CANINE solves with an efficient
+downsampling strategy, before applying a deep Transformer encoder.
+
+The abstract from the paper is the following:
+
+*Pipelined NLP systems have largely been superseded by end-to-end neural modeling, yet nearly all commonly-used models
+still require an explicit tokenization step. While recent tokenization approaches based on data-derived subword
+lexicons are less brittle than manually engineered tokenizers, these techniques are not equally suited to all
+languages, and the use of any fixed vocabulary may limit a model's ability to adapt. In this paper, we present CANINE,
+a neural encoder that operates directly on character sequences, without explicit tokenization or vocabulary, and a
+pre-training strategy that operates either directly on characters or optionally uses subwords as a soft inductive bias.
+To use its finer-grained input effectively and efficiently, CANINE combines downsampling, which reduces the input
+sequence length, with a deep transformer stack, which encodes context. CANINE outperforms a comparable mBERT model by
+2.8 F1 on TyDi QA, a challenging multilingual benchmark, despite having 28% fewer model parameters.*
+
+Tips:
+
+- CANINE uses no less than 3 Transformer encoders internally: 2 "shallow" encoders (which only consist of a single
+  layer) and 1 "deep" encoder (which is a regular BERT encoder). First, a "shallow" encoder is used to contextualize
+  the character embeddings, using local attention. Next, after downsampling, a "deep" encoder is applied. Finally,
+  after upsampling, a "shallow" encoder is used to create the final character embeddings. Details regarding up- and
+  downsampling can be found in the paper.
+- CANINE uses a max sequence length of 2048 characters by default. One can use :class:`~transformers.CanineTokenizer`
+  to prepare text for the model.
+- Classification can be done by placing a linear layer on top of the final hidden state of the special [CLS] token
+  (which has a predefined Unicode code point). For token classification tasks however, the downsampled sequence of
+  tokens needs to be upsampled again to match the length of the original character sequence (which is 2048). The
+  details for this can be found in the paper.
+-  Models:
+
+      - `google/canine-c <https://huggingface.co/google/canine-c>`__: Pre-trained with autoregressive character loss,
+        12-layer, 768-hidden, 12-heads, 121M parameters (size ~500 MB).
+      - `google/canine-s <https://huggingface.co/google/canine-s>`__: Pre-trained with subword loss, 12-layer,
+        768-hidden, 12-heads, 121M parameters (size ~500 MB).
+
+This model was contributed by `nielsr <https://huggingface.co/nielsr>`__. The original code can be found `here
+<https://github.com/google-research/language/tree/master/language/canine>`__.
+
+
+Example
+_______________________________________________________________________________________________________________________
+
+CANINE works on raw characters, so it can be used without a tokenizer:
+
+.. code-block::
+
+    from transformers import CanineModel
+    import torch
+
+    model = CanineModel.from_pretrained('google/canine-c') # model pre-trained with autoregressive character loss
+
+    text = "hello world"
+    # use Python's built-in ord() function to turn each character into its unicode code point id
+    input_ids = torch.tensor([[ord(char) for char in text]])
+
+    outputs = model(input_ids) # forward pass
+    pooled_output = outputs.pooler_output
+    sequence_output = outputs.last_hidden_state
+
+
+For batched inference and training, it is however recommended to make use of the tokenizer (to pad/truncate all
+sequences to the same length):
+
+.. code-block::
+
+    from transformers import CanineTokenizer, CanineModel
+
+    model = CanineModel.from_pretrained('google/canine-c')
+    tokenizer = CanineTokenizer.from_pretrained('google/canine-c')
+
+    inputs = ["Life is like a box of chocolates.", "You never know what you gonna get."]
+    encoding = tokenizer(inputs, padding="longest", truncation=True, return_tensors="pt")
+
+    outputs = model(**encoding) # forward pass
+    pooled_output = outputs.pooler_output
+    sequence_output = outputs.last_hidden_state
+
+
+CANINE specific outputs
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.models.canine.modeling_canine.CanineModelOutputWithPooling
+    :members:
+
+
+CanineConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CanineConfig
+    :members:
+
+
+CanineTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CanineTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences
+
+
+CanineModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CanineModel
+    :members: forward
+
+
+CanineForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CanineForSequenceClassification
+    :members: forward
+
+
+CanineForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CanineForMultipleChoice
+    :members: forward
+
+
+CanineForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CanineForTokenClassification
+    :members: forward
+
+
+CanineForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CanineForQuestionAnswering
+    :members: forward

docs/source/model_doc/clip.rst

@@ -60,7 +60,6 @@ encode the text and prepare the images. The following example shows how to get t
 
 .. code-block::
 
-        >>> import torch
         >>> from PIL import Image
         >>> import requests
 
@@ -152,3 +151,24 @@ CLIPVisionModel
 
 .. autoclass:: transformers.CLIPVisionModel
     :members: forward
+
+
+FlaxCLIPModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxCLIPModel
+    :members: __call__, get_text_features, get_image_features
+
+
+FlaxCLIPTextModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxCLIPTextModel
+    :members: __call__
+
+
+FlaxCLIPVisionModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxCLIPVisionModel
+    :members: __call__

docs/source/model_doc/detr.rst

@@ -0,0 +1,207 @@
+.. 
+    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.
+
+DETR
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The DETR model was proposed in `End-to-End Object Detection with Transformers <https://arxiv.org/abs/2005.12872>`__ by
+Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov and Sergey Zagoruyko. DETR
+consists of a convolutional backbone followed by an encoder-decoder Transformer which can be trained end-to-end for
+object detection. It greatly simplifies a lot of the complexity of models like Faster-R-CNN and Mask-R-CNN, which use
+things like region proposals, non-maximum suppression procedure and anchor generation. Moreover, DETR can also be
+naturally extended to perform panoptic segmentation, by simply adding a mask head on top of the decoder outputs.
+
+The abstract from the paper is the following:
+
+*We present a new method that views object detection as a direct set prediction problem. Our approach streamlines the
+detection pipeline, effectively removing the need for many hand-designed components like a non-maximum suppression
+procedure or anchor generation that explicitly encode our prior knowledge about the task. The main ingredients of the
+new framework, called DEtection TRansformer or DETR, are a set-based global loss that forces unique predictions via
+bipartite matching, and a transformer encoder-decoder architecture. Given a fixed small set of learned object queries,
+DETR reasons about the relations of the objects and the global image context to directly output the final set of
+predictions in parallel. The new model is conceptually simple and does not require a specialized library, unlike many
+other modern detectors. DETR demonstrates accuracy and run-time performance on par with the well-established and
+highly-optimized Faster RCNN baseline on the challenging COCO object detection dataset. Moreover, DETR can be easily
+generalized to produce panoptic segmentation in a unified manner. We show that it significantly outperforms competitive
+baselines.*
+
+This model was contributed by `nielsr <https://huggingface.co/nielsr>`__. The original code can be found `here
+<https://github.com/facebookresearch/detr>`__.
+
+The quickest way to get started with DETR is by checking the `example notebooks
+<https://github.com/NielsRogge/Transformers-Tutorials/tree/master/DETR>`__ (which showcase both inference and
+fine-tuning on custom data).
+
+Here's a TLDR explaining how :class:`~transformers.DetrForObjectDetection` works:
+
+First, an image is sent through a pre-trained convolutional backbone (in the paper, the authors use
+ResNet-50/ResNet-101). Let's assume we also add a batch dimension. This means that the input to the backbone is a
+tensor of shape :obj:`(batch_size, 3, height, width)`, assuming the image has 3 color channels (RGB). The CNN backbone
+outputs a new lower-resolution feature map, typically of shape :obj:`(batch_size, 2048, height/32, width/32)`. This is
+then projected to match the hidden dimension of the Transformer of DETR, which is :obj:`256` by default, using a
+:obj:`nn.Conv2D` layer. So now, we have a tensor of shape :obj:`(batch_size, 256, height/32, width/32).` Next, the
+feature map is flattened and transposed to obtain a tensor of shape :obj:`(batch_size, seq_len, d_model)` =
+:obj:`(batch_size, width/32*height/32, 256)`. So a difference with NLP models is that the sequence length is actually
+longer than usual, but with a smaller :obj:`d_model` (which in NLP is typically 768 or higher).
+
+Next, this is sent through the encoder, outputting :obj:`encoder_hidden_states` of the same shape (you can consider
+these as image features). Next, so-called **object queries** are sent through the decoder. This is a tensor of shape
+:obj:`(batch_size, num_queries, d_model)`, with :obj:`num_queries` typically set to 100 and initialized with zeros.
+These input embeddings are learnt positional encodings that the authors refer to as object queries, and similarly to
+the encoder, they are added to the input of each attention layer. Each object query will look for a particular object
+in the image. The decoder updates these embeddings through multiple self-attention and encoder-decoder attention layers
+to output :obj:`decoder_hidden_states` of the same shape: :obj:`(batch_size, num_queries, d_model)`. Next, two heads
+are added on top for object detection: a linear layer for classifying each object query into one of the objects or "no
+object", and a MLP to predict bounding boxes for each query.
+
+The model is trained using a **bipartite matching loss**: so what we actually do is compare the predicted classes +
+bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N
+(so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as
+bounding box). The `Hungarian matching algorithm <https://en.wikipedia.org/wiki/Hungarian_algorithm>`__ is used to find
+an optimal one-to-one mapping of each of the N queries to each of the N annotations. Next, standard cross-entropy (for
+the classes) and a linear combination of the L1 and `generalized IoU loss <https://giou.stanford.edu/>`__ (for the
+bounding boxes) are used to optimize the parameters of the model.
+
+DETR can be naturally extended to perform panoptic segmentation (which unifies semantic segmentation and instance
+segmentation). :class:`~transformers.DetrForSegmentation` adds a segmentation mask head on top of
+:class:`~transformers.DetrForObjectDetection`. The mask head can be trained either jointly, or in a two steps process,
+where one first trains a :class:`~transformers.DetrForObjectDetection` model to detect bounding boxes around both
+"things" (instances) and "stuff" (background things like trees, roads, sky), then freeze all the weights and train only
+the mask head for 25 epochs. Experimentally, these two approaches give similar results. Note that predicting boxes is
+required for the training to be possible, since the Hungarian matching is computed using distances between boxes.
+
+Tips:
+
+- DETR uses so-called **object queries** to detect objects in an image. The number of queries determines the maximum
+  number of objects that can be detected in a single image, and is set to 100 by default (see parameter
+  :obj:`num_queries` of :class:`~transformers.DetrConfig`). Note that it's good to have some slack (in COCO, the
+  authors used 100, while the maximum number of objects in a COCO image is ~70).
+- The decoder of DETR updates the query embeddings in parallel. This is different from language models like GPT-2,
+  which use autoregressive decoding instead of parallel. Hence, no causal attention mask is used.
+- DETR adds position embeddings to the hidden states at each self-attention and cross-attention layer before projecting
+  to queries and keys. For the position embeddings of the image, one can choose between fixed sinusoidal or learned
+  absolute position embeddings. By default, the parameter :obj:`position_embedding_type` of
+  :class:`~transformers.DetrConfig` is set to :obj:`"sine"`.
+- During training, the authors of DETR did find it helpful to use auxiliary losses in the decoder, especially to help
+  the model output the correct number of objects of each class. If you set the parameter :obj:`auxiliary_loss` of
+  :class:`~transformers.DetrConfig` to :obj:`True`, then prediction feedforward neural networks and Hungarian losses
+  are added after each decoder layer (with the FFNs sharing parameters).
+- If you want to train the model in a distributed environment across multiple nodes, then one should update the
+  `num_boxes` variable in the `DetrLoss` class of `modeling_detr.py`. When training on multiple nodes, this should be
+  set to the average number of target boxes across all nodes, as can be seen in the original implementation `here
+  <https://github.com/facebookresearch/detr/blob/a54b77800eb8e64e3ad0d8237789fcbf2f8350c5/models/detr.py#L227-L232>`__.
+- :class:`~transformers.DetrForObjectDetection` and :class:`~transformers.DetrForSegmentation` can be initialized with
+  any convolutional backbone available in the `timm library <https://github.com/rwightman/pytorch-image-models>`__.
+  Initializing with a MobileNet backbone for example can be done by setting the :obj:`backbone` attribute of
+  :class:`~transformers.DetrConfig` to :obj:`"tf_mobilenetv3_small_075"`, and then initializing the model with that
+  config.
+- DETR resizes the input images such that the shortest side is at least a certain amount of pixels while the longest is
+  at most 1333 pixels. At training time, scale augmentation is used such that the shortest side is randomly set to at
+  least 480 and at most 800 pixels. At inference time, the shortest side is set to 800. One can use
+  :class:`~transformers.DetrFeatureExtractor` to prepare images (and optional annotations in COCO format) for the
+  model. Due to this resizing, images in a batch can have different sizes. DETR solves this by padding images up to the
+  largest size in a batch, and by creating a pixel mask that indicates which pixels are real/which are padding.
+  Alternatively, one can also define a custom :obj:`collate_fn` in order to batch images together, using
+  :meth:`~transformers.DetrFeatureExtractor.pad_and_create_pixel_mask`.
+- The size of the images will determine the amount of memory being used, and will thus determine the :obj:`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.
+
+As a summary, consider the following table:
+
++---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
+| **Task**                                    | **Object detection**                                    | **Instance segmentation**                                            | **Panoptic segmentation**                                              |
++---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
+| **Description**                             | Predicting bounding boxes and class labels around       | Predicting masks around objects (i.e. instances) in an image         | Predicting masks around both objects (i.e. instances) as well as       |
+|                                             | objects in an image                                     |                                                                      | "stuff" (i.e. background things like trees and roads) in an image      |
++---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
+| **Model**                                   | :class:`~transformers.DetrForObjectDetection`           | :class:`~transformers.DetrForSegmentation`                           | :class:`~transformers.DetrForSegmentation`                             |
++---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
+| **Example dataset**                         | COCO detection                                          | COCO detection,                                                      | COCO panoptic                                                          |
+|                                             |                                                         | COCO panoptic                                                        |                                                                        |
++---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
+| **Format of annotations to provide to**     | {‘image_id’: int,                                       | {‘image_id’: int,                                                    | {‘file_name: str,                                                      |
+| :class:`~transformers.DetrFeatureExtractor` | ‘annotations’: List[Dict]}, each Dict being a COCO      | ‘annotations’: [List[Dict]] } (in case of COCO detection)            | ‘image_id: int,                                                        |
+|                                             | object annotation                                       |                                                                      | ‘segments_info’: List[Dict] }                                          |
+|                                             |                                                         | or                                                                   |                                                                        |
+|                                             |                                                         |                                                                      | and masks_path (path to directory containing PNG files of the masks)   |
+|                                             |                                                         | {‘file_name’: str,                                                   |                                                                        |
+|                                             |                                                         | ‘image_id’: int,                                                     |                                                                        |
+|                                             |                                                         | ‘segments_info’: List[Dict]} (in case of COCO panoptic)              |                                                                        |
++---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
+| **Postprocessing** (i.e. converting the     | :meth:`~transformers.DetrFeatureExtractor.post_process` | :meth:`~transformers.DetrFeatureExtractor.post_process_segmentation` | :meth:`~transformers.DetrFeatureExtractor.post_process_segmentation`,  |
+| output of the model to COCO API)            |                                                         |                                                                      | :meth:`~transformers.DetrFeatureExtractor.post_process_panoptic`       |
++---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
+| **evaluators**                              | :obj:`CocoEvaluator` with iou_types = “bbox”            | :obj:`CocoEvaluator` with iou_types = “bbox”, “segm”                 | :obj:`CocoEvaluator` with iou_tupes = “bbox, “segm”                    |
+|                                             |                                                         |                                                                      |                                                                        |
+|                                             |                                                         |                                                                      | :obj:`PanopticEvaluator`                                               |
++---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
+
+In short, one should prepare the data either in COCO detection or COCO panoptic format, then use
+:class:`~transformers.DetrFeatureExtractor` to create :obj:`pixel_values`, :obj:`pixel_mask` and optional
+:obj:`labels`, which can then be used to train (or fine-tune) a model. For evaluation, one should first convert the
+outputs of the model using one of the postprocessing methods of :class:`~transformers.DetrFeatureExtractor`. These can
+be be provided to either :obj:`CocoEvaluator` or :obj:`PanopticEvaluator`, which allow you to calculate metrics like
+mean Average Precision (mAP) and Panoptic Quality (PQ). The latter objects are implemented in the `original repository
+<https://github.com/facebookresearch/detr>`__. See the `example notebooks
+<https://github.com/NielsRogge/Transformers-Tutorials/tree/master/DETR>`__ for more info regarding evaluation.
+
+
+DETR specific outputs
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.models.detr.modeling_detr.DetrModelOutput
+    :members:
+
+.. autoclass:: transformers.models.detr.modeling_detr.DetrObjectDetectionOutput
+    :members:
+
+.. autoclass:: transformers.models.detr.modeling_detr.DetrSegmentationOutput
+    :members:
+
+
+DetrConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DetrConfig
+    :members:
+
+
+DetrFeatureExtractor
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DetrFeatureExtractor
+    :members: __call__, pad_and_create_pixel_mask, post_process, post_process_segmentation, post_process_panoptic
+
+
+DetrModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DetrModel
+    :members: forward
+
+
+DetrForObjectDetection
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DetrForObjectDetection
+    :members: forward
+
+
+DetrForSegmentation
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DetrForSegmentation
+    :members: forward

docs/source/model_doc/gpt2.rst

@@ -139,3 +139,17 @@ TFSequenceClassifierOutputWithPast
 
 .. autoclass:: transformers.modeling_tf_outputs.TFSequenceClassifierOutputWithPast
     :members:
+
+
+FlaxGPT2Model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxGPT2Model
+    :members: __call__
+
+
+FlaxGPT2LMHeadModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxGPT2LMHeadModel
+    :members: __call__

docs/source/model_doc/gpt_neo.rst

@@ -65,3 +65,22 @@ GPTNeoForCausalLM
 
 .. autoclass:: transformers.GPTNeoForCausalLM
     :members: forward
+
+GPTNeoForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.GPTNeoForSequenceClassification
+    :members: forward
+
+FlaxGPTNeoModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxGPTNeoModel
+    :members: __call__
+
+
+FlaxGPTNeoForCausalLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxGPTNeoForCausalLM
+    :members: __call__

docs/source/model_doc/hubert.rst

@@ -0,0 +1,78 @@
+.. 
+    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.
+
+Hubert
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Hubert was proposed in `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.
+
+The abstract from the paper is the following:
+
+*Self-supervised approaches for speech representation learning are challenged by three unique problems: (1) there are
+multiple sound units in each input utterance, (2) there is no lexicon of input sound units during the pre-training
+phase, and (3) sound units have variable lengths with no explicit segmentation. To deal with these three problems, we
+propose the Hidden-Unit BERT (HuBERT) approach for self-supervised speech representation learning, which utilizes an
+offline clustering step to provide aligned target labels for a BERT-like prediction loss. A key ingredient of our
+approach is applying the prediction loss over the masked regions only, which forces the model to learn a combined
+acoustic and language model over the continuous inputs. HuBERT relies primarily on the consistency of the unsupervised
+clustering step rather than the intrinsic quality of the assigned cluster labels. Starting with a simple k-means
+teacher of 100 clusters, and using two iterations of clustering, the HuBERT model either matches or improves upon the
+state-of-the-art wav2vec 2.0 performance on the Librispeech (960h) and Libri-light (60,000h) benchmarks with 10min, 1h,
+10h, 100h, and 960h fine-tuning subsets. Using a 1B parameter model, HuBERT shows up to 19% and 13% relative WER
+reduction on the more challenging dev-other and test-other evaluation subsets.*
+
+Tips:
+
+- Hubert is a speech model that accepts a float array corresponding to the raw waveform of the speech signal.
+- Hubert model was fine-tuned using connectionist temporal classification (CTC) so the model output has to be decoded
+  using :class:`~transformers.Wav2Vec2CTCTokenizer`.
+
+This model was contributed by `patrickvonplaten <https://huggingface.co/patrickvonplaten>`__.
+
+
+HubertConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.HubertConfig
+    :members:
+
+
+HubertModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.HubertModel
+    :members: forward
+
+
+HubertForCTC
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.HubertForCTC
+    :members: forward
+
+TFHubertModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFHubertModel
+    :members: call
+
+
+TFHubertForCTC
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFHubertForCTC
+    :members: call

docs/source/model_doc/marian.rst

@@ -216,3 +216,17 @@ TFMarianMTModel
 
 .. autoclass:: transformers.TFMarianMTModel
     :members: call
+
+
+FlaxMarianModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxMarianModel
+    :members: __call__
+
+
+FlaxMarianMTModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxMarianMTModel
+    :members: __call__

docs/source/model_doc/mbart.rst

@@ -240,3 +240,31 @@ TFMBartForConditionalGeneration
 
 .. autoclass:: transformers.TFMBartForConditionalGeneration
     :members: call
+
+
+FlaxMBartModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxMBartModel
+    :members: __call__, encode, decode
+
+
+FlaxMBartForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxMBartForConditionalGeneration
+    :members: __call__, encode, decode
+
+
+FlaxMBartForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxMBartForSequenceClassification
+    :members: __call__, encode, decode
+
+
+FlaxMBartForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxMBartForQuestionAnswering
+    :members: __call__, encode, decode

docs/source/model_doc/pegasus.rst

@@ -90,7 +90,7 @@ Usage Example
     >>> device = 'cuda' if torch.cuda.is_available() else 'cpu'
     >>> tokenizer = PegasusTokenizer.from_pretrained(model_name)
     >>> model = PegasusForConditionalGeneration.from_pretrained(model_name).to(device)
-    >>> batch = tokenizer(src_text, truncation=True, padding='longest', return_tensors="pt").to(torch_device)
+    >>> batch = tokenizer(src_text, truncation=True, padding='longest', return_tensors="pt").to(device)
     >>> translated = model.generate(**batch)
     >>> tgt_text = tokenizer.batch_decode(translated, skip_special_tokens=True)
     >>> assert tgt_text[0] == "California's largest electricity provider has turned off power to hundreds of thousands of customers."

docs/source/model_doc/roformer.rst

@@ -0,0 +1,161 @@
+.. 
+    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.
+
+RoFormer
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The RoFormer model was proposed in `RoFormer: Enhanced Transformer with Rotary Position Embedding
+<https://arxiv.org/pdf/2104.09864v1.pdf>`__ by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
+
+The abstract from the paper is the following:
+
+*Position encoding in transformer architecture provides supervision for dependency modeling between elements at
+different positions in the sequence. We investigate various methods to encode positional information in
+transformer-based language models and propose a novel implementation named Rotary Position Embedding(RoPE). The
+proposed RoPE encodes absolute positional information with rotation matrix and naturally incorporates explicit relative
+position dependency in self-attention formulation. Notably, RoPE comes with valuable properties such as flexibility of
+being expand to any sequence lengths, decaying inter-token dependency with increasing relative distances, and
+capability of equipping the linear self-attention with relative position encoding. As a result, the enhanced
+transformer with rotary position embedding, or RoFormer, achieves superior performance in tasks with long texts. We
+release the theoretical analysis along with some preliminary experiment results on Chinese data. The undergoing
+experiment for English benchmark will soon be updated.*
+
+Tips:
+
+- RoFormer is a BERT-like autoencoding model with rotary position embeddings. Rotary position embeddings have shown
+  improved performance on classification tasks with long texts.
+
+
+This model was contributed by `junnyu <https://huggingface.co/junnyu>`__. The original code can be found `here
+<https://github.com/ZhuiyiTechnology/roformer>`__.
+
+RoFormerConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RoFormerConfig
+    :members:
+
+
+RoFormerTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RoFormerTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+RoFormerTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RoFormerTokenizerFast
+    :members: build_inputs_with_special_tokens
+
+
+RoFormerModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RoFormerModel
+    :members: forward
+
+
+RoFormerForCausalLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RoFormerForCausalLM
+    :members: forward
+
+
+RoFormerForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RoFormerForMaskedLM
+    :members: forward
+
+
+RoFormerForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RoFormerForSequenceClassification
+    :members: forward
+
+
+RoFormerForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RoFormerForMultipleChoice
+    :members: forward
+
+
+RoFormerForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RoFormerForTokenClassification
+    :members: forward
+
+
+RoFormerForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RoFormerForQuestionAnswering
+    :members: forward
+
+
+TFRoFormerModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRoFormerModel
+    :members: call
+
+
+TFRoFormerForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRoFormerForMaskedLM
+    :members: call
+
+
+TFRoFormerForCausalLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRoFormerForCausalLM
+    :members: call
+
+
+TFRoFormerForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRoFormerForSequenceClassification
+    :members: call
+
+
+TFRoFormerForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRoFormerForMultipleChoice
+    :members: call
+
+
+TFRoFormerForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRoFormerForTokenClassification
+    :members: call
+
+
+TFRoFormerForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRoFormerForQuestionAnswering
+    :members: call

docs/source/model_doc/t5.rst

@@ -1,4 +1,4 @@
-.. 
+..
     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
@@ -44,9 +44,9 @@ Tips:
 
   For more information about which prefix to use, it is easiest to look into Appendix D of the `paper
   <https://arxiv.org/pdf/1910.10683.pdf>`__. - For sequence-to-sequence generation, it is recommended to use
-  :obj:`T5ForConditionalGeneration.generate()`. This method takes care of feeding the encoded input via cross-attention
-  layers to the decoder and auto-regressively generates the decoder output. - T5 uses relative scalar embeddings.
-  Encoder input padding can be done on the left and on the right.
+  :meth:`~transformers.generation_utils.GenerationMixin.generate`. This method takes care of feeding the encoded input
+  via cross-attention layers to the decoder and auto-regressively generates the decoder output. - T5 uses relative
+  scalar embeddings. Encoder input padding can be done on the left and on the right.
 
 This model was contributed by `thomwolf <https://huggingface.co/thomwolf>`__. The original code can be found `here
 <https://github.com/google-research/text-to-text-transfer-transformer>`__.
@@ -74,6 +74,10 @@ token. T5 can be trained / fine-tuned both in a supervised and unsupervised fash
 
 .. code-block::
 
+    from transformers import T5ForConditionalGeneration, T5Tokenizer
+    model = T5ForConditionalGeneration.from_pretrained("t5-small")
+    tokenizer = T5Tokenizer.from_pretrained("t5-small")
+
     input_ids = tokenizer('The <extra_id_0> walks in <extra_id_1> park', return_tensors='pt').input_ids
     labels = tokenizer('<extra_id_0> cute dog <extra_id_1> the <extra_id_2>', return_tensors='pt').input_ids
     # the forward function automatically creates the correct decoder_input_ids
@@ -87,6 +91,10 @@ token. T5 can be trained / fine-tuned both in a supervised and unsupervised fash
 
 .. code-block::
 
+    from transformers import T5ForConditionalGeneration, T5Tokenizer
+    model = T5ForConditionalGeneration.from_pretrained("t5-small")
+    tokenizer = T5Tokenizer.from_pretrained("t5-small")
+
     input_ids = tokenizer('translate English to German: The house is wonderful.', return_tensors='pt').input_ids
     labels = tokenizer('Das Haus ist wunderbar.', return_tensors='pt').input_ids
     # the forward function automatically creates the correct decoder_input_ids
@@ -152,3 +160,15 @@ TFT5EncoderModel
 
 .. autoclass:: transformers.TFT5EncoderModel
     :members: call
+
+FlaxT5Model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxT5Model
+    :members: __call__, encode, decode
+
+FlaxT5ForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxT5ForConditionalGeneration
+    :members: __call__, encode, decode

docs/source/model_doc/visual_bert.rst

@@ -0,0 +1,128 @@
+.. 
+    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.
+
+VisualBERT
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The VisualBERT model was proposed in `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.
+VisualBERT is a neural network trained on a variety of (image, text) pairs.
+
+The abstract from the paper is the following:
+
+*We propose VisualBERT, a simple and flexible framework for modeling a broad range of vision-and-language tasks.
+VisualBERT consists of a stack of Transformer layers that implicitly align elements of an input text and regions in an
+associated input image with self-attention. We further propose two visually-grounded language model objectives for
+pre-training VisualBERT on image caption data. Experiments on four vision-and-language tasks including VQA, VCR, NLVR2,
+and Flickr30K show that VisualBERT outperforms or rivals with state-of-the-art models while being significantly
+simpler. Further analysis demonstrates that VisualBERT can ground elements of language to image regions without any
+explicit supervision and is even sensitive to syntactic relationships, tracking, for example, associations between
+verbs and image regions corresponding to their arguments.*
+
+Tips:
+
+1. Most of the checkpoints provided work with the :class:`~transformers.VisualBertForPreTraining` configuration. Other
+   checkpoints provided are the fine-tuned checkpoints for down-stream tasks - VQA ('visualbert-vqa'), VCR
+   ('visualbert-vcr'), NLVR2 ('visualbert-nlvr2'). Hence, if you are not working on these downstream tasks, it is
+   recommended that you use the pretrained checkpoints.
+
+2. For the VCR task, the authors use a fine-tuned detector for generating visual embeddings, for all the checkpoints.
+   We do not provide the detector and its weights as a part of the package, but it will be available in the research
+   projects, and the states can be loaded directly into the detector provided.
+
+Usage
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+VisualBERT is a multi-modal vision and language model. It can be used for visual question answering, multiple choice,
+visual reasoning and region-to-phrase correspondence tasks. VisualBERT uses a BERT-like transformer to prepare
+embeddings for image-text pairs. Both the text and visual features are then projected to a latent space with identical
+dimension.
+
+To feed images to the model, each image is passed through a pre-trained object detector and the regions and the
+bounding boxes are extracted. The authors use the features generated after passing these regions through a pre-trained
+CNN like ResNet as visual embeddings. They also add absolute position embeddings, and feed the resulting sequence of
+vectors to a standard BERT model. The text input is concatenated in the front of the visual embeddings in the embedding
+layer, and is expected to be bound by [CLS] and a [SEP] tokens, as in BERT. The segment IDs must also be set
+appropriately for the textual and visual parts.
+
+The :class:`~transformers.BertTokenizer` is used to encode the text. A custom detector/feature extractor must be used
+to get the visual embeddings. For an example on how to generate visual embeddings, see the `colab notebook
+<https://colab.research.google.com/drive/1bLGxKdldwqnMVA5x4neY7-l_8fKGWQYI?usp=sharing>`__. The following example shows
+how to get the last hidden state using :class:`~transformers.VisualBertModel`:
+
+.. code-block::
+
+        >>> import torch
+        >>> from transformers import BertTokenizer, VisualBertModel
+
+        >>> model = VisualBertModel.from_pretrained("uclanlp/visualbert-vqa-coco-pre")
+        >>> tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+
+        >>> inputs = tokenizer("What is the man eating?", return_tensors="pt")
+        >>> # this is a custom function that returns the visual embeddings given the image path
+        >>> visual_embeds = get_visual_embeddings(image_path)
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+
+This model was contributed by `gchhablani <https://huggingface.co/gchhablani>`__. The original code can be found `here
+<https://github.com/uclanlp/visualbert>`__.
+
+VisualBertConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.VisualBertConfig
+    :members:
+
+VisualBertModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.VisualBertModel
+    :members: forward
+
+
+VisualBertForPreTraining
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.VisualBertForPreTraining
+    :members: forward
+
+
+VisualBertForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.VisualBertForQuestionAnswering
+    :members: forward
+
+
+VisualBertForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.VisualBertForMultipleChoice
+    :members: forward
+
+
+VisualBertForVisualReasoning
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.VisualBertForVisualReasoning
+    :members: forward
+
+
+VisualBertForRegionToPhraseAlignment
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.VisualBertForRegionToPhraseAlignment
+    :members: forward

docs/source/model_doc/vit.rst

@@ -101,3 +101,18 @@ ViTForImageClassification
 
 .. autoclass:: transformers.ViTForImageClassification
     :members: forward
+
+
+FlaxVitModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxViTModel
+    :members: __call__
+
+
+FlaxViTForImageClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxViTForImageClassification
+    :members: __call__
+

docs/source/model_doc/wav2vec2.rst

@@ -79,3 +79,43 @@ Wav2Vec2ForCTC
 
 .. autoclass:: transformers.Wav2Vec2ForCTC
     :members: forward
+
+Wav2Vec2ForPreTraining
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.Wav2Vec2ForPreTraining
+    :members: forward
+
+
+TFWav2Vec2Model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFWav2Vec2Model
+    :members: call
+
+
+TFWav2Vec2ForCTC
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFWav2Vec2ForCTC
+    :members: call
+
+
+FlaxWav2Vec2Model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxWav2Vec2Model
+    :members: __call__
+
+
+FlaxWav2Vec2ForCTC
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxWav2Vec2ForCTC
+    :members: __call__
+
+FlaxWav2Vec2ForPreTraining
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxWav2Vec2ForPreTraining
+    :members: __call__

docs/source/model_sharing.rst

@@ -16,6 +16,12 @@ Model sharing and uploading
 In this page, we will show you how to share a model you have trained or fine-tuned on new data with the community on
 the `model hub <https://huggingface.co/models>`__.
 
+.. raw:: html
+
+   <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>
+
 .. note::
 
     You will need to create an account on `huggingface.co <https://huggingface.co/join>`__ for this.
@@ -77,6 +83,12 @@ token that you can just copy.
 Directly push your model to the hub
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/Z1-XMy-GNLQ" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 Once you have an API token (either stored in the cache or copied and pasted in your notebook), you can directly push a
 finetuned model you saved in :obj:`save_drectory` by calling:
 
@@ -131,7 +143,7 @@ directly create a PyTorch version of your TensorFlow model:
 
 .. code-block:: python
 
-    from transfomers import AutoModel
+    from transformers import AutoModel
 
     model = AutoModel.from_pretrained(save_directory, from_tf=True)
 
@@ -152,6 +164,12 @@ or
 Use your terminal and git
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/rkCly_cbMBk" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 Basic steps
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 

docs/source/model_summary.rst

@@ -28,6 +28,12 @@ Each one of the models in the library falls into one of the following categories
   * :ref:`multimodal-models`
   * :ref:`retrieval-based-models`
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/H39Z_720T5s" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 Autoregressive models are pretrained on the classic language modeling task: guess the next token having read all the
 previous ones. They correspond to the decoder of the original transformer model, and a mask is used on top of the full
 sentence so that the attention heads can only see what was before in the text, and not what’s after. Although those
@@ -54,12 +60,18 @@ Multimodal models mix text inputs with other kinds (e.g. images) and are more sp
 
 .. _autoregressive-models:
 
-Autoregressive models
+Decoders or autoregressive models
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 As mentioned before, these models rely on the decoder part of the original transformer and use an attention mask so
 that at each position, the model can only look at the tokens before the attention heads.
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/d_ixlCubqQw" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 Original GPT
 -----------------------------------------------------------------------------------------------------------------------
 
@@ -215,13 +227,19 @@ multiple choice classification and question answering.
 
 .. _autoencoding-models:
 
-Autoencoding models
+Encoders or autoencoding models
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 As mentioned before, these models rely on the encoder part of the original transformer and use no mask so the model can
 look at all the tokens in the attention heads. For pretraining, targets are the original sentences and inputs are their
 corrupted versions.
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/MUqNwgPjJvQ" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 BERT
 -----------------------------------------------------------------------------------------------------------------------
 
@@ -526,6 +544,12 @@ Sequence-to-sequence models
 
 As mentioned before, these models keep both the encoder and the decoder of the original transformer.
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/0_4KEb08xrE" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 BART
 -----------------------------------------------------------------------------------------------------------------------
 

docs/source/parallelism.md

@@ -0,0 +1,354 @@
+<!---
+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.
+-->
+
+# Model Parallelism
+
+
+## Parallelism overview
+
+In the modern machine learning the various approaches to parallelism are used to:
+1. fit very large models onto limited hardware - e.g. t5-11b is 45GB in just model params
+2. significantly speed up training - finish training that would take a year in hours
+
+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.
+
+While the main concepts most likely will apply to any other framework, this article is focused on PyTorch-based implementations.
+
+
+## Concepts
+
+The following is the brief description of the main concepts that will be described later in depth in this document.
+
+1. DataParallel (DP) - the same setup is replicated multiple times, and each being fed a slice of the data. The processing is done in parallel and all setups are synchronized at the end of each training step.
+2. TensorParallel (TP) - each tensor is split up into multiple chunks, so instead of having the whole tensor reside on a single gpu, each shard of the tensor resides on its designated gpu. During processing each shard gets processed separately and in parallel on different GPUs and the results are synced at the end of the step. This is what one may call horizontal parallelism, as the splitting happens on horizontal level.
+3. PipelineParallel (PP) - the model is split up vertically (layer-level) across multiple GPUs, so that only one or several layers of the model are places on a single gpu. Each gpu processes in parallel different stages of the pipeline and working on a small chunk of the batch.
+4. Zero Redundancy Optimizer (ZeRO) - Also performs sharding of the tensors somewhat similar to TP, except the whole tensor gets reconstructed in time for a forward or backward computation, therefore the model does't need to be modified. It also supports various offloading techniques to compensate for limited GPU memory.
+5. Sharded DDP - is another name for the foundational ZeRO concept as used by various other implementations of ZeRO.
+
+
+## Data Parallel
+
+Most users with just 2 GPUs already enjoy the increased training speed up thanks to DataParallel (DP) and DistributedDataParallel (DDP) that are almost trivial to use. This is a built-in feature of Pytorch.
+
+## ZeRO Data Parallel
+
+ZeRO-powered data parallelism (ZeRO-DP) is described on the following diagram from this [blog post](https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/)
+![DeepSpeed-Image-1](imgs/parallelism-zero.png)
+
+It can be difficult to wrap one's head around it, but in reality the concept is quite simple. This is just the usual DataParallel (DP), except, instead of replicating the full model params, gradients and optimizer states, each GPU stores only a slice of it.  And then at run-time when the full layer params are needed just for the given layer, all GPUs synchronize to give each other parts that they miss - this is it.
+
+Consider this simple model with 3 layers, where each layer has 3 params:
+```
+La | Lb | Lc
+---|----|---
+a0 | b0 | c0
+a1 | b1 | c1
+a2 | b2 | c2
+```
+Layer La has weights a0, at and a2.
+
+If we have 3 GPUs, the Sharded DDP (= Zero-DP) splits the model onto 3 GPUs like so:
+
+```
+GPU0:
+La | Lb | Lc
+---|----|---
+a0 | b0 | c0
+
+GPU1:
+La | Lb | Lc
+---|----|---
+a1 | b1 | c1
+
+GPU2:
+La | Lb | Lc
+---|----|---
+a2 | b2 | c2
+```
+
+In a way this is the same horizontal slicing, as tensor parallelism, if you imagine the typical DNN diagram. Vertical slicing is where one puts whole layer-groups on different GPUs. But it's just the starting point.
+
+Now each of these GPUs will get the usual mini-batch as it works in DP:
+```
+x0 => GPU0
+x1 => GPU1
+x2 => GPU2
+```
+
+The inputs are unmodified - they think they are going to be processed by the normal model.
+
+First, the inputs hit the layer La.
+
+Let's focus just on GPU0: x0 needs a0, a1, a2 params to do its forward path, but GPU0 has only a0 - it gets sent a1 from GPU1 and a2 from GPU2, bringing all pieces of the model together.
+
+In parallel, GPU1 gets mini-batch x1 and it only has a1, but needs a0 and a2 params, so it gets those from GPU0 and GPU2.
+
+Same happens to GPU2 that gets input x2. It gets a0 and a1 from GPU0 and GPU1, and with its a2 it reconstructs the full tensor.
+
+All 3 GPUs get the full tensors reconstructed and a forward happens.
+
+As soon as the calculation is done, the data that is no longer needed gets dropped - it's only used during the calculation. The reconstruction is done efficiently via a pre-fetch.
+
+And the whole process is repeated for layer Lb, then Lc forward-wise, and then backward Lc -> Lb -> La.
+
+To me this sounds like an efficient group backpacking weight distribution strategy:
+
+1. person A carries the tent
+2. person B carries the stove
+3. person C carries the axe
+
+Now each night they all share what they have with others and get from others what the don't have, and in the morning they pack up their allocated type of gear and continue on their way. This is Sharded DDP / Zero DP.
+
+Compare this strategy to the simple one where each person has to carry their own tent, stove and axe, which would be far more inefficient. This is DataParallel (DP and DDP) in Pytorch.
+
+While reading the literature on this topic you may encounter the following synonyms: Sharded, Partitioned.
+
+If you pay close attention the way ZeRO partitions the model's weights - it looks very similar to tensor parallelism which will be discussed later. This is because it partitions/shards each layer's weights, unlike vertical model parallelism which is discussed next.
+
+Implementations:
+
+- [DeepSpeed](https://www.deepspeed.ai/features/#the-zero-redundancy-optimizer) ZeRO-DP stages 1+2+3
+- [Fairscale](https://github.com/facebookresearch/fairscale/#optimizer-state-sharding-zero) ZeRO-DP stages 1+2+3
+- [`transformers` integration](https://huggingface.co/transformers/master/main_classes/trainer.html#trainer-integrations)
+
+## Naive Model Parallel (Vertical) and Pipeline Parallel
+
+Naive Model Parallel (MP) is where one spreads groups of model layers across multiple GPUs. The mechanism is relatively simple - switch the desired layers `.to()` the desired devices and now whenever the data goes in and out those layers switch the data to the same device as the layer and leave the rest unmodified.
+
+We refer to it as Vertical MP, because if you remember how most models are drawn, we slice the layers vertically. For example, if the following diagram shows an 8-layer model:
+
+```
+===================  ===================
+|  0 | 1 | 2 | 3  |  |  4 | 5 | 6 | 7  |
+===================  ===================
+        gpu0                 gpu1
+```
+we just sliced it in 2 vertically, placing layers 0-3 onto GPU0 and 4-7 to GPU1.
+
+Now while data travels from layer 0 to 1, 1 to 2 and 2 to 3 this is just the normal model. But when data needs to pass from layer 3 to layer 4 it needs to travel from GPU0 to GPU1 which introduces a communication overhead. If the participating GPUs are on the same compute node (e.g. same physical machine) this copying is pretty fast, but if the GPUs are located on different compute nodes (e.g. multiple machines) the communication overhead could be significantly larger.
+
+Then layers 4 to 5 to 6 to 7 are as a normal model would have and when the 7th layer completes we often need to send the data back to layer 0 where the labels are (or alternatively send the labels to the the last layer). Now the loss can be computed and the optimizer can do its work.
+
+Problems:
+- the main deficiency and why this one is called "naive" MP, is that all but one GPU is idle at any given moment. So if 4 GPUs are used, it's almost identical to quadrupling the amount of memory of a single GPU, and ignoring the rest of the hardware. Plus there is the overhead of copying the data between devices. So 4x 6GB cards will be able to accommodate the same size as 1x 24GB card using naive MP, except the latter will complete the training faster, since it doesn't have the data copying overhead. But, say, if you have 40GB cards and need to fit a 45GB model you can with 4x 40GB cards (but barely because of the gradient and optimizer states)
+- shared embeddings may need to get copied back and forth between GPUs.
+
+Pipeline Parallel (PP) is almost identical to a naive MP, but it solves the GPU idling problem, by chunking the incoming batch into micro-batches and artificially creating a pipeline, which allows different GPUs to concurrently participate in the computation process.
+
+The following illustration from the [GPipe paper](https://ai.googleblog.com/2019/03/introducing-gpipe-open-source-library.html) shows the naive MP on the top, and PP on the bottom:
+
+![mp-pp](imgs/parallelism-gpipe-bubble.png)
+
+It's easy to see from the bottom diagram how PP has less dead zones, where GPUs are idle. The idle parts are referred to as the "bubble".
+
+Both parts of the diagram show a parallelism that is of degree 4. That is 4 GPUs are participating in the pipeline. So there is the forward path of 4 pipe stages F0, F1, F2 and F3 and then the return reverse order backward path of B3, B2, B1 and B0.
+
+PP introduces a new hyper-parameter to tune and it's `chunks` which defines how many chunks of data are sent in a sequence through the same pipe stage. For example, in the bottomw diagram you can see that `chunks=4`. GPU0 performs the same forward path on chunk 0, 1, 2 and 3 (F0,0, F0,1, F0,2, F0,3) and then it waits for other GPUs to do their work and only when their work is starting to be complete, GPU0 starts to work again doing the backward path for chunks 3, 2, 1 and 0 (B0,3, B0,2, B0,1, B0,0).
+
+Note that conceptually this is the same concept as gradient accumulation steps (GAS). Pytorch uses `chunks`, whereas DeepSpeed refers to the same hyper-parameter as GAS.
+
+Because of the chunks, PP introduces the concept of micro-batches (MBS). DP splits the global data batch size into mini-batches, so if you have a DP degree of 4, a global batch size of 1024 gets split up into 4 mini-batches of 256 each (1024/4). And if the number of `chunks` (or GAS) is 32 we end up with a micro-batch size of 8 (256/32). Each Pipeline stage works with a single micro-batch at a time.
+
+To calculate the global batch size of the DP + PP setup we then do: `mbs*chunks*dp_degree` (`8*32*4=1024`).
+
+Let's go back to the diagram.
+
+With `chunks=1` you end up with the naive MP, which is very inefficient. With a very large `chunks` value you end up with tiny micro-batch sizes which could be not every efficient either. So one has to experiment to find the value that leads to the highest efficient utilization of the gpus.
+
+While the diagram shows that there is a bubble of "dead" time that can't be parallelized because the last `forward` stage has to wait for `backward` to complete the pipeline, the purpose of finding the best value for `chunks` is to enable a high concurrent GPU utilization across all participating GPUs which translates to minimizing the size of the bubble.
+
+Problems:
+- have to modify the model quite heavily, because Pipeline requires one to rewrite the normal flow of modules into a `nn.Sequential` sequence of the same, which may require changes to the design of the model.
+- currently the Pipeline API is very restricted. If you had a bunch of python variables being passed in the very first stage of the Pipeline, you will have to find a way around it. Currently, the pipeline interface requires either a single Tensor or a tuple of Tensors as the only input and output. These tensors must have a batch size as the very first dimension, since pipeline is going to chunk the mini batch into micro-batches. Possible improvements are being discussed here https://github.com/pytorch/pytorch/pull/50693
+- have to arrange each layer so that the output of one model becomes an input to the other model
+
+Implementations:
+- [Pytorch](https://pytorch.org/docs/stable/pipeline.html) (initial support in pytorch-1.8, and progressively getting improved in 1.9 and more so in 1.10). Some [examples](https://github.com/pytorch/pytorch/blob/master/benchmarks/distributed/pipeline/pipe.py)
+- [FairScale](https://fairscale.readthedocs.io/en/latest/tutorials/pipe.html)
+- [DeepSpeed](https://www.deepspeed.ai/tutorials/pipeline/)
+- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM) has an internal implementation - no API.
+
+🤗 Transformers status: as of this writing none of the models supports full-PP. GPT2 and T5 models have naive PP support. The main obstacle is being unable to convert the models to `nn.Sequential` and have all the inputs to be Tensors. This is because currently the models include many features that make the conversion very complicated, and will need to be removed to accomplish that.
+
+Other approaches:
+
+DeepSpeed and SageMaker use the concept of an [Interleaved Pipeline](https://docs.aws.amazon.com/sagemaker/latest/dg/model-parallel-core-features.html)
+![interleaved-pipeline-execution](imgs/parallelism-sagemaker-interleaved-pipeline.png)
+
+Here the bubble (idle time) is further minimized by prioritizing backward passes.
+
+According to [the same document](https://docs.aws.amazon.com/sagemaker/latest/dg/model-parallel-core-features.html), it might be able to automate the non `nn.Sequential` model conversion to pipeline. The only problem is that this is currently only available at AWS, so you can't run it on your own hardware.
+
+
+## Tensor Parallelism
+
+In Tensor Parallelism each GPU processes only a slice of a tensor and only aggregates the full tensor for operations that require the whole thing.
+
+In this section we use concepts and diagrams from the [Megatron-LM](https://github.com/NVIDIA/Megatron-LM) paper: [Efficient Large-Scale Language Model Training on GPU Clusters](https://arxiv.org/abs/2104.04473).
+
+The main building block of any transformer is a fully connected `nn.Linear` followed by a nonlinear activation `GeLU`.
+
+Following the Megatron's paper notation, we can write the dot-product part of it as `Y = GeLU(XA)`, where `X` and `Y` are the input and output vectors, and `A` is the weight matrix.
+
+If we look at the computation in matrix form, it's easy to see how the matrix multiplication can be split between multiple GPUs:
+![Parallel GEMM](imgs/parallelism-tp-parallel_gemm.png)
+
+If we split the weight matrix `A` column-wise across `N` GPUs and perform matrix multiplications `XA_1` through `XA_n` in parallel, then we will end up with `N` output vectors `Y_1, Y_2, ..., Y_n` which can be fed into `GeLU` independently:
+![independent GeLU](imgs/parallelism-tp-independent-gelu.png)
+
+Using this principle, we can update an MLP of arbitrary depth, without the need for any synchronization between GPUs until the very end, where we need to reconstruct the output vector from shards. The Megatron-LM paper authors provide a helpful illustration for that:
+![parallel shard processing](imgs/parallelism-tp-parallel_shard_processing.png)
+
+Parallelizing the multi-headed attention layers is even simpler, since they are already inherently parallel, due to having multiple independent heads!
+![parallel self-attention](imgs/parallelism-tp-parallel_self_attention.png)
+
+Special considerations: TP requires very fast network, and therefore it's not advisable to do TP across more than one node. Practically, if a node has 4 GPUs, the highest TP degree is therefore 4. If you need a TP degree of 8, you need to use nodes that have at least 8 GPUs.
+
+This section is based on the original much more [detailed TP overview](https://github.com/huggingface/transformers/issues/10321#issuecomment-783543530).
+by [@anton-l](https://github.com/anton-l).
+
+Implementations:
+- DeepSpeed calls it [tensor slicing](https://www.deepspeed.ai/features/#model-parallelism)
+- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM) has an internal implementation.
+
+🤗 Transformers status:
+- core: not yet implemented in the core
+- but if you want inference [parallelformers](https://github.com/tunib-ai/parallelformers) provides this support for most of our models. So until this is implemented in the core you can use theirs. And hopefully training mode will be supported too.
+- Deepspeed-Inference also supports our BERT, GPT-2, and GPT-Neo models in their super-fast CUDA-kernel-based inference mode, see more [here](https://www.deepspeed.ai/tutorials/inference-tutorial/)
+
+
+
+## DP+PP
+
+The following diagram from the DeepSpeed [pipeline tutorial](https://www.deepspeed.ai/tutorials/pipeline/) demonstrates how one combines DP with PP.
+
+![dp-pp-2d](imgs/parallelism-zero-dp-pp.png)
+
+Here it's important to see how DP rank 0 doesn't see GPU2 and DP rank 1 doesn't see GPU3. To DP there is just GPUs 0 and 1 where it feeds data as if there were just 2 GPUs. GPU0 "secretly" offloads some of its load to GPU2 using PP. And GPU1 does the same by enlisting GPU3 to its aid.
+
+Since each dimension requires at least 2 GPUs, here you'd need at least 4 GPUs.
+
+Implementations:
+- [DeepSpeed](https://github.com/microsoft/DeepSpeed)
+- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
+
+🤗 Transformers status: not yet implemented
+
+
+## DP+PP+TP
+
+To get an even more efficient training a 3D parallelism is used where PP is combined with TP and DP. This can be seen in the following diagram.
+
+![dp-pp-tp-3d](imgs/parallelism-deepspeed-3d.png)
+
+This diagram is from a blog post [3D parallelism: Scaling to trillion-parameter models](https://www.microsoft.com/en-us/research/blog/deepspeed-extreme-scale-model-training-for-everyone/), which is a good read as well.
+
+Since each dimension requires at least 2 GPUs, here you'd need at least 8 GPUs.
+
+Implementations:
+- [DeepSpeed](https://github.com/microsoft/DeepSpeed) - DeepSpeed also includes an even more efficient DP, which they call ZeRO-DP.
+- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
+
+🤗 Transformers status: not yet implemented, since we have no PP and TP.
+
+
+## DP+PP+TP+ZeRO
+
+One of the main features of DeepSpeed is ZeRO, which is a super-scalable extension of DP. It has already been discussed in [ZeRO Data Parallel](#zero-data-parallel). Normally it's a standalone feature that doesn't require PP or TP. But it can be combined with PP and TP.
+
+When ZeRO-DP is combined with PP (and optinally TP) it typically enables only ZeRO stage 1 (optimizer sharding).
+
+While it's theoretically possible to use ZeRO stage 2 (gradient sharding) with Pipeline Parallelism, it will have bad performance impacts. There would need to be an additional reduce-scatter collective for every micro-batch to aggregate the gradients before sharding, which adds a potentially significant communication overhead. By nature of Pipeline Parallelism, small micro-batches are used and instead the focus is on trying to balance arithmetic intensity (micro-batch size) with minimizing the Pipeline bubble (number of micro-batches). Therefore those communication costs are going to hurt.
+
+In addition, There are already fewer layers than normal due to PP and so the memory savings won't be huge. PP already reduces gradient size by ``1/PP``, and so gradient sharding savings on top of that are less significant than pure DP.
+
+ZeRO stage 3 is not a good choice either for the same reason - more inter-node communications required.
+
+And since we have ZeRO, the other benefit is ZeRO-Offload. Since this is stage 1 optimizer states can be offloaded to CPU.
+
+Implementations:
+- [Megatron-DeepSpeed](https://github.com/microsoft/Megatron-DeepSpeed)
+
+🤗 Transformers status: not yet implemented, since we have no PP and TP.
+
+
+## FlexFlow
+
+[FlexFlow](https://github.com/flexflow/FlexFlow) also solves the parallelization problem in a slightly different approach.
+
+Paper: ["Beyond Data and Model Parallelism for Deep Neural Networks" by Zhihao Jia, Matei Zaharia, Alex Aiken](https://arxiv.org/abs/1807.05358)
+
+It performs a sort of 4D Parallelism over Sample-Operator-Attribute-Parameter.
+
+1. Sample = Data Parallelism
+2. Operator = part vertical Layer Parallelism, but it can split the layer too - more refined level
+3. Attribute = horizontal Model Parallelism (Megatron-LM style)
+4. Parameter = Sharded model params
+
+and they are working on Pipeline Parallelism. I guess ZeRO-DP is Sample+Parameter in this context.
+
+![flex-flow-soap](imgs/parallelism-flexflow.jpeg)
+
+The significance of this framework is that it takes resources like (1) GPU/TPU/CPU vs. (2) RAM/DRAM vs. (3) fast-intra-connect/slow-inter-connect and it automatically optimizes all these  algorithmically deciding which parallelisation to use where.
+
+One very important aspect is that FlexFlow is designed for optimizing DNN parallelizations for models with static and fixed workloads, since models with dynamic behavior may prefer different parallelization strategies across iterations.
+
+So the promise is very attractive - it runs a 30min simulation on the cluster of choice and it comes up with the best strategy to utilise this specific environment. If you add/remove/replace any parts it'll run and re-optimize the plan for that. And then you can train. A different setup will have its own custom optimization.
+
+🤗 Transformers status: not yet integrated. We already have our models FX-trace-able via [transformers.utils.fx](https://github.com/huggingface/transformers/blob/master/src/transformers/utils/fx.py), which is a prerequisite for FlexFlow, so someone needs to figure out what needs to be done to make FlexFlow work with our models.
+
+
+## Which Strategy To Use When
+
+Here is a very rough outlook at which parallelism strategy to use when. The first on the list is typically faster.
+
+**⇨ Single GPU**
+
+* Model fits onto a single GPU:
+
+    1. Normal use
+
+* Model doesn't fit onto a single GPU:
+
+    1. ZeRO + Offload CPU and optionally NVMe
+
+
+**⇨ Single Node / Multi-GPU**
+
+* Model fits onto a single GPU:
+
+    1. DDP - Distributed DP
+    2. ZeRO - may or may not be faster depending on the situation and configuration used
+
+* Model doesn't fit onto a single GPU:
+
+    1. PP
+    2. ZeRO
+    3. TP
+
+    With very fast intra-node connectivity of NVLINK or NVSwitch all three should be mostly on par, without these PP will be faster than TP and ZeRO. The degree of TP may also make a difference. Best to experiment to find the winner on your particular setup.
+
+
+**⇨ Multi-Node / Multi-GPU**
+
+* When you have fast inter-node connectivity:
+
+    1. ZeRO - as it requires close to no modifications to the model
+    2. PP+TP+DP - less communications, but requires massive changes to the model
+
+* when you have slow inter-node connectivity and still low on GPU memory:
+
+    1. DP+PP+TP+ZeRO-1

docs/source/performance.md

@@ -0,0 +1,343 @@
+<!---
+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.
+-->
+
+# Performance and Scalability: How To Fit a Bigger Model and Train It Faster
+
+For now the software sections of this document are mainly Pytorch-specific, but the guide can be extended to other frameworks in the future.
+
+## Quick notes
+
+This section gives brief ideas on how to make training faster and support bigger models. Later sections will expand, demonstrate and elucidate each of these.
+
+### Faster Training
+
+Hardware:
+
+- fast connectivity between GPUs
+    * intra-node: NVLink
+    * inter-node: Infiniband / Intel OPA
+
+Software:
+
+- Data Parallel / Distributed Data Parallel
+- fp16 (autocast caching)
+
+
+### Bigger Models
+
+Hardware:
+
+- bigger GPUs
+- more GPUs
+- more CPU and NVMe (offloaded to by DeepSpeed)
+
+Software:
+
+- Deepspeed ZeRO
+- Deepspeed ZeRO-Offload
+- Megatron-LM 3D Parallelism
+- Pipeline Parallelism
+- Tensor Parallelism
+- Low-memory Optimizers
+- fp16/bf16 (smaller data)
+
+
+
+## Hardware
+
+### Multi-GPU Connectivity
+
+If you use multiple GPUs the way cards are inter-connected can have a huge impact on the total training time.
+
+If the GPUs are on the same physical node, you can run:
+
+```
+nvidia-smi topo -m
+```
+
+and it will tell you how the GPUs are inter-connected.
+
+On a machine with dual-GPU and which are connected with NVLink, you will most likely see something like:
+
+```
+        GPU0    GPU1    CPU Affinity    NUMA Affinity
+GPU0     X      NV2     0-23            N/A
+GPU1    NV2      X      0-23            N/A
+```
+
+on a different machine w/o NVLink we may see:
+```
+        GPU0    GPU1    CPU Affinity    NUMA Affinity
+GPU0     X      PHB     0-11            N/A
+GPU1    PHB      X      0-11            N/A
+```
+
+The report includes this legend:
+
+```
+  X    = Self
+  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
+  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
+  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
+  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
+  PIX  = Connection traversing at most a single PCIe bridge
+  NV#  = Connection traversing a bonded set of # NVLinks
+```
+
+So the first report `NV2` tells us the GPUs are interconnected with 2 NVLinks, and the second report `PHB` we have a typical consumer-level PCIe+Bridge setup.
+
+Check what type of connectivity you have on your setup. Some of these will make the communication between cards faster (e.g. NVLink), others slower (e.g. PHB).
+
+Depending on the type of scalability solution used, the connectivity speed could have a major or a minor impact. If the GPUs need to sync rarely, as in DDP, the impact of a slower connection will be less significant. If the GPUs need to send messages to each other often, as in ZeRO-DP, then faster connectivity becomes super important to achieve faster training.
+
+### NVlink
+
+[NVLink](https://en.wikipedia.org/wiki/NVLink) is a wire-based serial multi-lane near-range communications link developed by Nvidia.
+
+Each new generation provides a faster bandwidth, e.g. here is a quote from [Nvidia Ampere GA102 GPU Architecture](https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/ampere/pdf/NVIDIA-ampere-GA102-GPU-Architecture-Whitepaper-V1.pdf):
+
+> Third-Generation NVLink®
+> GA102 GPUs utilize NVIDIA’s third-generation NVLink interface, which includes four x4 links,
+> with each link providing 14.0625 GB/sec bandwidth in each direction between two GPUs. Four
+> links provide 56.25 GB/sec bandwidth in each direction, and 112.5 GB/sec total bandwidth
+> between two GPUs. Two RTX 3090 GPUs can be connected together for SLI using NVLink.
+> (Note that 3-Way and 4-Way SLI configurations are not supported.)
+
+So the higher `X` you get in the report of `NVX` in the output of `nvidia-smi topo -m` the better. The generation will depend on your GPU architecture.
+
+Let's compare the execution of a gpt2 language model training over a small sample of wikitext.
+
+The results are:
+
+
+| NVlink | Time |
+| -----  | ---: |
+| Y      | 101s |
+| N      | 131s |
+
+
+You can see that NVLink completes the training ~23% faster.
+
+In the second benchmark we use `NCCL_P2P_DISABLE=1` to tell the GPUs not to use NVLink.
+
+Here is the full benchmark code and outputs:
+
+```
+# DDP w/ NVLink
+
+rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch \
+--nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py --model_name_or_path gpt2 \
+--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train \
+--output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
+
+{'train_runtime': 101.9003, 'train_samples_per_second': 1.963, 'epoch': 0.69}
+
+# DDP w/o NVLink
+
+rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 NCCL_P2P_DISABLE=1 python -m torch.distributed.launch \
+--nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py --model_name_or_path gpt2 \
+--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train
+--output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
+
+{'train_runtime': 131.4367, 'train_samples_per_second': 1.522, 'epoch': 0.69}
+```
+
+Hardware: 2x TITAN RTX 24GB each + NVlink with 2 NVLinks (`NV2` in `nvidia-smi topo -m`)
+Software: `pytorch-1.8-to-be` + `cuda-11.0` / `transformers==4.3.0.dev0`
+
+## Software
+
+### Anatomy of Model's Memory
+
+The components on GPU memory are the following:
+- the model weights
+- the forward activations saved for gradient computation
+- the gradients
+- the optimizer state
+
+### `forward` vs `backward` Execution Speed
+
+For convolutions and linear layers there are 2x flops in the backward compared to the forward, which generally translates into ~2x slower (sometimes more, because sizes in the backward tend to be more awkward). Activations are usually bandwidth-limited, and it’s typical for an activation to have to read more data in the backward than in the forward (e.g. activation forward reads once, writes once, activation backward reads twice, gradOutput and output of the forward, and writes once, gradInput).
+
+### fp16
+
+AMP = Automatic Mixed Precision
+
+If we look at what's happening with FP16 training (mixed precision) we have:
+- the model has two copies in memory: one in half-precision for the forward/backward computations and one in full precision - no memory saved here
+- the forward activations saved for gradient computation are in half-precision - memory is saved here
+- the gradients are computed in half-precision *but* converted to full-precision for the update, no saving there
+- the optimizer states are in full precision as all the updates are done in full-precision
+
+So the savings only happen for the forward activations saved for the backward computation, and there is a slight overhead because the model weights are stored both in half- and full-precision.
+
+Now let's look at a simple text-classification fine-tuning on 2 GPUs (I'm giving the command for reference):
+```
+export BS=16
+python -m torch.distributed.launch \
+    --nproc_per_node 2 examples/pytorch/text-classification/run_glue.py \
+    --model_name_or_path bert-base-cased \
+    --task_name mrpc \
+    --do_train \
+    --do_eval \
+    --max_seq_length 128 \
+    --per_device_train_batch_size $BS \
+    --learning_rate 2e-5 \
+    --num_train_epochs 3.0 \
+    --output_dir /tmp/mrpc \
+    --overwrite_output_dir \
+    --fp16
+```
+Since the only savings we get are in the model activations saved for the backward passed, it's logical that the bigger those activations are, the bigger the saving will be. If we try different batch sizes, I indeed get (this is with `nvidia-smi` so not completely reliable as said above but it will be a fair comparison):
+
+| batch size | w/o --fp16 | w/ --fp16 | savings |
+| ---------: | ---------: | --------: | ------: |
+|          8 |       4247 |      4163 |      84 |
+|         16 |       4971 |      4793 |     178 |
+|         32 |       6827 |      6207 |     620 |
+|         64 |      10037 |      8061 |    1976 |
+
+So there is only a real memory saving if we train at a high batch size (and it's not half) and at batch sizes lower than 8, you actually get a bigger memory footprint (because of the overhead mentioned above). The gain for FP16 training is that in each of those cases, the training with the flag `--fp16` is twice as fast, which does require every tensor to have every dimension be a multiple of 8 (examples pad the tensors to a sequence length that is a multiple of 8).
+
+Summary: FP16 with apex or AMP will only give you some memory savings with a reasonably high batch size.
+
+Additionally, under mixed precision when possible, it's important that the batch size is a multiple of 8 to efficiently use tensor cores.
+
+Some amazing tutorials to read on mixed precision:
+- @sgugger wrote a great explanation of mixed precision [here](https://docs.fast.ai/callback.fp16.html#A-little-bit-of-theory)
+- Aleksey Bilogur's [A developer-friendly guide to mixed precision training with PyTorch](https://spell.ml/blog/mixed-precision-training-with-pytorch-Xuk7YBEAACAASJam)
+
+### fp16 caching
+
+pytorch `autocast` which performs AMP include a caching feature, which speed things up by caching fp16-converted values. Here is the full description from this [comment](https://discuss.pytorch.org/t/autocast-and-torch-no-grad-unexpected-behaviour/93475/3):
+
+Autocast maintains a cache of the FP16 casts of model params (leaves). This helps streamline parameter reuse: if the same FP32 param is used in several different FP16list ops, like several matmuls, instead of re-casting the param to FP16 on entering each matmul, the cast will occur on the first matmul, the casted FP16 copy will be cached, and for all later matmuls the FP16 copy will be reused. The cache is maintained only within a particular outermost autocast context. When you exit the autocast context the cache is dropped. For recommended usage, in which autocast wraps the forward pass, and then you exit the context before calling backward(), this means the cache only lasts the duration of the forward pass each iteration, and will be rebuilt next iteration. (The cache of FP16-casted copies MUST be rebuilt each iteration. The FP32 params get updated by the optimizer, so the FP16 copies must be recreated, otherwise the FP16 values will be stale.)
+
+### Batch sizes
+
+One gets the most efficient performance when batch sizes and input/output neuron counts are divisible by a certain number, which typically starts at 8, but can be much higher as well. That number varies a lot depending on the specific hardware being used and the dtype of the model.
+
+For example for fully connected layers (which correspond to GEMMs), NVIDIA provides recommendations for [input/output neuron counts](
+https://docs.nvidia.com/deeplearning/performance/dl-performance-fully-connected/index.html#input-features) and [batch size](https://docs.nvidia.com/deeplearning/performance/dl-performance-fully-connected/index.html#batch-size).
+
+[Tensor Core Requirements](https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html#requirements-tc) define the multiplier based on the dtype and the hardware. For example, for fp16 a multiple of 8 is recommended, but on A100 it's 64!
+
+For parameters that are small, there is also [Dimension Quantization Effects](https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html#dim-quantization) to consider, this is where tiling happens and the right multiplier can have a significant speedup.
+
+
+### DP vs DDP
+
+`DistributedDataParallel` (DDP) is typically faster than `DataParallel` (DP), but it is not always the case:
+* while DP is python threads-based, DDP is multiprocess-based - and as such it has no python threads limitations, such as GIL
+* on the other hand a slow inter-connectivity between the GPU cards could lead to an actual slower outcome with DDP
+
+Here are the main differences in the inter-GPU communication overhead between the two modes:
+
+[DDP](https://pytorch.org/docs/master/notes/ddp.html):
+
+- At the start time the main process replicates the model once from gpu 0 to the rest of gpus
+- Then for each batch:
+   1. each gpu consumes each own mini-batch of data directly
+   2. during `backward`, once the local gradients are ready, they are then averaged across all processes
+
+[DP](https://pytorch.org/docs/master/generated/torch.nn.DataParallel.html):
+
+For each batch:
+   1. gpu 0 reads the batch of data and then sends a mini-batch to each gpu
+   2. replicates the up-to-date model from gpu 0 to each gpu
+   3. runs `forward` and sends output from each gpu to gpu 0, computes loss
+   4. scatters loss from gpu 0 to all gpus, runs `backward`
+   5. sends gradients from each gpu to gpu 0 and averages those
+
+The only communication DDP performs per batch is sending gradients, whereas DP does 5 different data exchanges per batch.
+
+DP copies data within the process via python threads, whereas DDP copies data via [torch.distributed](https://pytorch.org/docs/master/distributed.html).
+
+Under DP gpu 0 performs a lot more work than the rest of the gpus, thus resulting in under-utilization of gpus.
+
+You can use DDP across multiple machines, but this is not the case with DP.
+
+There are other differences between DP and DDP but they aren't relevant to this discussion.
+
+If you want to go really deep into understanding these 2 modes, this [article](https://www.telesens.co/2019/04/04/distributed-data-parallel-training-using-pytorch-on-aws/) is highly recommended, as it has great diagrams, includes multiple benchmarks and profiler outputs on various hardware, explains all the nuances that you may need to know.
+
+Let's look at an actual benchmark:
+
+| Type   | NVlink | Time |
+| :----- | -----  | ---: |
+| 2:DP   | Y      | 110s |
+| 2:DDP  | Y      | 101s |
+| 2:DDP  | N      | 131s |
+
+
+Analysis:
+
+Here DP is ~10% slower than DDP w/ NVlink, but ~15% faster than DDP w/o NVlink
+
+The real difference will depend on how much data each GPU needs to sync with the others - the more there is to sync, the more a slow link will slow down the total runtime.
+
+Here is the full benchmark code and outputs:
+
+`NCCL_P2P_DISABLE=1` was used to disable the NVLink feature on the corresponding benchmark.
+
+```
+
+# DP
+rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 \
+python examples/pytorch/language-modeling/run_clm.py \
+--model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
+--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
+
+{'train_runtime': 110.5948, 'train_samples_per_second': 1.808, 'epoch': 0.69}
+
+# DDP w/ NVlink
+rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 \
+python -m torch.distributed.launch --nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py \
+--model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
+--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
+
+{'train_runtime': 101.9003, 'train_samples_per_second': 1.963, 'epoch': 0.69}
+
+# DDP w/o NVlink
+rm -r /tmp/test-clm; NCCL_P2P_DISABLE=1 CUDA_VISIBLE_DEVICES=0,1 \
+python -m torch.distributed.launch --nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py \
+--model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
+--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
+
+{'train_runtime': 131.4367, 'train_samples_per_second': 1.522, 'epoch': 0.69}
+```
+
+Hardware: 2x TITAN RTX 24GB each + NVlink with 2 NVLinks (`NV2` in `nvidia-smi topo -m`)
+Software: `pytorch-1.8-to-be` + `cuda-11.0` / `transformers==4.3.0.dev0`
+
+
+### DataLoader
+
+One of the important requirements to reach great training speed is the ability to feed the GPU at the maximum speed it can handle. By default everything happens in the main process and it might not be able to read the data from disk fast enough, and thus create a bottleneck, leading to GPU under-utilization.
+
+- `DataLoader(pin_memory=True, ...)` which ensures that the data gets preloaded into the pinned memory on CPU and typically leads to much faster transfers from CPU to GPU memory.
+-  `DataLoader(num_workers=4, ...)` - spawn several workers to pre-load data faster - during training watch the GPU utilization stats and if it's far from 100% experiment with raising the number of workers. Of course, the problem could be elsewhere so a very big number of workers won't necessarily lead to a better performance.
+
+### Faster optimizer
+
+pytorch-nightly introduced `torch.optim._multi_tensor` which should significantly speed up the optimizers for situations with lots of small feature tensors. It should eventually become the default, but if you want to experiment with it sooner and don't mind using the bleed-edge, see: https://github.com/huggingface/transformers/issues/9965
+
+
+## Contribute
+
+This document is far from being complete and a lot more needs to be added, so if you have additions or corrections to make please don't hesitate to open a PR or if you aren't sure start an Issue and we can discuss the details there.
+
+When making contributions that A is better than B, please try to include a reproducible benchmark and/or a link to the source of that information (unless it comes directly from you).

docs/source/preprocessing.rst

@@ -39,6 +39,12 @@ To automatically download the vocab used during pretraining or fine-tuning a giv
 Base use
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/Yffk5aydLzg" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 A :class:`~transformers.PreTrainedTokenizer` has many methods, but the only one you need to remember for preprocessing
 is its ``__call__``: you just need to feed your sentence to your tokenizer object.
 
@@ -138,6 +144,12 @@ can safely ignore it. You can also pass ``verbose=False`` to stop the tokenizer
 Preprocessing pairs of sentences
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/0u3ioSwev3s" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 Sometimes you need to feed a pair of sentences to your model. For instance, if you want to classify if two sentences in
 a pair are similar, or for question-answering models, which take a context and a question. For BERT models, the input
 is then represented like this: :obj:`[CLS] Sequence A [SEP] Sequence B [SEP]`

docs/source/quicktour.rst

@@ -28,8 +28,15 @@ will dig a little bit more and see how the library gives you access to those mod
 Getting started on a task with a pipeline
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-The easiest way to use a pretrained model on a given task is to use :func:`~transformers.pipeline`. 🤗 Transformers
-provides the following tasks out of the box:
+The easiest way to use a pretrained model on a given task is to use :func:`~transformers.pipeline`.
+
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/tiZFewofSLM" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
+🤗 Transformers provides the following tasks out of the box:
 
 - Sentiment analysis: is a text positive or negative?
 - Text generation (in English): provide a prompt and the model will generate what follows.
@@ -137,8 +144,15 @@ to share your fine-tuned model on the hub with the community, using :doc:`this t
 Under the hood: pretrained models
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-Let's now see what happens beneath the hood when using those pipelines. As we saw, the model and tokenizer are created
-using the :obj:`from_pretrained` method:
+Let's now see what happens beneath the hood when using those pipelines.
+
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/AhChOFRegn4" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
+As we saw, the model and tokenizer are created using the :obj:`from_pretrained` method:
 
 .. code-block::
 
@@ -265,8 +279,8 @@ Let's apply the SoftMax activation to get predictions.
 .. code-block::
 
     >>> ## PYTORCH CODE
-    >>> import torch.nn.functional as F
-    >>> pt_predictions = F.softmax(pt_outputs.logits, dim=-1)
+    >>> from torch import nn
+    >>> pt_predictions = nn.functional.softmax(pt_outputs.logits, dim=-1)
     >>> ## TENSORFLOW CODE
     >>> import tensorflow as tf
     >>> tf.nn.softmax(tf_outputs.logits, axis=-1)

docs/source/sagemaker.md

@@ -16,378 +16,10 @@ limitations under the License.
 
 # Run training on Amazon SageMaker
 
-Hugging Face and Amazon are introducing new [Hugging Face Deep Learning Containers (DLCs)](https://github.com/aws/deep-learning-containers/blob/master/available_images.md#huggingface-training-containers) to make it easier than ever to train Hugging Face Transformer models in [Amazon SageMaker](https://aws.amazon.com/sagemaker/).
+The documentation has been moved to [hf.co/docs/sagemaker](https://huggingface.co/docs/sagemaker). This page will be removed in `transformers` 5.0. 
 
-To learn how to access and use the new Hugging Face DLCs with the Amazon SageMaker Python SDK, check out the guides and resources below.
+### Table of Content
 
----
-
-## Deep Learning Container (DLC) overview
-
-The Deep Learning Container are in every available where Amazon SageMaker is available. You can see the [AWS region table](https://aws.amazon.com/about-aws/global-infrastructure/regional-product-services/) for all AWS global infrastructure. To get an detailed overview of all included packages look [here in the release notes](https://docs.aws.amazon.com/deep-learning-containers/latest/devguide/deep-learning-containers-images.html).
-
-| 🤗 Transformers version | 🤗 Datasets version | PyTorch/TensorFlow version | type     | device | Python Version | Example `image_uri`                                                                                                               |
-| ----------------------- | ------------------- | -------------------------- | -------- | ------ | -------------- | --------------------------------------------------------------------------------------------------------------------------------- |
-| 4.4.2                   | 1.5.0               | PyTorch 1.6.0              | training | GPU    | 3.6            | `763104351884.dkr.ecr.us-west-2.amazonaws.com/huggingface-pytorch-training:1.6.0-transformers4.4.2-gpu-py36-cu110-ubuntu18.04`    |
-| 4.4.2                   | 1.5.0               | TensorFlow 2.4.1           | training | GPU    | 3.7            | `763104351884.dkr.ecr.us-west-2.amazonaws.com/huggingface-tensorflow-training:2.4.1-transformers4.4.2-gpu-py37-cu110-ubuntu18.04` |
-
----
-
-## Getting Started: Train a 🤗 Transformers Model
-
-To train a 🤗 Transformers model by using the `HuggingFace` SageMaker Python SDK you need to:
-
-- [Prepare a training script](#prepare-a-transformers-fine-tuning-script)
-- [Create a `HuggingFace` Estimator](#create-an-huggingface-estimator)
-- [Run training by calling the `fit` method](#execute-training)
-- [Access you model](#access-trained-model)
-
-### Setup & Installation
-
-Before you can train a transformers models with Amazon SageMaker you need to sign up for an AWS account. If you do not have an AWS account yet learn more [here](https://docs.aws.amazon.com/sagemaker/latest/dg/gs-set-up.html).
-
-After you complete these tasks you can get started using either [SageMaker Studio](https://docs.aws.amazon.com/sagemaker/latest/dg/gs-studio-onboard.html), [SageMaker Notebook Instances](https://docs.aws.amazon.com/sagemaker/latest/dg/gs-console.html), or a local environment. To start training locally you need configure the right [IAM permission](https://docs.aws.amazon.com/sagemaker/latest/dg/sagemaker-roles.html).
-
-Upgrade to the latest `sagemaker` version.
-
-```bash
-pip install sagemaker --upgrade
-```
-
-**SageMaker environment**
-
-_Note: The execution role is intended to be available only when running a notebook within SageMaker. If you run `get_execution_role` in a notebook not on SageMaker, expect a "region" error._
-
-```python
-import sagemaker
-sess = sagemaker.Session()
-role = sagemaker.get_execution_role()
-```
-
-**Local environment**
-
-```python
-import sagemaker
-import boto3
-
-iam_client = boto3.client('iam')
-role = iam_client.get_role(RoleName='role-name-of-your-iam-role-with-right-permissions')['Role']['Arn']
-sess = sagemaker.Session()
-```
-
-### Prepare a 🤗 Transformers fine-tuning script.
-
-The training script is very similar to a training script you might run outside of SageMaker, but you can access useful properties about the training environment through various environment variables, including the following:
-
-- `SM_MODEL_DIR`: A string that represents the path where the training job writes the model artifacts to. After training, artifacts in this directory are uploaded to S3 for model hosting. `SM_MODEL_DIR` is always set to `/opt/ml/model`.
-
-- `SM_NUM_GPUS`: An integer representing the number of GPUs available to the host.
-
-- `SM_CHANNEL_XXXX:` A string that represents the path to the directory that contains the input data for the specified channel. For example, if you specify two input channels in the HuggingFace estimator’s fit call, named `train` and `test`, the environment variables `SM_CHANNEL_TRAIN` and `SM_CHANNEL_TEST` are set.
-
-You can find a full list of the exposed environment variables [here](https://github.com/aws/sagemaker-training-toolkit/blob/master/ENVIRONMENT_VARIABLES.md).
-
-Later we define `hyperparameters` in the [HuggingFace Estimator](#create-an-huggingface-estimator), which are passed in as named arguments and and can be processed with the `ArgumentParser()`.
-
-```python
-import transformers
-import datasets
-import argparse
-import os
-
-if __name__ == "__main__":
-
-    parser = argparse.ArgumentParser()
-
-    # hyperparameters sent by the client are passed as command-line arguments to the script.
-    parser.add_argument("--epochs", type=int, default=3)
-    parser.add_argument("--per_device_train_batch_size", type=int, default=32)
-    parser.add_argument("--model_name_or_path", type=str)
-
-    # Data, model, and output directories
-    parser.add_argument("--model-dir", type=str, default=os.environ["SM_MODEL_DIR"])
-    parser.add_argument("--training_dir", type=str, default=os.environ["SM_CHANNEL_TRAIN"])
-    parser.add_argument("--test_dir", type=str, default=os.environ["SM_CHANNEL_TEST"])
-```
-
-_Note that SageMaker doesn’t support argparse actions. For example, if you want to use a boolean hyperparameter, specify `type` as `bool` in your script and provide an explicit `True` or `False` value._
-
-For a complete example of a 🤗 Transformers training script, see [train.py](https://github.com/huggingface/notebooks/blob/master/sagemaker/01_getting_started_pytorch/scripts/train.py)
-
-### Create an HuggingFace Estimator
-
-You run 🤗 Transformers training scripts on SageMaker by creating `HuggingFace` Estimators. The Estimator handles end-to-end Amazon SageMaker training. The training of your script is invoked when you call `fit` on a `HuggingFace` Estimator. In the Estimator you define, which fine-tuning script should be used as `entry_point`, which `instance_type` should be used, which `hyperparameters` are passed in, you can find all possible `HuggingFace` Parameter [here](https://sagemaker.readthedocs.io/en/stable/frameworks/huggingface/sagemaker.huggingface.html#huggingface-estimator). and an example of a fine-tuning script [here](https://github.com/huggingface/notebooks/blob/master/sagemaker/01_getting_started_pytorch/scripts/train.py).
-You can find all useable `instance_types` [here](https://aws.amazon.com/de/sagemaker/pricing/).
-
-The following code sample shows how you train a custom `HuggingFace` script `train.py`, passing in three hyperparameters (`epochs`, `per_device_train_batch_size`, and `model_name_or_path`).
-
-```python
-from sagemaker.huggingface import HuggingFace
-
-
-# hyperparameters, which are passed into the training job
-hyperparameters={'epochs': 1,
-                 'per_device_train_batch_size': 32,
-                 'model_name_or_path': 'distilbert-base-uncased'
-                 }
-
-# create the Estimator
-huggingface_estimator = HuggingFace(
-        entry_point='train.py',
-        source_dir='./scripts',
-        instance_type='ml.p3.2xlarge',
-        instance_count=1,
-        role=role,
-        transformers_version='4.4',
-        pytorch_version='1.6',
-        py_version='py36',
-        hyperparameters = hyperparameters
-)
-```
-
-To run the `TrainingJob` locally you can define `instance_type='local'` or `instance_type='local-gpu'` for gpu usage. _Note: this does not working within SageMaker Studio_
-
-### Execute Training
-
-You start your `TrainingJob` by calling `fit` on a `HuggingFace` Estimator. In the `fit` method you specify your input training data, like a string S3 URI `s3://my-bucket/my-training-data` or a `FileSystemInput` for [EFS or FSx Lustre](https://sagemaker.readthedocs.io/en/stable/overview.html?highlight=FileSystemInput#use-file-systems-as-training-inputs), see [here](https://sagemaker.readthedocs.io/en/stable/overview.html?highlight=FileSystemInput#use-file-systems-as-training-inputs).
-
-```python
-huggingface_estimator.fit(
-  {'train': 's3://sagemaker-us-east-1-558105141721/samples/datasets/imdb/train',
-   'test': 's3://sagemaker-us-east-1-558105141721/samples/datasets/imdb/test'}
-)
-
-```
-
-SageMaker takes care of starting and managing all the required ec2 instances for ands starts the training job by running.
-
-```bash
-/opt/conda/bin/python train.py --epochs 1 --model_name_or_path distilbert-base-uncased --per_device_train_batch_size 32
-```
-
-### Access trained model
-
-After training is done you can access your model either through the [AWS console](https://console.aws.amazon.com/console/home?nc2=h_ct&src=header-signin) or downloading it directly from S3.
-
-```python
-from sagemaker.s3 import S3Downloader
-
-S3Downloader.download(
-    s3_uri=huggingface_estimator.model_data, # s3 uri where the trained model is located
-    local_path='.', # local path where *.targ.gz is saved
-    sagemaker_session=sess # sagemaker session used for training the model
-)
-```
-
----
-
-## Sample Notebooks
-
-You can find here a list of the official notebooks provided by Hugging Face.
-
-| Notebook                                                                                                                                                                                        | Description                                                                                                      |
-| ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------- |
-| [Getting Started Pytorch](https://github.com/huggingface/notebooks/blob/master/sagemaker/01_getting_started_pytorch/sagemaker-notebook.ipynb)                                                   | End-to-End binary Text-Classification example using `Trainer` and `imdb` dataset                                 |
-| [Getting Started Tensorflow](https://github.com/huggingface/notebooks/blob/master/sagemaker/02_getting_started_tensorflow/sagemaker-notebook.ipynb)                                             | End-to-End binary Text-Classification example using `Keras` and `imdb` dataset                                   |
-| [Distributed Training Data Parallelism](https://github.com/huggingface/notebooks/blob/master/sagemaker/03_distributed_training_data_parallelism/sagemaker-notebook.ipynb)                       | End-to-End distributed Question-Answering example using `Trainer` and 🤗 Transformers example script for `SQAuD` |
-| [Distributed Training Model Parallelism](https://github.com/huggingface/notebooks/blob/master/sagemaker/04_distributed_training_model_parallelism/sagemaker-notebook.ipynb)                     | End-to-End model parallelism example using `SageMakerTrainer` and `run_glue.py` script                           |
-| [Spot Instances and continues training](https://github.com/huggingface/notebooks/blob/master/sagemaker/05_spot_instances/sagemaker-notebook.ipynb)                                              | End-to-End to Text-Classification example using spot instances with continued training.                          |
-| [SageMaker Metrics](https://github.com/huggingface/notebooks/blob/master/sagemaker/06_sagemaker_metrics/sagemaker-notebook.ipynb)                                                               | End-to-End to Text-Classification example using SageMaker Metrics to extract and log metrics during training     |
-| [Distributed Training Data Parallelism Tensorflow](https://github.com/huggingface/notebooks/blob/master/sagemaker/07_tensorflow_distributed_training_data_parallelism/sagemaker-notebook.ipynb) | End-to-End distributed binary Text-Classification example using `Keras` and `TensorFlow`                    
-| [Distributed Seq2Seq Training with Data Parallelism and BART](https://github.com/huggingface/notebooks/blob/master/sagemaker/08_distributed_summarization_bart_t5/sagemaker-notebook.ipynb) | End-to-End distributed summarization example `BART-large` and 🤗 Transformers example script for `summarization`                        |
-
-
----
-
-## Advanced Features
-
-In addition to the Deep Learning Container and the SageMaker SDK, we have implemented other additional features.
-
-### Distributed Training: Data-Parallel
-
-You can use [SageMaker Data Parallelism Library](https://aws.amazon.com/blogs/aws/managed-data-parallelism-in-amazon-sagemaker-simplifies-training-on-large-datasets/) out of the box for distributed training. We added the functionality of Data Parallelism directly into the [Trainer](https://huggingface.co/transformers/main_classes/trainer.html). If your `train.py` uses the [Trainer](https://huggingface.co/transformers/main_classes/trainer.html) API you only need to define the distribution parameter in the HuggingFace Estimator.
-
-- [Example Notebook PyTorch](https://github.com/huggingface/notebooks/blob/master/sagemaker/04_distributed_training_model_parallelism/sagemaker-notebook.ipynb)
-- [Example Notebook TensorFlow](https://github.com/huggingface/notebooks/blob/master/sagemaker/07_tensorflow_distributed_training_data_parallelism/sagemaker-notebook.ipynb)
-
-```python
-# configuration for running training on smdistributed Data Parallel
-distribution = {'smdistributed':{'dataparallel':{ 'enabled': True }}}
-
-# create the Estimator
-huggingface_estimator = HuggingFace(
-        entry_point='train.py',
-        source_dir='./scripts',
-        instance_type='ml.p3dn.24xlarge',
-        instance_count=2,
-        role=role,
-        transformers_version='4.4.2',
-        pytorch_version='1.6.0',
-        py_version='py36',
-        hyperparameters = hyperparameters
-        distribution = distribution
-)
-
-```
-
-### Distributed Training: Model-Parallel
-
-You can use [SageMaker Model Parallelism Library](https://aws.amazon.com/blogs/aws/amazon-sagemaker-simplifies-training-deep-learning-models-with-billions-of-parameters/) out of the box for distributed training. We added the functionality of Model Parallelism directly into the [Trainer](https://huggingface.co/transformers/main_classes/trainer.html). If your `train.py` uses the [Trainer](https://huggingface.co/transformers/main_classes/trainer.html) API you only need to define the distribution parameter in the HuggingFace Estimator.  
-For detailed information about the adjustments take a look [here](https://sagemaker.readthedocs.io/en/stable/api/training/smd_model_parallel_general.html?highlight=modelparallel#required-sagemaker-python-sdk-parameters).
-
-
-- [Example Notebook](https://github.com/huggingface/notebooks/blob/master/sagemaker/04_distributed_training_model_parallelism/sagemaker-notebook.ipynb)
-
-
-```python
-# configuration for running training on smdistributed Model Parallel
-mpi_options = {
-    "enabled" : True,
-    "processes_per_host" : 8
-}
-
-smp_options = {
-    "enabled":True,
-    "parameters": {
-        "microbatches": 4,
-        "placement_strategy": "spread",
-        "pipeline": "interleaved",
-        "optimize": "speed",
-        "partitions": 4,
-        "ddp": True,
-    }
-}
-
-distribution={
-    "smdistributed": {"modelparallel": smp_options},
-    "mpi": mpi_options
-}
-
- # create the Estimator
-huggingface_estimator = HuggingFace(
-        entry_point='train.py',
-        source_dir='./scripts',
-        instance_type='ml.p3dn.24xlarge',
-        instance_count=2,
-        role=role,
-        transformers_version='4.4.2',
-        pytorch_version='1.6.0',
-        py_version='py36',
-        hyperparameters = hyperparameters,
-        distribution = distribution
-)
-```
-
-### Spot Instances
-
-With the creation of HuggingFace Framework extension for the SageMaker Python SDK we can also leverage the benefit of [fully-managed EC2 spot instances](https://docs.aws.amazon.com/sagemaker/latest/dg/model-managed-spot-training.html) and save up to 90% of our training cost.
-
-_Note: Unless your training job completes quickly, we recommend you use [checkpointing](https://docs.aws.amazon.com/sagemaker/latest/dg/model-checkpoints.html) with managed spot training, therefore you need to define the `checkpoint_s3_uri`._
-
-To use spot instances with the `HuggingFace` Estimator we have to set the `use_spot_instances` parameter to `True` and define your `max_wait` and `max_run` time. You can read more about the [managed spot training lifecycle here](https://docs.aws.amazon.com/sagemaker/latest/dg/model-managed-spot-training.html).
-
-- [Example Notebook](https://github.com/huggingface/notebooks/blob/master/sagemaker/05_spot_instances/sagemaker-notebook.ipynb)
-
-```python
-# hyperparameters, which are passed into the training job
-hyperparameters={'epochs': 1,
-                 'train_batch_size': 32,
-                 'model_name':'distilbert-base-uncased',
-                 'output_dir':'/opt/ml/checkpoints'
-                 }
-# create the Estimator
-
-huggingface_estimator = HuggingFace(
-        entry_point='train.py',
-        source_dir='./scripts',
-        instance_type='ml.p3.2xlarge',
-        instance_count=1,
-	    checkpoint_s3_uri=f's3://{sess.default_bucket()}/checkpoints'
-        use_spot_instances=True,
-        max_wait=3600, # This should be equal to or greater than max_run in seconds'
-        max_run=1000,
-        role=role,
-        transformers_version='4.4',
-        pytorch_version='1.6',
-        py_version='py36',
-        hyperparameters = hyperparameters
-)
-
-# Training seconds: 874
-# Billable seconds: 262
-# Managed Spot Training savings: 70.0%
-
-```
-
-### Git Repository
-
-When you create a `HuggingFace` Estimator, you can specify a [training script that is stored in a GitHub repository](https://sagemaker.readthedocs.io/en/stable/overview.html#use-scripts-stored-in-a-git-repository) as the entry point for the estimator, so that you don’t have to download the scripts locally. If Git support is enabled, the `entry_point` and `source_dir` should be relative paths in the Git repo if provided. 
-
-If you are using `git_config` to run the [🤗 Transformers examples scripts](https://github.com/huggingface/transformers/tree/master/examples) keep in mind that you need to configure the right `'branch'` for you `transformers_version`, e.g. if you use `transformers_version='4.4.2` you have to use `'branch':'v4.4.2'`. 
-
-As an example to use `git_config` with an [example script from the transformers repository](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification).
-
-_Tip: define `output_dir` as `/opt/ml/model` in the hyperparameter for the script to save your model to S3 after training._
-
-- [Example Notebook](https://github.com/huggingface/notebooks/blob/master/sagemaker/02_getting_started_tensorflow/sagemaker-notebook.ipynb)
-
-```python
-# configure git settings
-git_config = {'repo': 'https://github.com/huggingface/transformers.git','branch': 'v4.4.2'} # v4.4.2 is referring to the `transformers_version you use in the estimator.
-
- # create the Estimator
-huggingface_estimator = HuggingFace(
-        entry_point='run_glue.py',
-        source_dir='./examples/pytorch/text-classification',
-        git_config=git_config,
-        instance_type='ml.p3.2xlarge',
-        instance_count=1,
-        role=role,
-        transformers_version='4.4',
-        pytorch_version='1.6',
-        py_version='py36',
-        hyperparameters=hyperparameters
-)
-
-```
-
-### SageMaker Metrics
-
-[SageMaker Metrics](https://docs.aws.amazon.com/sagemaker/latest/dg/training-metrics.html#define-train-metrics) can automatically parse the logs for metrics and send those metrics to CloudWatch. If you want SageMaker to parse logs you have to specify the metrics that you want SageMaker to send to CloudWatch when you configure the training job. You specify the name of the metrics that you want to send and the regular expressions that SageMaker uses to parse the logs that your algorithm emits to find those metrics.
-
-- [Example Notebook](https://github.com/huggingface/notebooks/blob/master/sagemaker/06_sagemaker_metrics/sagemaker-notebook.ipynb)
-
-```python
-# define metrics definitions
-
-metric_definitions = [
-{"Name": "train_runtime", "Regex": "train_runtime.*=\D*(.*?)$"},
-{"Name": "eval_accuracy", "Regex": "eval_accuracy.*=\D*(.*?)$"},
-{"Name": "eval_loss", "Regex": "eval_loss.*=\D*(.*?)$"},
-]
-
-# create the Estimator
-
-huggingface_estimator = HuggingFace(
-        entry_point='train.py',
-        source_dir='./scripts',
-        instance_type='ml.p3.2xlarge',
-        instance_count=1,
-        role=role,
-        transformers_version='4.4',
-        pytorch_version='1.6',
-        py_version='py36',
-        metric_definitions=metric_definitions,
-        hyperparameters = hyperparameters)
-
-```
-
-## Additional Resources
-
-- [Announcement Blog Post](https://huggingface.co/blog/the-partnership-amazon-sagemaker-and-hugging-face)
-
-- [AWS and Hugging Face collaborate to simplify and accelerate adoption of natural language processing](https://aws.amazon.com/blogs/machine-learning/aws-and-hugging-face-collaborate-to-simplify-and-accelerate-adoption-of-natural-language-processing-models/)
-
-- [Amazon SageMaker documentation for Hugging Face](https://docs.aws.amazon.com/sagemaker/latest/dg/hugging-face.html)
-
-- [SageMaker Python SDK](https://sagemaker.readthedocs.io/en/stable/frameworks/huggingface/index.html)
+- [Train Hugging Face models on Amazon SageMaker with the SageMaker Python SDK](https://huggingface.co/docs/sagemaker/train)
+- [Deploy Hugging Face models to Amazon SageMaker with the SageMaker Python SDK](https://huggingface.co/docs/sagemaker/inference)
+- [Frequently Asked Questions](https://huggingface.co/docs/sagemaker/faq)

docs/source/serialization.rst

@@ -21,11 +21,165 @@ Projects `ONNX (Open Neural Network eXchange) <http://onnx.ai>`_ and `ONNXRuntim
 unified and community-driven format to store and, by extension, efficiently execute neural network leveraging a variety
 of hardware and dedicated optimizations.
 
+
 Starting from transformers v2.10.0 we partnered with ONNX Runtime to provide an easy export of transformers models to
 the ONNX format. You can have a look at the effort by looking at our joint blog post `Accelerate your NLP pipelines
 using Hugging Face Transformers and ONNX Runtime
 <https://medium.com/microsoftazure/accelerate-your-nlp-pipelines-using-hugging-face-transformers-and-onnx-runtime-2443578f4333>`_.
 
+
+Configuration-based approach
+-----------------------------------------------------------------------------------------------------------------------
+
+Transformers v4.9.0 introduces a new package: ``transformers.onnx``. This package allows converting checkpoints to an
+ONNX graph by leveraging configuration objects. These configuration objects come ready made for a number of model
+architectures, and are made to be easily extendable to other architectures.
+
+Ready-made configurations include the following models:
+
+- ALBERT
+- BART
+- BERT
+- DistilBERT
+- GPT-2
+- RoBERTa
+- T5
+- XLM-RoBERTa
+
+This conversion is handled with the PyTorch version of models - it, therefore, requires PyTorch to be installed. If you
+would like to be able to convert from TensorFlow, please let us know by opening an issue.
+
+.. note::
+    The models showcased here are close to fully feature complete, but do lack some features that are currently in
+    development. Namely, the ability to handle the past key values for decoder models is currently in the works.
+
+
+Converting an ONNX model using the ``transformers.onnx`` package
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+The package may be used as a Python module:
+
+.. code-block::
+
+    python -m transformers.onnx --help
+
+    usage: Hugging Face ONNX Exporter tool [-h] -m MODEL -f {pytorch} [--features {default}] [--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's name of path on disk to load.
+      --features {default}  Export the model with some additional features.
+      --opset OPSET         ONNX opset version to export the model with (default 12).
+      --atol ATOL           Absolute difference tolerance when validating the model.
+
+Exporting a checkpoint using a ready-made configuration can be done as follows:
+
+.. code-block::
+
+    python -m transformers.onnx --model=bert-base-cased onnx/bert-base-cased/
+
+This exports an ONNX graph of the mentioned checkpoint. Here it is `bert-base-cased`, but it can be any model from the
+hub, or a local path.
+
+It will be exported under ``onnx/bert-base-cased``. You should see similar logs:
+
+.. code-block::
+
+    Validating ONNX model...
+            -[✓] ONNX model outputs' name match reference model ({'pooler_output', 'last_hidden_state'}
+            - Validating ONNX Model output "last_hidden_state":
+                    -[✓] (2, 8, 768) matchs (2, 8, 768)
+                    -[✓] all values close (atol: 0.0001)
+            - Validating ONNX Model output "pooler_output":
+                    -[✓] (2, 768) matchs (2, 768)
+                    -[✓] all values close (atol: 0.0001)
+    All good, model saved at: onnx/bert-base-cased/model.onnx
+
+This export can now be used in the ONNX inference runtime:
+
+.. code-block::
+
+    import onnxruntime as ort
+
+    from transformers import BertTokenizerFast
+    tokenizer = BertTokenizerFast.from_pretrained("bert-base-cased")
+
+    ort_session = ort.InferenceSession("onnx/bert-base-cased/model.onnx")
+
+    inputs = tokenizer("Using BERT in ONNX!", return_tensors="np")
+    outputs = ort_session.run(["last_hidden_state", "pooler_output"], dict(inputs))
+
+The outputs used (:obj:`["last_hidden_state", "pooler_output"]`) can be obtained by taking a look at the ONNX
+configuration of each model. For example, for BERT:
+
+.. code-block::
+
+    from transformers.models.bert import BertOnnxConfig, BertConfig
+
+    config = BertConfig()
+    onnx_config = BertOnnxConfig(config)
+    output_keys = list(onnx_config.outputs.keys())
+
+Implementing a custom configuration for an unsupported architecture
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Let's take a look at the changes necessary to add a custom configuration for an unsupported architecture. Firstly, we
+will need a custom ONNX configuration object that details the model inputs and outputs. The BERT ONNX configuration is
+visible below:
+
+.. code-block::
+
+    class BertOnnxConfig(OnnxConfig):
+        @property
+        def inputs(self) -> Mapping[str, Mapping[int, str]]:
+            return OrderedDict(
+                [
+                    ("input_ids", {0: "batch", 1: "sequence"}),
+                    ("attention_mask", {0: "batch", 1: "sequence"}),
+                    ("token_type_ids", {0: "batch", 1: "sequence"}),
+                ]
+            )
+
+        @property
+        def outputs(self) -> Mapping[str, Mapping[int, str]]:
+            return OrderedDict([("last_hidden_state", {0: "batch", 1: "sequence"}), ("pooler_output", {0: "batch"})])
+
+Let's understand what's happening here. This configuration has two properties: the inputs, and the outputs.
+
+The inputs return a dictionary, where each key corresponds to an expected input, and each value indicates the axis of
+that input.
+
+For BERT, there are three necessary inputs. These three inputs are of similar shape, which is made up of two
+dimensions: the batch is the first dimension, and the second is the sequence.
+
+The outputs return a similar dictionary, where, once again, each key corresponds to an expected output, and each value
+indicates the axis of that output.
+
+Once this is done, a single step remains: adding this configuration object to the initialisation of the model class,
+and to the general ``transformers`` initialisation.
+
+An important fact to notice is the use of `OrderedDict` in both inputs and outputs properties. This is a requirements
+as inputs are matched against their relative position within the `PreTrainedModel.forward()` prototype and outputs are
+match against there position in the returned `BaseModelOutputX` instance.
+
+An example of such an addition is visible here, for the MBart model: `Making MBART ONNX-convertible
+<https://github.com/huggingface/transformers/pull/13049/commits/d097adcebd89a520f04352eb215a85916934204f>`__
+
+If you would like to contribute your addition to the library, we recommend you implement tests. An example of such
+tests is visible here: `Adding tests to the MBART ONNX conversion
+<https://github.com/huggingface/transformers/pull/13049/commits/5d642f65abf45ceeb72bd855ca7bfe2506a58e6a>`__
+
+Graph conversion
+-----------------------------------------------------------------------------------------------------------------------
+
+.. note::
+    The approach detailed here is bing deprecated. We recommend you follow the part above for an up to date approach.
+
+
 Exporting a model is done through the script `convert_graph_to_onnx.py` at the root of the transformers sources. The
 following command shows how easy it is to export a BERT model from the library, simply run:
 

docs/source/task_summary.rst

@@ -1,4 +1,4 @@
-.. 
+..
     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
@@ -69,13 +69,13 @@ This returns a label ("POSITIVE" or "NEGATIVE") alongside a score, as follows:
 
     >>> from transformers import pipeline
 
-    >>> nlp = pipeline("sentiment-analysis")
+    >>> classifier = pipeline("sentiment-analysis")
 
-    >>> result = nlp("I hate you")[0]
+    >>> result = classifier("I hate you")[0]
     >>> print(f"label: {result['label']}, with score: {round(result['score'], 4)}")
     label: NEGATIVE, with score: 0.9991
 
-    >>> result = nlp("I love you")[0]
+    >>> result = classifier("I love you")[0]
     >>> print(f"label: {result['label']}, with score: {round(result['score'], 4)}")
     label: POSITIVE, with score: 0.9999
 
@@ -182,7 +182,7 @@ leverages a fine-tuned model on SQuAD.
 
     >>> from transformers import pipeline
 
-    >>> nlp = pipeline("question-answering")
+    >>> question_answerer = pipeline("question-answering")
 
     >>> context = r"""
     ... Extractive Question Answering is the task of extracting an answer from a text given a question. An example of a
@@ -195,11 +195,11 @@ positions of the extracted answer in the text.
 
 .. code-block::
 
-    >>> result = nlp(question="What is extractive question answering?", context=context)
+    >>> result = question_answerer(question="What is extractive question answering?", context=context)
     >>> print(f"Answer: '{result['answer']}', score: {round(result['score'], 4)}, start: {result['start']}, end: {result['end']}")
     Answer: 'the task of extracting an answer from a text given a question.', score: 0.6226, start: 34, end: 96
 
-    >>> result = nlp(question="What is a good example of a question answering dataset?", context=context)
+    >>> result = question_answerer(question="What is a good example of a question answering dataset?", context=context)
     >>> print(f"Answer: '{result['answer']}', score: {round(result['score'], 4)}, start: {result['start']}, end: {result['end']}")
     Answer: 'SQuAD dataset,', score: 0.5053, start: 147, end: 161
 
@@ -336,14 +336,14 @@ Here is an example of using pipelines to replace a mask from a sequence:
 
     >>> from transformers import pipeline
 
-    >>> nlp = pipeline("fill-mask")
+    >>> unmasker = pipeline("fill-mask")
 
 This outputs the sequences with the mask filled, the confidence score, and the token id in the tokenizer vocabulary:
 
 .. code-block::
 
     >>> from pprint import pprint
-    >>> pprint(nlp(f"HuggingFace is creating a {nlp.tokenizer.mask_token} that the community uses to solve NLP tasks."))
+    >>> pprint(unmasker(f"HuggingFace is creating a {unmasker.tokenizer.mask_token} that the community uses to solve NLP tasks."))
     [{'score': 0.1792745739221573,
       'sequence': '<s>HuggingFace is creating a tool that the community uses to '
                   'solve NLP tasks.</s>',
@@ -451,7 +451,7 @@ of tokens.
     >>> ## PYTORCH CODE
     >>> from transformers import AutoModelWithLMHead, AutoTokenizer, top_k_top_p_filtering
     >>> import torch
-    >>> from torch.nn import functional as F
+    >>> from torch import nn
 
     >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
     >>> model = AutoModelWithLMHead.from_pretrained("gpt2")
@@ -467,7 +467,7 @@ of tokens.
     >>> filtered_next_token_logits = top_k_top_p_filtering(next_token_logits, top_k=50, top_p=1.0)
 
     >>> # sample
-    >>> probs = F.softmax(filtered_next_token_logits, dim=-1)
+    >>> probs = nn.functional.softmax(filtered_next_token_logits, dim=-1)
     >>> next_token = torch.multinomial(probs, num_samples=1)
 
     >>> generated = torch.cat([input_ids, next_token], dim=-1)
@@ -505,8 +505,8 @@ This outputs a (hopefully) coherent next token following the original sequence,
     >>> print(resulting_string)
     Hugging Face is based in DUMBO, New York City, and has
 
-In the next section, we show how :func:`~transformers.PreTrainedModel.generate` can be used to generate multiple tokens
-up to a specified length instead of one token at a time.
+In the next section, we show how :func:`~transformers.generation_utils.GenerationMixin.generate` can be used to
+generate multiple tokens up to a specified length instead of one token at a time.
 
 Text Generation
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -627,9 +627,9 @@ It leverages a fine-tuned model on CoNLL-2003, fine-tuned by `@stefan-it <https:
 
     >>> from transformers import pipeline
 
-    >>> nlp = pipeline("ner")
+    >>> ner_pipe = pipeline("ner")
 
-    >>> sequence = """Hugging Face Inc. is a company based in New York City. Its headquarters are in DUMBO, 
+    >>> sequence = """Hugging Face Inc. is a company based in New York City. Its headquarters are in DUMBO,
     ... therefore very close to the Manhattan Bridge which is visible from the window."""
 
 
@@ -638,7 +638,7 @@ Here are the expected results:
 
 .. code-block::
 
-    >>> print(nlp(sequence))
+    >>> print(ner_pipe(sequence))
     [
         {'word': 'Hu', 'score': 0.9995632767677307, 'entity': 'I-ORG'},
         {'word': '##gging', 'score': 0.9915938973426819, 'entity': 'I-ORG'},
@@ -827,18 +827,18 @@ CNN / Daily Mail), it yields very good results.
 .. code-block::
 
     >>> ## PYTORCH CODE
-    >>> from transformers import AutoModelWithLMHead, AutoTokenizer
+    >>> from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
 
-    >>> model = AutoModelWithLMHead.from_pretrained("t5-base")
+    >>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")
     >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
 
     >>> # T5 uses a max_length of 512 so we cut the article to 512 tokens.
-    >>> inputs = tokenizer.encode("summarize: " + ARTICLE, return_tensors="pt", max_length=512)
+    >>> inputs = tokenizer.encode("summarize: " + ARTICLE, return_tensors="pt", max_length=512, truncation=True)
     >>> outputs = model.generate(inputs, max_length=150, min_length=40, length_penalty=2.0, num_beams=4, early_stopping=True)
     >>> ## TENSORFLOW CODE
-    >>> from transformers import TFAutoModelWithLMHead, AutoTokenizer
+    >>> from transformers import TFAutoModelForSeq2SeqLM, AutoTokenizer
 
-    >>> model = TFAutoModelWithLMHead.from_pretrained("t5-base")
+    >>> model = TFAutoModelForSeq2SeqLM.from_pretrained("t5-base")
     >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
 
     >>> # T5 uses a max_length of 512 so we cut the article to 512 tokens.

docs/source/testing.rst

@@ -431,6 +431,7 @@ decorators are used to set the requirements of tests CPU/GPU/TPU-wise:
 * ``require_torch_gpu`` - as ``require_torch`` plus requires at least 1 GPU
 * ``require_torch_multi_gpu`` - as ``require_torch`` plus requires at least 2 GPUs
 * ``require_torch_non_multi_gpu`` - as ``require_torch`` plus requires 0 or 1 GPUs
+* ``require_torch_up_to_2_gpus`` - as ``require_torch`` plus requires 0 or 1 or 2 GPUs
 * ``require_torch_tpu`` - as ``require_torch`` plus requires at least 1 TPU
 
 Let's depict the GPU requirements in the following table:
@@ -447,6 +448,8 @@ Let's depict the GPU requirements in the following table:
 +----------+----------------------------------+
 | ``< 2``  | ``@require_torch_non_multi_gpu`` |
 +----------+----------------------------------+
+| ``< 3``  | ``@require_torch_up_to_2_gpus``  |
++----------+----------------------------------+
 
 
 For example, here is a test that must be run only when there are 2 or more GPUs available and pytorch is installed:
@@ -1167,6 +1170,23 @@ To start a debugger at the point of the warning, do this:
     pytest tests/test_logging.py -W error::UserWarning --pdb
 
 
+Working with github actions workflows
+-----------------------------------------------------------------------------------------------------------------------
+
+To trigger a self-push workflow CI job, you must:
+
+1. Create a new branch on ``transformers`` origin (not a fork!).
+2. The branch name has to start with either ``ci_`` or ``ci-`` (``master`` triggers it too, but we can't do PRs on
+   ``master``). It also gets triggered only for specific paths - you can find the up-to-date definition in case it
+   changed since this document has been written `here
+   <https://github.com/huggingface/transformers/blob/master/.github/workflows/self-push.yml>`__ under `push:`
+3. Create a PR from this branch.
+4. Then you can see the job appear `here
+   <https://github.com/huggingface/transformers/actions/workflows/self-push.yml>`__. It may not run right away if there
+   is a backlog.
+
+
+
 
 Testing Experimental CI Features
 -----------------------------------------------------------------------------------------------------------------------

docs/source/tokenizer_summary.rst

@@ -13,12 +13,20 @@
 Summary of the tokenizers
 -----------------------------------------------------------------------------------------------------------------------
 
-On this page, we will have a closer look at tokenization. As we saw in :doc:`the preprocessing tutorial
-<preprocessing>`, tokenizing a text is splitting it into words or subwords, which then are converted to ids through a
-look-up table. Converting words or subwords to ids is straightforward, so in this summary, we will focus on splitting a
-text into words or subwords (i.e. tokenizing a text). More specifically, we will look at the three main types of
-tokenizers used in 🤗 Transformers: :ref:`Byte-Pair Encoding (BPE) <byte-pair-encoding>`, :ref:`WordPiece <wordpiece>`,
-and :ref:`SentencePiece <sentencepiece>`, and show examples of which tokenizer type is used by which model.
+On this page, we will have a closer look at tokenization.
+
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/VFp38yj8h3A" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
+As we saw in :doc:`the preprocessing tutorial <preprocessing>`, tokenizing a text is splitting it into words or
+subwords, which then are converted to ids through a look-up table. Converting words or subwords to ids is
+straightforward, so in this summary, we will focus on splitting a text into words or subwords (i.e. tokenizing a text).
+More specifically, we will look at the three main types of tokenizers used in 🤗 Transformers: :ref:`Byte-Pair Encoding
+(BPE) <byte-pair-encoding>`, :ref:`WordPiece <wordpiece>`, and :ref:`SentencePiece <sentencepiece>`, and show examples
+of which tokenizer type is used by which model.
 
 Note that on each model page, you can look at the documentation of the associated tokenizer to know which tokenizer
 type was used by the pretrained model. For instance, if we look at :class:`~transformers.BertTokenizer`, we can see
@@ -28,8 +36,15 @@ Introduction
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 Splitting a text into smaller chunks is a task that is harder than it looks, and there are multiple ways of doing so.
-For instance, let's look at the sentence ``"Don't you love 🤗 Transformers? We sure do."`` A simple way of tokenizing
-this text is to split it by spaces, which would give:
+For instance, let's look at the sentence ``"Don't you love 🤗 Transformers? We sure do."``
+
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/nhJxYji1aho" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
+A simple way of tokenizing this text is to split it by spaces, which would give:
 
 .. code-block::
 
@@ -69,16 +84,30 @@ Such a big vocabulary size forces the model to have an enormous embedding matrix
 causes both an increased memory and time complexity. In general, transformers models rarely have a vocabulary size
 greater than 50,000, especially if they are pretrained only on a single language.
 
-So if simple space and punctuation tokenization is unsatisfactory, why not simply tokenize on characters? While
-character tokenization is very simple and would greatly reduce memory and time complexity it makes it much harder for
-the model to learn meaningful input representations. *E.g.* learning a meaningful context-independent representation
-for the letter ``"t"`` is much harder than learning a context-independent representation for the word ``"today"``.
-Therefore, character tokenization is often accompanied by a loss of performance. So to get the best of both worlds,
-transformers models use a hybrid between word-level and character-level tokenization called **subword** tokenization.
+So if simple space and punctuation tokenization is unsatisfactory, why not simply tokenize on characters?
+
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/ssLq_EK2jLE" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
+While character tokenization is very simple and would greatly reduce memory and time complexity it makes it much harder
+for the model to learn meaningful input representations. *E.g.* learning a meaningful context-independent
+representation for the letter ``"t"`` is much harder than learning a context-independent representation for the word
+``"today"``. Therefore, character tokenization is often accompanied by a loss of performance. So to get the best of
+both worlds, transformers models use a hybrid between word-level and character-level tokenization called **subword**
+tokenization.
 
 Subword tokenization
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/zHvTiHr506c" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 Subword tokenization algorithms rely on the principle that frequently used words should not be split into smaller
 subwords, but rare words should be decomposed into meaningful subwords. For instance ``"annoyingly"`` might be
 considered a rare word and could be decomposed into ``"annoying"`` and ``"ly"``. Both ``"annoying"`` and ``"ly"`` as

docs/source/training.rst

@@ -27,6 +27,12 @@ negative. For examples of other tasks, refer to the :ref:`additional-resources`
 Preparing the datasets
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/_BZearw7f0w" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 We will use the `🤗 Datasets <https:/github.com/huggingface/datasets/>`__ library to download and preprocess the IMDB
 datasets. We will go over this part pretty quickly. Since the focus of this tutorial is on training, you should refer
 to the 🤗 Datasets `documentation <https://huggingface.co/docs/datasets/>`__ or the :doc:`preprocessing` tutorial for
@@ -95,6 +101,12 @@ them by their `full` equivalent to train or evaluate on the full dataset.
 Fine-tuning in PyTorch with the Trainer API
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/nvBXf7s7vTI" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 Since PyTorch does not provide a training loop, the 🤗 Transformers library provides a :class:`~transformers.Trainer`
 API that is optimized for 🤗 Transformers models, with a wide range of training options and with built-in features like
 logging, gradient accumulation, and mixed precision.
@@ -140,7 +152,7 @@ To fine-tune our model, we just need to call
     trainer.train()
 
 which will start a training that you can follow with a progress bar, which should take a couple of minutes to complete
-(as long as you hav access to a GPU). It won't actually tell you anything useful about how well (or badly) your model
+(as long as you have access to a GPU). It won't actually tell you anything useful about how well (or badly) your model
 is performing however as by default, there is no evaluation during training, and we didn't tell the
 :class:`~transformers.Trainer` to compute any metrics. Let's have a look on how to do that now!
 
@@ -200,6 +212,12 @@ See the documentation of :class:`~transformers.TrainingArguments` for more optio
 Fine-tuning with Keras
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/rnTGBy2ax1c" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 Models can also be trained natively in TensorFlow using the Keras API. First, let's define our model:
 
 .. code-block:: python
@@ -257,6 +275,12 @@ as a PyTorch model (or vice-versa):
 Fine-tuning in native PyTorch
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
+.. raw:: html
+
+   <iframe width="560" height="315" src="https://www.youtube.com/embed/Dh9CL8fyG80" title="YouTube video player"
+   frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope;
+   picture-in-picture" allowfullscreen></iframe>
+
 You might need to restart your notebook at this stage to free some memory, or excute the following code:
 
 .. code-block:: python

examples/flax/README.md

@@ -0,0 +1,61 @@
+<!---
+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.
+-->
+
+# JAX/Flax Examples
+
+This folder contains actively maintained examples of 🤗 Transformers using the JAX/Flax backend. Porting models and examples to JAX/Flax is an ongoing effort, and more will be added in the coming months. In particular, these examples are all designed to run fast on Cloud TPUs, and we include step-by-step guides to getting started with Cloud TPU.
+
+*NOTE*: Currently, there is no "Trainer" abstraction for JAX/Flax -- all examples contain an explicit training loop.
+
+The following table lists all of our examples on how to use 🤗 Transformers with the JAX/Flax backend:
+- with information about the model and dataset used,
+- whether or not they leverage the [🤗 Datasets](https://github.com/huggingface/datasets) library,
+- links to **Colab notebooks** to walk through the scripts and run them easily.
+
+| Task | Example model | Example dataset | 🤗 Datasets | Colab
+|---|---|---|:---:|:---:|
+| [**`causal-language-modeling`**](https://github.com/huggingface/transformers/tree/master/examples/flax/language-modeling) | GPT2 | OSCAR | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/master/examples/causal_language_modeling_flax.ipynb)
+| [**`masked-language-modeling`**](https://github.com/huggingface/transformers/tree/master/examples/flax/language-modeling) | RoBERTa | OSCAR | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/master/examples/masked_language_modeling_flax.ipynb)
+| [**`text-classification`**](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) | BERT | GLUE | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/master/examples/text_classification_flax.ipynb)
+
+## Intro: JAX and Flax
+
+[JAX](https://github.com/google/jax) is a numerical computation library that exposes a NumPy-like API with tracing capabilities. With JAX's `jit`, you can
+trace pure functions and compile them into efficient, fused accelerator code on both GPU and TPU. JAX
+supports additional transformations such as `grad` (for arbitrary gradients), `pmap` (for parallelizing computation on multiple devices), `remat` (for gradient checkpointing), `vmap` (automatic
+efficient vectorization), and `pjit` (for automatically sharded model parallelism). All JAX transformations compose arbitrarily with each other -- e.g., efficiently
+computing per-example gradients is simply `vmap(grad(f))`.
+
+[Flax](https://github.com/google/flax) builds on top of JAX with an ergonomic
+module abstraction using Python dataclasses that leads to concise and explicit code. Flax's "lifted" JAX transformations (e.g. `vmap`, `remat`) allow you to nest JAX transformation and modules in any way you wish. Flax is the most widely used JAX library, with [129 dependent projects](https://github.com/google/flax/network/dependents?package_id=UGFja2FnZS01MjEyMjA2MA%3D%3D) as of May 2021. It is also the library underlying all of the official Cloud TPU JAX examples.
+
+## Running on Cloud TPU
+
+All of our JAX/Flax models are designed to run efficiently on Google
+Cloud TPUs. Here is [a guide for running JAX on Google Cloud TPU](https://cloud.google.com/tpu/docs/jax-quickstart-tpu-vm).
+
+Each example README contains more details on the specific model and training
+procedure.
+
+## Supported models
+
+Porting models from PyTorch to JAX/Flax is an ongoing effort. 
+Feel free to reach out if you are interested in contributing a model in JAX/Flax -- we'll 
+be adding a guide for porting models from PyTorch in the upcoming few weeks.
+
+For a complete overview of models that are supported in JAX/Flax, please have a look at [this](https://huggingface.co/transformers/master/index.html#supported-frameworks) table.
+
+Over 3000 pretrained checkpoints are supported in JAX/Flax as of May 2021.
+Click [here](https://huggingface.co/models?filter=jax) to see the full list on the 🤗 hub.

examples/flax/language-modeling/README.md

@@ -0,0 +1,498 @@
+<!---
+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.
+-->
+
+# Language model training examples
+
+The following example showcases how to train a language model from scratch 
+using the JAX/Flax backend.
+
+JAX/Flax allows you to trace pure functions and compile them into efficient, fused accelerator code on both GPU and TPU.
+Models written in JAX/Flax are **immutable** and updated in a purely functional
+way which enables simple and efficient model parallelism.
+
+## Masked language modeling
+
+In the following, we demonstrate how to train a bi-directional transformer model 
+using masked language modeling objective as introduced in [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805).
+More specifically, we demonstrate how JAX/Flax can be leveraged 
+to pre-train [**`roberta-base`**](https://huggingface.co/roberta-base)
+in Norwegian on a single TPUv3-8 pod.
+
+The example script uses the 🤗 Datasets library. You can easily customize them to your needs if you need extra processing on your datasets.
+
+Let's start by creating a model repository to save the trained model and logs.
+Here we call the model `"norwegian-roberta-base"`, but you can change the model name as you like.
+
+You can do this either directly on [huggingface.co](https://huggingface.co/new) (assuming that
+you are logged in) or via the command line:
+
+```
+huggingface-cli repo create norwegian-roberta-base
+```
+
+Next we clone the model repository to add the tokenizer and model files.
+
+```
+git clone https://huggingface.co/<your-username>/norwegian-roberta-base
+```
+
+To ensure that all tensorboard traces will be uploaded correctly, we need to 
+track them. You can run the following command inside your model repo to do so.
+
+```
+cd norwegian-roberta-base
+git lfs track "*tfevents*"
+```
+
+Great, we have set up our model repository. During training, we will automatically
+push the training logs and model weights to the repo.
+
+Next, let's add a symbolic link to the `run_mlm_flax.py`.
+
+```bash
+export MODEL_DIR="./norwegian-roberta-base"
+ln -s ~/transformers/examples/flax/language-modeling/run_mlm_flax.py run_mlm_flax.py
+```
+
+### Train tokenizer
+
+In the first step, we train a tokenizer to efficiently process the text input for the model. Similar to how it is shown in [How to train a new language model from scratch using Transformers and Tokenizers](https://huggingface.co/blog/how-to-train), we use a **`ByteLevelBPETokenizer`**.
+The tokenizer is trained on the complete Norwegian dataset of OSCAR
+and consequently saved in `${MODEL_DIR}`
+This can take up to 10 minutes depending on your hardware ☕.
+
+```python
+from datasets import load_dataset
+from tokenizers import trainers, Tokenizer, normalizers, ByteLevelBPETokenizer
+
+model_dir = "./norwegian-roberta-base"  # ${MODEL_DIR}
+
+# load dataset
+dataset = load_dataset("oscar", "unshuffled_deduplicated_no", split="train")
+
+# Instantiate tokenizer
+tokenizer = ByteLevelBPETokenizer()
+
+def batch_iterator(batch_size=1000):
+    for i in range(0, len(dataset), batch_size):
+        yield dataset[i: i + batch_size]["text"]
+
+# Customized training
+tokenizer.train_from_iterator(batch_iterator(), vocab_size=50265, min_frequency=2, special_tokens=[
+    "<s>",
+    "<pad>",
+    "</s>",
+    "<unk>",
+    "<mask>",
+])
+
+# Save files to disk
+tokenizer.save(f"{model_dir}/tokenizer.json")
+```
+
+### Create configuration
+
+Next, we create the model's configuration file. This is as simple 
+as loading and storing [`**roberta-base**`](https://huggingface.co/roberta-base)
+in the local model folder:
+
+```python
+from transformers import RobertaConfig
+
+model_dir = "./norwegian-roberta-base"  # ${MODEL_DIR}
+
+config = RobertaConfig.from_pretrained("roberta-base", vocab_size=tokenizer.vocab_size)
+config.save_pretrained(model_dir)
+```
+
+### Train model
+
+Next we can run the example script to pretrain the model:
+
+```bash
+./run_mlm_flax.py \
+    --output_dir="${MODEL_DIR}" \
+    --model_type="roberta" \
+    --config_name="${MODEL_DIR}" \
+    --tokenizer_name="${MODEL_DIR}" \
+    --dataset_name="oscar" \
+    --dataset_config_name="unshuffled_deduplicated_no" \
+    --max_seq_length="128" \
+    --weight_decay="0.01" \
+    --per_device_train_batch_size="128" \
+    --per_device_eval_batch_size="128" \
+    --learning_rate="3e-4" \
+    --warmup_steps="1000" \
+    --overwrite_output_dir \
+    --num_train_epochs="18" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --logging_steps="500" \
+    --save_steps="2500" \
+    --eval_steps="2500" \
+    --push_to_hub
+```
+
+Training should converge at a loss and accuracy 
+of 1.78 and 0.64 respectively after 18 epochs on a single TPUv3-8.
+This should take less than 18 hours.
+Training statistics can be accessed on [tfhub.de](https://tensorboard.dev/experiment/GdYmdak2TWeVz0DDRYOrrg).
+
+For a step-by-step walkthrough of how to do masked language modeling in Flax, please have a 
+look at [this](https://colab.research.google.com/github/huggingface/notebooks/blob/master/examples/masked_language_modeling_flax.ipynb) google colab.
+
+## Causal language modeling
+
+In the following, we demonstrate how to train an auto-regressive causal transformer model 
+in JAX/Flax.
+More specifically, we pretrain a randomely initialized [**`gpt2`**](https://huggingface.co/gpt2) model in Norwegian on a single TPUv3-8.
+to pre-train 124M [**`gpt2`**](https://huggingface.co/gpt2)
+in Norwegian on a single TPUv3-8 pod.
+
+The example script uses the 🤗 Datasets library. You can easily customize them to your needs if you need extra processing on your datasets.
+
+Let's start by creating a model repository to save the trained model and logs.
+Here we call the model `"norwegian-gpt2"`, but you can change the model name as you like.
+
+You can do this either directly on [huggingface.co](https://huggingface.co/new) (assuming that
+you are logged in) or via the command line:
+
+```
+huggingface-cli repo create norwegian-gpt2
+```
+
+Next we clone the model repository to add the tokenizer and model files.
+
+```
+git clone https://huggingface.co/<your-username>/norwegian-gpt2
+```
+
+To ensure that all tensorboard traces will be uploaded correctly, we need to 
+track them. You can run the following command inside your model repo to do so.
+
+```
+cd norwegian-gpt2
+git lfs track "*tfevents*"
+```
+
+Great, we have set up our model repository. During training, we will automatically
+push the training logs and model weights to the repo.
+
+Next, let's add a symbolic link to the `run_clm_flax.py`.
+
+```bash
+export MODEL_DIR="./norwegian-gpt2"
+ln -s ~/transformers/examples/flax/language-modeling/run_clm_flax.py run_clm_flax.py
+```
+
+Next, we'll follow the same steps as above in [Train tokenizer](#train-tokenizer) to train the tokenizer.
+
+### Create configuration
+
+Next, we create the model's configuration file. This is as simple 
+as loading and storing [`**gpt2**`](https://huggingface.co/gpt2)
+in the local model folder:
+
+```python
+from transformers import GPT2Config
+
+model_dir = "./norwegian-gpt2"  # ${MODEL_DIR}
+
+config = GPT2Config.from_pretrained("gpt2", resid_pdrop=0.0, embd_pdrop=0.0, attn_pdrop=0.0, vocab_size=tokenizer.vocab_size)
+config.save_pretrained(model_dir)
+```
+
+### Train model
+
+Next we can run the example script to pretrain the model:
+
+```bash
+./run_clm_flax.py \
+    --output_dir="${MODEL_DIR}" \
+    --model_type="gpt2" \
+    --config_name="${MODEL_DIR}" \
+    --tokenizer_name="${MODEL_DIR}" \
+    --dataset_name="oscar" \
+    --dataset_config_name="unshuffled_deduplicated_no" \
+    --do_train --do_eval \
+    --block_size="512" \
+    --per_device_train_batch_size="64" \
+    --per_device_eval_batch_size="64" \
+    --learning_rate="5e-3" --warmup_steps="1000" \
+    --adam_beta1="0.9" --adam_beta2="0.98" --weight_decay="0.01" \
+    --overwrite_output_dir \
+    --num_train_epochs="20" \
+    --logging_steps="500" \
+    --save_steps="2500" \
+    --eval_steps="2500" \
+    --push_to_hub
+```
+
+Training should converge at a loss and perplexity 
+of 3.24 and 25.72 respectively after 20 epochs on a single TPUv3-8.
+This should take less than ~21 hours.
+Training statistics can be accessed on [tfhub.de](https://tensorboard.dev/experiment/2zEhLwJ0Qp2FAkI3WVH9qA).
+
+## T5-like span-masked language modeling
+
+In the following, we demonstrate how to train a T5 model using the span-masked language model 
+objective as proposed in the [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683).
+More specifically, we demonstrate how JAX/Flax can be leveraged 
+to pre-train [**`google/t5-v1_1-base`**](https://huggingface.co/google/t5-v1_1-base)
+in Norwegian on a single TPUv3-8 pod.
+
+The example script uses the 🤗 Datasets library. You can easily customize them to your needs if you need extra processing on your datasets.
+
+Let's start by creating a model repository to save the trained model and logs.
+Here we call the model `"norwegian-t5-base"`, but you can change the model name as you like.
+
+You can do this either directly on [huggingface.co](https://huggingface.co/new) (assuming that
+you are logged in) or via the command line:
+
+```
+huggingface-cli repo create norwegian-t5-base
+```
+
+Next we clone the model repository to add the tokenizer and model files.
+
+```
+git clone https://huggingface.co/<your-username>/norwegian-t5-base
+```
+
+To ensure that all tensorboard traces will be uploaded correctly, we need to 
+track them. You can run the following command inside your model repo to do so.
+
+```
+cd norwegian-t5-base
+git lfs track "*tfevents*"
+```
+
+Great, we have set up our model repository. During training, we will automatically
+push the training logs and model weights to the repo.
+
+Next, let's add a symbolic link to the `run_t5_mlm_flax.py` and `t5_tokenizer_model` scripts.
+
+```bash
+export MODEL_DIR="./norwegian-t5-base"
+ln -s ~/transformers/examples/flax/language-modeling/run_t5_mlm_flax.py run_t5_mlm_flax.py
+ln -s ~/transformers/examples/flax/language-modeling/t5_tokenizer_model.py t5_tokenizer_model.py
+```
+
+### Train tokenizer
+
+In the first step, we train a tokenizer to efficiently process the text input for the model. 
+We make use of the [tokenizers](https://github.com/huggingface/tokenizers) library to train 
+a sentencepiece unigram tokenizer as shown in [t5_tokenizer_model.py](https://github.com/huggingface/transformers/tree/master/examples/flax/language-modeling/t5_tokenizer_model.py) 
+which is heavily inspired from [yandex-research/DeDLOC's tokenizer model](https://github.com/yandex-research/DeDLOC/blob/5c994bc64e573702a9a79add3ecd68b38f14b548/sahajbert/tokenizer/tokenizer_model.py) .
+
+The tokenizer is trained on the complete Norwegian dataset of OSCAR
+and consequently saved in `${MODEL_DIR}`
+This can take up to 120 minutes depending on your hardware ☕☕☕ .
+
+```python
+import datasets
+
+from t5_tokenizer_model import SentencePieceUnigramTokenizer
+
+
+vocab_size = 32_000
+input_sentence_size = None
+model_dir = "./norwegian-t5-base"  # ${MODEL_DIR}
+
+# Initialize a dataset
+dataset = datasets.load_dataset("oscar", name="unshuffled_deduplicated_no", split="train")
+
+tokenizer = SentencePieceUnigramTokenizer(unk_token="<unk>", eos_token="</s>", pad_token="<pad>")
+
+
+# Build an iterator over this dataset
+def batch_iterator(input_sentence_size=None):
+    if input_sentence_size is None:
+        input_sentence_size = len(dataset)
+    batch_length = 100
+    for i in range(0, input_sentence_size, batch_length):
+        yield dataset[i: i + batch_length]["text"]
+
+
+# Train tokenizer
+tokenizer.train_from_iterator(
+    iterator=batch_iterator(input_sentence_size=input_sentence_size),
+    vocab_size=vocab_size,
+    show_progress=True,
+)
+
+# Save files to disk
+tokenizer.save(f"{model_dir}/tokenizer.json")
+```
+
+### Create configuration
+
+Next, we create the model's configuration file. This is as simple 
+as loading and storing [`**google/t5-v1_1-base**`](https://huggingface.co/google/t5-v1_1-base)
+in the local model folder:
+
+```python
+from transformers import T5Config
+
+model_dir = "./norwegian-t5-base"  # ${MODEL_DIR}
+
+config = T5Config.from_pretrained("google/t5-v1_1-base", vocab_size=tokenizer.vocab_size)
+config.save_pretrained(model_dir)
+```
+
+### Train model
+
+Next we can run the example script to pretrain the model:
+
+```bash
+./run_t5_mlm_flax.py \
+	--output_dir="./" \
+	--model_type="t5" \
+	--config_name="./" \
+	--tokenizer_name="./" \
+	--dataset_name="oscar" \
+	--dataset_config_name="unshuffled_deduplicated_no" \
+	--max_seq_length="512" \
+	--per_device_train_batch_size="32" \
+	--per_device_eval_batch_size="32" \
+	--adafactor \
+	--learning_rate="0.005" \
+	--weight_decay="0.001" \
+	--warmup_steps="2000" \
+	--overwrite_output_dir \
+	--logging_steps="100" \
+	--save_steps="1000" \
+	--eval_steps="1000" \
+	--push_to_hub
+```
+
+Training should converge at a loss and accuracy 
+of 2.2 and 58.0 respectively after 2 epochs on a single TPUv3-8.
+This should take around 24 hours.
+Training statistics can be accessed on directly on the 🤗 [hub](https://huggingface.co/patrickvonplaten/t5-base-norwegian/tensorboard)
+
+## Runtime evaluation
+
+We also ran masked language modeling using PyTorch/XLA on a TPUv3-8, and PyTorch on 8 V100 GPUs. We report the
+overall training time below.
+For reproducibility, we state the training commands used for PyTorch/XLA and PyTorch further below.
+
+| Task  | [TPU v3-8 (Flax)](https://tensorboard.dev/experiment/GdYmdak2TWeVz0DDRYOrrg/)  | [TPU v3-8 (Pytorch/XLA)](https://tensorboard.dev/experiment/7Jq1kcQQRAmy12KOdXek7A/)| [8 GPU (PyTorch)](https://tensorboard.dev/experiment/PJneV8FQRxa2unPw1QnVHA)  |
+|-------|-----------|------------|------------|
+| MLM   |  15h32m   |  23h46m    | 44h14m     |
+
+*All experiments are ran on Google Cloud Platform. 
+GPU experiments are ran without further optimizations besides JAX
+transformations. GPU experiments are ran with full precision (fp32). "TPU v3-8"
+are 8 TPU cores on 4 chips (each chips has 2 cores), while "8 GPU" are 8 GPU chips.
+
+### Script to run MLM with PyTorch/XLA on TPUv3-8
+
+For comparison one can run the same pre-training with PyTorch/XLA on TPU. To set up PyTorch/XLA on Cloud TPU VMs, please 
+refer to [this](https://cloud.google.com/tpu/docs/pytorch-xla-ug-tpu-vm) guide.
+Having created the tokenzier and configuration in `norwegian-roberta-base`, we create the following symbolic links:
+
+```bash
+ln -s ~/transformers/examples/pytorch/language-modeling/run_mlm.py ./
+ln -s ~/transformers/examples/pytorch/xla_spawn.py ./
+```
+
+, set the following environment variables:
+
+```bash
+export XRT_TPU_CONFIG="localservice;0;localhost:51011"
+unset LD_PRELOAD
+
+export NUM_TPUS=8
+export TOKENIZERS_PARALLELISM=0
+export MODEL_DIR="./norwegian-roberta-base"
+mkdir -p ${MODEL_DIR}
+```
+
+, and start training as follows:
+
+```bash
+python3 xla_spawn.py --num_cores ${NUM_TPUS} run_mlm.py --output_dir="./runs" \
+    --model_type="roberta" \
+    --config_name="${MODEL_DIR}" \
+    --tokenizer_name="${MODEL_DIR}" \
+    --dataset_name="oscar" \
+    --dataset_config_name="unshuffled_deduplicated_no" \
+    --max_seq_length="128" \
+    --weight_decay="0.01" \
+    --per_device_train_batch_size="128" \
+    --per_device_eval_batch_size="128" \
+    --learning_rate="3e-4" \
+    --warmup_steps="1000" \
+    --overwrite_output_dir \
+    --num_train_epochs="18" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --do_train \
+    --do_eval \
+    --logging_steps="500" \
+    --evaluation_strategy="epoch" \
+    --report_to="tensorboard" \
+    --save_strategy="no"
+```
+
+### Script to compare pre-training with PyTorch on 8 GPU V100's
+
+For comparison you can run the same pre-training with PyTorch on GPU. Note that we have to make use of `gradient_accumulation` 
+because the maximum batch size that fits on a single V100 GPU is 32 instead of 128.
+Having created the tokenzier and configuration in `norwegian-roberta-base`, we create the following symbolic links:
+
+```bash
+ln -s ~/transformers/examples/pytorch/language-modeling/run_mlm.py ./
+```
+
+, set some environment variables:
+
+```bash
+export NUM_GPUS=8
+export TOKENIZERS_PARALLELISM=0
+export MODEL_DIR="./norwegian-roberta-base"
+mkdir -p ${MODEL_DIR}
+```
+
+, and can start training as follows:
+
+```bash
+python3 -m torch.distributed.launch --nproc_per_node ${NUM_GPUS} run_mlm.py \
+    --output_dir="${MODEL_DIR}" \
+    --model_type="roberta" \
+    --config_name="${MODEL_DIR}" \
+    --tokenizer_name="${MODEL_DIR}" \
+    --dataset_name="oscar" \
+    --dataset_config_name="unshuffled_deduplicated_no" \
+    --max_seq_length="128" \
+    --weight_decay="0.01" \
+    --per_device_train_batch_size="32" \
+    --per_device_eval_batch_size="32" \
+    --gradient_accumulation="4" \
+    --learning_rate="3e-4" \
+    --warmup_steps="1000" \
+    --overwrite_output_dir \
+    --num_train_epochs="18" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --do_train \
+    --do_eval \
+    --logging_steps="500" \
+    --evaluation_strategy="steps" \
+    --report_to="tensorboard" \
+    --save_strategy="no"
+```

examples/flax/language-modeling/requirements.txt

@@ -0,0 +1,5 @@
+datasets >= 1.1.3
+jax>=0.2.8
+jaxlib>=0.1.59
+flax>=0.3.4
+optax>=0.0.9

examples/flax/language-modeling/run_clm_flax.py

@@ -0,0 +1,654 @@
+#!/usr/bin/env python
+# coding=utf-8
+# 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.
+"""
+Pre-training/Fine-tuning the library models for causal language modeling (GPT, GPT-2, CTRL, ...) on a text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=causal-lm
+"""
+# You can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.
+
+import logging
+import math
+import os
+import sys
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable, Optional
+
+import datasets
+from datasets import Dataset, load_dataset
+from tqdm import tqdm
+
+import jax
+import jax.numpy as jnp
+import optax
+import transformers
+from flax import jax_utils, traverse_util
+from flax.jax_utils import unreplicate
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_CAUSAL_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForCausalLM,
+    HfArgumentParser,
+    TrainingArguments,
+    is_tensorboard_available,
+)
+from transformers.testing_utils import CaptureLogger
+
+
+logger = logging.getLogger(__name__)
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_CAUSAL_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+            "value if set."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    block_size: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Optional input sequence length after tokenization. "
+            "The training dataset will be truncated in block of this size for training. "
+            "Default to the model max input length for single sentence inputs (take into account special tokens)."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+class TrainState(train_state.TrainState):
+    dropout_rng: jnp.ndarray
+
+    def replicate(self):
+        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
+
+
+def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False):
+    """
+    Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
+    Shuffle batches if `shuffle` is `True`.
+    """
+    steps_per_epoch = len(dataset) // batch_size
+
+    if shuffle:
+        batch_idx = jax.random.permutation(rng, len(dataset))
+    else:
+        batch_idx = jnp.arange(len(dataset))
+
+    batch_idx = batch_idx[: steps_per_epoch * batch_size]  # Skip incomplete batch.
+    batch_idx = batch_idx.reshape((steps_per_epoch, batch_size))
+
+    for idx in batch_idx:
+        batch = dataset[idx]
+        batch = {k: jnp.array(v) for k, v in batch.items()}
+
+        batch = shard(batch)
+
+        yield batch
+
+
+def write_train_metric(summary_writer, train_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+
+def write_eval_metric(summary_writer, eval_metrics, step):
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+def create_learning_rate_fn(
+    train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, linear_decay learning rate function."""
+    steps_per_epoch = train_ds_size // train_batch_size
+    num_train_steps = steps_per_epoch * num_train_epochs
+    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
+    decay_fn = optax.linear_schedule(
+        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
+    )
+    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
+    return schedule_fn
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    #  Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        dataset = load_dataset(
+            data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir, keep_in_memory=False
+        )
+
+        if "validation" not in dataset.keys():
+            dataset["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            dataset["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        dataset = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForCausalLM.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForCausalLM.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = dataset["train"].column_names
+    else:
+        column_names = dataset["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    # since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
+    tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")
+
+    def tokenize_function(examples):
+        with CaptureLogger(tok_logger) as cl:
+            output = tokenizer(examples[text_column_name])
+        # clm input could be much much longer than block_size
+        if "Token indices sequence length is longer than the" in cl.out:
+            tok_logger.warning(
+                "^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model."
+            )
+        return output
+
+    tokenized_datasets = dataset.map(
+        tokenize_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        remove_columns=column_names,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if data_args.block_size is None:
+        block_size = tokenizer.model_max_length
+        if block_size > config.max_position_embeddings:
+            logger.warning(
+                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
+                "Picking 1024 instead. You can change that default value by passing --block_size xxx."
+            )
+            block_size = 1024
+    else:
+        if data_args.block_size > tokenizer.model_max_length:
+            logger.warning(
+                f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
+                f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
+            )
+        block_size = min(data_args.block_size, tokenizer.model_max_length)
+
+    # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
+    def group_texts(examples):
+        # Concatenate all texts.
+        concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
+        total_length = len(concatenated_examples[list(examples.keys())[0]])
+        # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+        # customize this part to your needs.
+        if total_length >= block_size:
+            total_length = (total_length // block_size) * block_size
+        # Split by chunks of max_len.
+        result = {
+            k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
+            for k, t in concatenated_examples.items()
+        }
+        result["labels"] = result["input_ids"].copy()
+        return result
+
+    # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
+    # for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
+    # to preprocess.
+    #
+    # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+    # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+
+    lm_datasets = tokenized_datasets.map(
+        group_texts,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if training_args.do_train:
+        if "train" not in tokenized_datasets:
+            raise ValueError("--do_train requires a train dataset")
+        train_dataset = lm_datasets["train"]
+        if data_args.max_train_samples is not None:
+            train_dataset = train_dataset.select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        if "validation" not in tokenized_datasets:
+            raise ValueError("--do_eval requires a validation dataset")
+        eval_dataset = lm_datasets["validation"]
+        if data_args.max_eval_samples is not None:
+            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+
+            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+    else:
+        logger.warning(
+            "Unable to display metrics through TensorBoard because the package is not installed: "
+            "Please run pip install tensorboard to enable."
+        )
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    rng, dropout_rng = jax.random.split(rng)
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+    steps_per_epoch = len(train_dataset) // train_batch_size
+    total_train_steps = steps_per_epoch * num_epochs
+
+    # Create learning rate schedule
+    linear_decay_lr_schedule_fn = create_learning_rate_fn(
+        len(train_dataset),
+        train_batch_size,
+        training_args.num_train_epochs,
+        training_args.warmup_steps,
+        training_args.learning_rate,
+    )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    # Note that this mask is specifically adapted for FlaxGPT2.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {
+            path: (path[-1] != "bias" and path[-2:] not in [("ln_1", "scale"), ("ln_2", "scale"), ("ln_f", "scale")])
+            for path in flat_params
+        }
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
+
+    # Setup train state
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng)
+
+    def loss_fn(logits, labels):
+        shift_logits = logits[..., :-1, :]
+        shift_labels = labels[..., 1:]
+        loss = optax.softmax_cross_entropy(shift_logits, onehot(shift_labels, shift_logits.shape[-1]))
+        return loss.mean()
+
+    # Define gradient update step fn
+    def train_step(state, batch):
+        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+
+        def compute_loss(params):
+            labels = batch.pop("labels")
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+            loss = loss_fn(logits, labels)
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+
+        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
+
+        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_state, metrics
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+        logits = model(**batch, params=params, train=False)[0]
+        loss = loss_fn(logits, labels)
+
+        # summarize metrics
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+        return metrics
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+    p_eval_step = jax.pmap(eval_step, "batch")
+
+    # Replicate the train state on each device
+    state = state.replicate()
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {num_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    train_time = 0
+    train_metrics = []
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True)
+        steps_per_epoch = len(train_dataset) // train_batch_size
+        # train
+        for step in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False):
+            batch = next(train_loader)
+            state, train_metric = p_train_step(state, batch)
+            train_metrics.append(train_metric)
+
+            cur_step = epoch * (len(train_dataset) // train_batch_size) + step
+
+            if cur_step % training_args.logging_steps == 0 and cur_step > 0:
+                # Save metrics
+                train_metric = unreplicate(train_metric)
+                train_time += time.time() - train_start
+                if has_tensorboard and jax.process_index() == 0:
+                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
+                )
+
+                train_metrics = []
+
+            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
+                # ======================== Evaluating ==============================
+                eval_metrics = []
+                eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
+                eval_steps = len(eval_dataset) // eval_batch_size
+                for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
+                    # Model forward
+                    batch = next(eval_loader)
+                    metrics = p_eval_step(state.params, batch)
+                    eval_metrics.append(metrics)
+
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
+
+                try:
+                    eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
+                except OverflowError:
+                    eval_metrics["perplexity"] = float("inf")
+
+                # Print metrics and update progress bar
+                desc = f"Step... ({cur_step} | Eval Loss: {eval_metrics['loss']} | Eval Perplexity: {eval_metrics['perplexity']})"
+                epochs.write(desc)
+                epochs.desc = desc
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # save checkpoint after each epoch and push checkpoint to the hub
+                if jax.process_index() == 0:
+                    params = jax.device_get(unreplicate(state.params))
+                    model.save_pretrained(
+                        training_args.output_dir,
+                        params=params,
+                        push_to_hub=training_args.push_to_hub,
+                        commit_message=f"Saving weights and logs of step {cur_step}",
+                    )
+
+
+if __name__ == "__main__":
+    main()

examples/flax/language-modeling/run_mlm_flax.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python
 # coding=utf-8
-# Copyright 2020 The HuggingFace Team All rights reserved.
+# 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.
@@ -23,6 +23,7 @@ https://huggingface.co/models?filter=masked-lm
 import logging
 import os
 import sys
+import time
 from dataclasses import dataclass, field
 
 # You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
@@ -33,16 +34,16 @@ import numpy as np
 from datasets import load_dataset
 from tqdm import tqdm
 
+import flax
 import jax
 import jax.numpy as jnp
-from flax import jax_utils
-from flax.optim import Adam
-from flax.training import common_utils
-from flax.training.common_utils import get_metrics
-from jax.nn import log_softmax
+import optax
+from flax import jax_utils, traverse_util
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
 from transformers import (
     CONFIG_MAPPING,
-    MODEL_FOR_MASKED_LM_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
     AutoConfig,
     AutoTokenizer,
     FlaxAutoModelForMaskedLM,
@@ -55,23 +56,7 @@ from transformers import (
 )
 
 
-# Cache the result
-has_tensorboard = is_tensorboard_available()
-if has_tensorboard:
-    try:
-        from flax.metrics.tensorboard import SummaryWriter
-    except ImportError as ie:
-        has_tensorboard = False
-        print(f"Unable to display metrics through TensorBoard because some package are not installed: {ie}")
-
-else:
-    print(
-        "Unable to display metrics through TensorBoard because the package is not installed: "
-        "Please run pip install tensorboard to enable."
-    )
-
-
-MODEL_CONFIG_CLASSES = list(MODEL_FOR_MASKED_LM_MAPPING.keys())
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
 MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
 
 
@@ -185,9 +170,7 @@ class DataTrainingArguments:
                 assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
 
 
-# Adapted from transformers/data/data_collator.py
-# Letting here for now, let's discuss where it should live
-@dataclass
+@flax.struct.dataclass
 class FlaxDataCollatorForLanguageModeling:
     """
     Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
@@ -196,12 +179,8 @@ class FlaxDataCollatorForLanguageModeling:
     Args:
         tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
             The tokenizer used for encoding the data.
-        mlm (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Whether or not to use masked language modeling. If set to :obj:`False`, the labels are the same as the
-            inputs with the padding tokens ignored (by setting them to -100). Otherwise, the labels are -100 for
-            non-masked tokens and the value to predict for the masked token.
         mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
-            The probability with which to (randomly) mask tokens in the input, when :obj:`mlm` is set to :obj:`True`.
+            The probability with which to (randomly) mask tokens in the input.
 
     .. note::
 
@@ -212,11 +191,10 @@ class FlaxDataCollatorForLanguageModeling:
     """
 
     tokenizer: PreTrainedTokenizerBase
-    mlm: bool = True
     mlm_probability: float = 0.15
 
     def __post_init__(self):
-        if self.mlm and self.tokenizer.mask_token is None:
+        if self.tokenizer.mask_token is None:
             raise ValueError(
                 "This tokenizer does not have a mask token which is necessary for masked language modeling. "
                 "You should pass `mlm=False` to train on causal language modeling instead."
@@ -228,15 +206,10 @@ class FlaxDataCollatorForLanguageModeling:
 
         # If special token mask has been preprocessed, pop it from the dict.
         special_tokens_mask = batch.pop("special_tokens_mask", None)
-        if self.mlm:
-            batch["input_ids"], batch["labels"] = self.mask_tokens(
-                batch["input_ids"], special_tokens_mask=special_tokens_mask
-            )
-        else:
-            labels = batch["input_ids"].copy()
-            if self.tokenizer.pad_token_id is not None:
-                labels[labels == self.tokenizer.pad_token_id] = -100
-            batch["labels"] = labels
+
+        batch["input_ids"], batch["labels"] = self.mask_tokens(
+            batch["input_ids"], special_tokens_mask=special_tokens_mask
+        )
         return batch
 
     def mask_tokens(
@@ -269,167 +242,32 @@ class FlaxDataCollatorForLanguageModeling:
         return inputs, labels
 
 
-def create_learning_rate_scheduler(
-    factors="constant * linear_warmup * rsqrt_decay",
-    base_learning_rate=0.5,
-    warmup_steps=1000,
-    decay_factor=0.5,
-    steps_per_decay=20000,
-    steps_per_cycle=100000,
-):
-    """Creates learning rate schedule.
-    Interprets factors in the factors string which can consist of:
-    * constant: interpreted as the constant value,
-    * linear_warmup: interpreted as linear warmup until warmup_steps,
-    * rsqrt_decay: divide by square root of max(step, warmup_steps)
-    * rsqrt_normalized_decay: divide by square root of max(step/warmup_steps, 1)
-    * decay_every: Every k steps decay the learning rate by decay_factor.
-    * cosine_decay: Cyclic cosine decay, uses steps_per_cycle parameter.
-    Args:
-      factors: string, factors separated by "*" that defines the schedule.
-      base_learning_rate: float, the starting constant for the lr schedule.
-      warmup_steps: int, how many steps to warm up for in the warmup schedule.
-      decay_factor: float, the amount to decay the learning rate by.
-      steps_per_decay: int, how often to decay the learning rate.
-      steps_per_cycle: int, steps per cycle when using cosine decay.
-    Returns:
-      a function learning_rate(step): float -> {"learning_rate": float}, the
-      step-dependent lr.
-    """
-    factors = [n.strip() for n in factors.split("*")]
-
-    def step_fn(step):
-        """Step to learning rate function."""
-        ret = 1.0
-        for name in factors:
-            if name == "constant":
-                ret *= base_learning_rate
-            elif name == "linear_warmup":
-                ret *= jnp.minimum(1.0, step / warmup_steps)
-            elif name == "rsqrt_decay":
-                ret /= jnp.sqrt(jnp.maximum(step, warmup_steps))
-            elif name == "rsqrt_normalized_decay":
-                ret *= jnp.sqrt(warmup_steps)
-                ret /= jnp.sqrt(jnp.maximum(step, warmup_steps))
-            elif name == "decay_every":
-                ret *= decay_factor ** (step // steps_per_decay)
-            elif name == "cosine_decay":
-                progress = jnp.maximum(0.0, (step - warmup_steps) / float(steps_per_cycle))
-                ret *= jnp.maximum(0.0, 0.5 * (1.0 + jnp.cos(jnp.pi * (progress % 1.0))))
-            else:
-                raise ValueError(f"Unknown factor {name}.")
-        return jnp.asarray(ret, dtype=jnp.float32)
-
-    return step_fn
-
-
-def compute_metrics(logits, labels, weights, label_smoothing=0.0):
-    """Compute summary metrics."""
-    loss, normalizer = cross_entropy(logits, labels, weights, label_smoothing)
-    acc, _ = accuracy(logits, labels, weights)
-    metrics = {"loss": loss, "accuracy": acc, "normalizer": normalizer}
-    metrics = jax.lax.psum(metrics, axis_name="batch")
-    return metrics
-
-
-def accuracy(logits, targets, weights=None):
-    """Compute weighted accuracy for log probs and targets.
-    Args:
-     logits: [batch, length, num_classes] float array.
-     targets: categorical targets [batch, length] int array.
-     weights: None or array of shape [batch, length]
-    Returns:
-      Tuple of scalar loss and batch normalizing factor.
-    """
-    if logits.ndim != targets.ndim + 1:
-        raise ValueError(f"Incorrect shapes. Got shape {logits.shape} logits and {targets.shape} targets")
-
-    loss = jnp.equal(jnp.argmax(logits, axis=-1), targets)
-    loss *= weights
-
-    return loss.sum(), weights.sum()
-
-
-def cross_entropy(logits, targets, weights=None, label_smoothing=0.0):
-    """Compute cross entropy and entropy for log probs and targets.
-    Args:
-     logits: [batch, length, num_classes] float array.
-     targets: categorical targets [batch, length] int array.
-     weights: None or array of shape [batch, length]
-     label_smoothing: label smoothing constant, used to determine the on and off values.
-    Returns:
-      Tuple of scalar loss and batch normalizing factor.
-    """
-    if logits.ndim != targets.ndim + 1:
-        raise ValueError(f"Incorrect shapes. Got shape {logits.shape} logits and {targets.shape} targets")
-
-    vocab_size = logits.shape[-1]
-    confidence = 1.0 - label_smoothing
-    low_confidence = (1.0 - confidence) / (vocab_size - 1)
-    normalizing_constant = -(
-        confidence * jnp.log(confidence) + (vocab_size - 1) * low_confidence * jnp.log(low_confidence + 1e-20)
-    )
-    soft_targets = common_utils.onehot(targets, vocab_size, on_value=confidence, off_value=low_confidence)
-
-    loss = -jnp.sum(soft_targets * log_softmax(logits), axis=-1)
-    loss = loss - normalizing_constant
-
-    if weights is not None:
-        loss = loss * weights
-        normalizing_factor = weights.sum()
-    else:
-        normalizing_factor = np.prod(targets.shape)
-
-    return loss.sum(), normalizing_factor
-
-
-def training_step(optimizer, batch, dropout_rng):
-    dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
-
-    def loss_fn(params):
-        targets = batch.pop("labels")
-
-        # Hide away tokens which doesn't participate in the optimization
-        token_mask = jnp.where(targets > 0, 1.0, 0.0)
-
-        logits = model(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
-        loss, weight_sum = cross_entropy(logits, targets, token_mask)
-        return loss / weight_sum
-
-    step = optimizer.state.step
-    lr = lr_scheduler_fn(step)
-    grad_fn = jax.value_and_grad(loss_fn)
-    loss, grad = grad_fn(optimizer.target)
-    grad = jax.lax.pmean(grad, "batch")
-    optimizer = optimizer.apply_gradient(grad, learning_rate=lr)
-
-    return loss, optimizer, new_dropout_rng
-
-
-def eval_step(params, batch):
-    """
-    Calculate evaluation metrics on a batch.
-    """
-    targets = batch.pop("labels")
-
-    # Hide away tokens which doesn't participate in the optimization
-    token_mask = jnp.where(targets > 0, 1.0, 0.0)
-    logits = model(**batch, params=params, train=False)[0]
-
-    return compute_metrics(logits, targets, token_mask)
-
-
 def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
-    nb_samples = len(samples_idx)
-    samples_to_remove = nb_samples % batch_size
+    num_samples = len(samples_idx)
+    samples_to_remove = num_samples % batch_size
 
     if samples_to_remove != 0:
         samples_idx = samples_idx[:-samples_to_remove]
-    sections_split = nb_samples // batch_size
+    sections_split = num_samples // batch_size
     batch_idx = np.split(samples_idx, sections_split)
     return batch_idx
 
 
+def write_train_metric(summary_writer, train_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+
+def write_eval_metric(summary_writer, eval_metrics, step):
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
 if __name__ == "__main__":
     # See all possible arguments in src/transformers/training_args.py
     # or by passing the --help flag to this script.
@@ -456,17 +294,13 @@ if __name__ == "__main__":
 
     # Setup logging
     logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
         level="NOTSET",
         datefmt="[%X]",
     )
 
     # Log on each process the small summary:
     logger = logging.getLogger(__name__)
-    logger.warning(
-        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
-        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
-    )
 
     # Set the verbosity to info of the Transformers logger (on main process only):
     logger.info(f"Training/evaluation parameters {training_args}")
@@ -486,6 +320,7 @@ if __name__ == "__main__":
     if data_args.dataset_name is not None:
         # Downloading and loading a dataset from the hub.
         datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
+
         if "validation" not in datasets.keys():
             datasets["validation"] = load_dataset(
                 data_args.dataset_name,
@@ -509,6 +344,20 @@ if __name__ == "__main__":
         if extension == "txt":
             extension = "text"
         datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
     # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
     # https://huggingface.co/docs/datasets/loading_datasets.html.
 
@@ -596,7 +445,8 @@ if __name__ == "__main__":
             total_length = len(concatenated_examples[list(examples.keys())[0]])
             # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
             # customize this part to your needs.
-            total_length = (total_length // max_seq_length) * max_seq_length
+            if total_length >= max_seq_length:
+                total_length = (total_length // max_seq_length) * max_seq_length
             # Split by chunks of max_len.
             result = {
                 k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
@@ -610,7 +460,6 @@ if __name__ == "__main__":
         #
         # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
         # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
-
         tokenized_datasets = tokenized_datasets.map(
             group_texts,
             batched=True,
@@ -619,8 +468,22 @@ if __name__ == "__main__":
         )
 
     # Enable tensorboard only on the master node
-    if has_tensorboard and jax.host_id() == 0:
-        summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir).joinpath("logs").as_posix())
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+
+            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+    else:
+        logger.warning(
+            "Unable to display metrics through TensorBoard because the package is not installed: "
+            "Please run pip install tensorboard to enable."
+        )
 
     # Data collator
     # This one will take care of randomly masking the tokens.
@@ -630,88 +493,199 @@ if __name__ == "__main__":
     rng = jax.random.PRNGKey(training_args.seed)
     dropout_rngs = jax.random.split(rng, jax.local_device_count())
 
-    model = FlaxAutoModelForMaskedLM.from_config(config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype))
-
-    # Setup optimizer
-    optimizer = Adam(
-        learning_rate=training_args.learning_rate,
-        weight_decay=training_args.weight_decay,
-        beta1=training_args.adam_beta1,
-        beta2=training_args.adam_beta2,
-    ).create(model.params)
-
-    # Create learning rate scheduler
-    # warmup_steps = 0 causes the Flax optimizer to return NaNs; warmup_steps = 1 is functionally equivalent.
-    lr_scheduler_fn = create_learning_rate_scheduler(
-        base_learning_rate=training_args.learning_rate, warmup_steps=max(training_args.warmup_steps, 1)
-    )
-
-    # Create parallel version of the training and evaluation steps
-    p_training_step = jax.pmap(training_step, "batch", donate_argnums=(0,))
-    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
-
-    # Replicate the optimizer on each device
-    optimizer = jax_utils.replicate(optimizer)
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForMaskedLM.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForMaskedLM.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
 
     # Store some constant
-    nb_epochs = int(training_args.num_train_epochs)
-    batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
     eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
 
-    epochs = tqdm(range(nb_epochs), desc=f"Epoch ... (1/{nb_epochs})", position=0)
-    for epoch in epochs:
+    num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
 
+    # Create learning rate schedule
+    warmup_fn = optax.linear_schedule(
+        init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
+    )
+    decay_fn = optax.linear_schedule(
+        init_value=training_args.learning_rate,
+        end_value=0,
+        transition_steps=num_train_steps - training_args.warmup_steps,
+    )
+    linear_decay_lr_schedule_fn = optax.join_schedules(
+        schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
+    )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    # Note that this mask is specifically adapted for FlaxBERT-like models.
+    # For other models, one should correct the layer norm parameter naming
+    # accordingly.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
+
+    # Setup train state
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
+
+    # Define gradient update step fn
+    def train_step(state, batch, dropout_rng):
+        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+        def loss_fn(params):
+            labels = batch.pop("labels")
+
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+
+            # compute loss, ignore padded input tokens
+            label_mask = jnp.where(labels > 0, 1.0, 0.0)
+            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+            # take average
+            loss = loss.sum() / label_mask.sum()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(loss_fn)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+        new_state = state.apply_gradients(grads=grad)
+
+        metrics = jax.lax.pmean(
+            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
+        )
+
+        return new_state, metrics, new_dropout_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+
+        logits = model(**batch, params=params, train=False)[0]
+
+        # compute loss, ignore padded input tokens
+        label_mask = jnp.where(labels > 0, 1.0, 0.0)
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask
+
+        # compute accuracy
+        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask
+
+        # summarize metrics
+        metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
+        metrics = jax.lax.psum(metrics, axis_name="batch")
+
+        return metrics
+
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+
+    train_time = 0
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
         # ======================== Training ================================
+        train_start = time.time()
+        train_metrics = []
+
         # Create sampling rng
-        rng, training_rng, eval_rng = jax.random.split(rng, 3)
+        rng, input_rng = jax.random.split(rng)
 
         # Generate an epoch by shuffling sampling indices from the train dataset
-        nb_training_samples = len(tokenized_datasets["train"])
-        training_samples_idx = jax.random.permutation(training_rng, jnp.arange(nb_training_samples))
-        training_batch_idx = generate_batch_splits(training_samples_idx, batch_size)
+        num_train_samples = len(tokenized_datasets["train"])
+        train_samples_idx = jax.random.permutation(input_rng, jnp.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
-        for batch_idx in tqdm(training_batch_idx, desc="Training...", position=1):
+        for step, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1)):
             samples = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
             model_inputs = data_collator(samples, pad_to_multiple_of=16)
 
             # Model forward
-            model_inputs = common_utils.shard(model_inputs.data)
-            loss, optimizer, dropout_rngs = p_training_step(optimizer, model_inputs, dropout_rngs)
+            model_inputs = shard(model_inputs.data)
+            state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+            train_metrics.append(train_metric)
 
-        epochs.write(f"Loss: {loss}")
+            cur_step = epoch * (num_train_samples // train_batch_size) + step
 
-        # ======================== Evaluating ==============================
-        nb_eval_samples = len(tokenized_datasets["validation"])
-        eval_samples_idx = jnp.arange(nb_eval_samples)
-        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+            if cur_step % training_args.logging_steps == 0 and cur_step > 0:
+                # Save metrics
+                train_metric = jax_utils.unreplicate(train_metric)
+                train_time += time.time() - train_start
+                if has_tensorboard and jax.process_index() == 0:
+                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
 
-        eval_metrics = []
-        for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
-            samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
-            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+                )
 
-            # Model forward
-            model_inputs = common_utils.shard(model_inputs.data)
-            metrics = p_eval_step(optimizer.target, model_inputs)
-            eval_metrics.append(metrics)
+                train_metrics = []
 
-        eval_metrics_np = get_metrics(eval_metrics)
-        eval_metrics_np = jax.tree_map(jnp.sum, eval_metrics_np)
-        eval_normalizer = eval_metrics_np.pop("normalizer")
-        eval_summary = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics_np)
+            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)
+                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
 
-        # Update progress bar
-        epochs.desc = (
-            f"Epoch... ({epoch + 1}/{nb_epochs} | Loss: {eval_summary['loss']}, Acc: {eval_summary['accuracy']})"
-        )
+                eval_metrics = []
+                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+                    samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+                    model_inputs = data_collator(samples, pad_to_multiple_of=16)
 
-        # Save metrics
-        if has_tensorboard and jax.host_id() == 0:
-            for name, value in eval_summary.items():
-                summary_writer.scalar(name, value, epoch)
+                    # Model forward
+                    model_inputs = shard(model_inputs.data)
+                    metrics = p_eval_step(state.params, model_inputs)
+                    eval_metrics.append(metrics)
 
-        # save last checkpoint
-        if jax.host_id() == 0:
-            params = jax.device_get(jax.tree_map(lambda x: x[0], optimizer.target))
-            model.save_pretrained(training_args.output_dir, params=params)
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
+                eval_normalizer = eval_metrics.pop("normalizer")
+                eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+                # Update progress bar
+                epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # save checkpoint after each epoch and push checkpoint to the hub
+                if jax.process_index() == 0:
+                    params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+                    model.save_pretrained(
+                        training_args.output_dir,
+                        params=params,
+                        push_to_hub=training_args.push_to_hub,
+                        commit_message=f"Saving weights and logs of step {cur_step}",
+                    )

examples/flax/language-modeling/run_t5_mlm_flax.py

@@ -0,0 +1,798 @@
+#!/usr/bin/env python
+# coding=utf-8
+# 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.
+"""
+Pretraining the library models for T5-like span-masked language modeling on a text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be pretrained by this script:
+https://huggingface.co/models?filter=t5
+"""
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+import logging
+import os
+import sys
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Dict, List, Optional
+
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import flax
+import jax
+import jax.numpy as jnp
+import optax
+from flax import jax_utils, traverse_util
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    AutoTokenizer,
+    BatchEncoding,
+    FlaxT5ForConditionalGeneration,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    T5Config,
+    TrainingArguments,
+    is_tensorboard_available,
+    set_seed,
+)
+from transformers.models.t5.modeling_flax_t5 import shift_tokens_right
+
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    train_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
+    )
+    validation_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    validation_split_percentage: Optional[int] = field(
+        default=5,
+        metadata={
+            "help": "The percentage of the train set used as validation set in case there's no validation split"
+        },
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization and masking. Sequences longer than this will be truncated. Default to the max input length of the model."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    mlm_probability: float = field(
+        default=0.15, metadata={"help": "Ratio of tokens to mask for span masked language modeling loss"}
+    )
+    mean_noise_span_length: float = field(
+        default=3.0,
+        metadata={"help": "Mean span length of masked tokens"},
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+def compute_input_and_target_lengths(inputs_length, noise_density, mean_noise_span_length):
+    """This function is copy of `random_spans_helper <https://github.com/google-research/text-to-text-transfer-transformer/blob/84f8bcc14b5f2c03de51bd3587609ba8f6bbd1cd/t5/data/preprocessors.py#L2466>`__ .
+
+    Training parameters to avoid padding with random_spans_noise_mask.
+    When training a model with random_spans_noise_mask, we would like to set the other
+    training hyperparmeters in a way that avoids padding.
+    This function helps us compute these hyperparameters.
+    We assume that each noise span in the input is replaced by extra_tokens_per_span_inputs sentinel tokens,
+    and each non-noise span in the targets is replaced by extra_tokens_per_span_targets sentinel tokens.
+    This function tells us the required number of tokens in the raw example (for split_tokens())
+    as well as the length of the encoded targets. Note that this function assumes
+    the inputs and targets will have EOS appended and includes that in the reported length.
+
+    Args:
+        inputs_length: an integer - desired length of the tokenized inputs sequence
+        noise_density: a float
+        mean_noise_span_length: a float
+    Returns:
+        tokens_length: length of original text in tokens
+        targets_length: an integer - length in tokens of encoded targets sequence
+    """
+
+    def _tokens_length_to_inputs_length_targets_length(tokens_length):
+        num_noise_tokens = int(round(tokens_length * noise_density))
+        num_nonnoise_tokens = tokens_length - num_noise_tokens
+        num_noise_spans = int(round(num_noise_tokens / mean_noise_span_length))
+        # inputs contain all nonnoise tokens, sentinels for all noise spans
+        # and one EOS token.
+        _input_length = num_nonnoise_tokens + num_noise_spans + 1
+        _output_length = num_noise_tokens + num_noise_spans + 1
+        return _input_length, _output_length
+
+    tokens_length = inputs_length
+
+    while _tokens_length_to_inputs_length_targets_length(tokens_length + 1)[0] <= inputs_length:
+        tokens_length += 1
+
+    inputs_length, targets_length = _tokens_length_to_inputs_length_targets_length(tokens_length)
+
+    # minor hack to get the targets length to be equal to inputs length
+    # which is more likely to have been set to a nice round number.
+    if noise_density == 0.5 and targets_length > inputs_length:
+        tokens_length -= 1
+        targets_length -= 1
+    return tokens_length, targets_length
+
+
+@flax.struct.dataclass
+class FlaxDataCollatorForT5MLM:
+    """
+    Data collator used for T5 span-masked language modeling.
+    It is made sure that after masking the inputs are of length `data_args.max_seq_length` and targets are also of fixed length.
+    For more information on how T5 span-masked language modeling works, one can take a look
+    at the `official paper <https://arxiv.org/pdf/1910.10683.pdf>`__
+    or the `official code for preprocessing <https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py>`__ .
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        noise_density (:obj:`float`):
+            The probability with which to (randomly) mask tokens in the input.
+        mean_noise_span_length (:obj:`float`):
+            The average span length of the masked tokens.
+        input_length (:obj:`int`):
+            The expected input length after masking.
+        target_length (:obj:`int`):
+            The expected target length after masking.
+        pad_token_id: (:obj:`int`):
+            The pad token id of the model
+        decoder_start_token_id: (:obj:`int):
+            The decoder start token id of the model
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    noise_density: float
+    mean_noise_span_length: float
+    input_length: int
+    target_length: int
+    pad_token_id: int
+    decoder_start_token_id: int
+
+    def __call__(self, examples: List[Dict[str, np.ndarray]]) -> Dict[str, np.ndarray]:
+
+        # convert list to dict and tensorize input
+        batch = BatchEncoding(
+            {k: np.array([examples[i][k] for i in range(len(examples))]) for k, v in examples[0].items()}
+        )
+
+        input_ids = batch["input_ids"]
+        batch_size, expandend_input_length = input_ids.shape
+
+        mask_indices = np.asarray([self.random_spans_noise_mask(expandend_input_length) for i in range(batch_size)])
+        labels_mask = ~mask_indices
+
+        input_ids_sentinel = self.create_sentinel_ids(mask_indices.astype(np.int8))
+        labels_sentinel = self.create_sentinel_ids(labels_mask.astype(np.int8))
+
+        batch["input_ids"] = self.filter_input_ids(input_ids, input_ids_sentinel)
+        batch["labels"] = self.filter_input_ids(input_ids, labels_sentinel)
+
+        if batch["input_ids"].shape[-1] != self.input_length:
+            raise ValueError(
+                f"`input_ids` are incorrectly preprocessed. `input_ids` length is {batch['input_ids'].shape[-1]}, but should be {self.target_length}."
+            )
+
+        if batch["labels"].shape[-1] != self.target_length:
+            raise ValueError(
+                f"`labels` are incorrectly preprocessed. `labels` length is {batch['labels'].shape[-1]}, but should be {self.target_length}."
+            )
+
+        # to check that tokens are correctly proprocessed, one can run `self.tokenizer.batch_decode(input_ids)` and `self.tokenizer.batch_decode(labels)` here...
+        batch["decoder_input_ids"] = shift_tokens_right(
+            batch["labels"], self.pad_token_id, self.decoder_start_token_id
+        )
+
+        return batch
+
+    def create_sentinel_ids(self, mask_indices):
+        """
+        Sentinel ids creation given the indices that should be masked.
+        The start indices of each mask are replaced by the sentinel ids in increasing
+        order. Consecutive mask indices to be deleted are replaced with `-1`.
+        """
+        start_indices = mask_indices - np.roll(mask_indices, 1, axis=-1) * mask_indices
+        start_indices[:, 0] = mask_indices[:, 0]
+
+        sentinel_ids = np.where(start_indices != 0, np.cumsum(start_indices, axis=-1), start_indices)
+        sentinel_ids = np.where(sentinel_ids != 0, (sentinel_ids + self.tokenizer.vocab_size - 1), 0)
+        sentinel_ids -= mask_indices - start_indices
+
+        return sentinel_ids
+
+    def filter_input_ids(self, input_ids, sentinel_ids):
+        """
+        Puts sentinel mask on `input_ids` and fuse consecutive mask tokens into a single mask token by deleting.
+        This will reduce the sequence length from `expanded_inputs_length` to `input_length`.
+        """
+        batch_size = input_ids.shape[0]
+
+        input_ids_full = np.where(sentinel_ids != 0, sentinel_ids, input_ids)
+        input_ids = input_ids_full[input_ids_full > 0].reshape((batch_size, -1))
+        input_ids = np.concatenate(
+            [input_ids, np.full((batch_size, 1), self.tokenizer.eos_token_id, dtype=np.int32)], axis=-1
+        )
+        return input_ids
+
+    def random_spans_noise_mask(self, length):
+
+        """This function is copy of `random_spans_helper <https://github.com/google-research/text-to-text-transfer-transformer/blob/84f8bcc14b5f2c03de51bd3587609ba8f6bbd1cd/t5/data/preprocessors.py#L2682>`__ .
+
+        Noise mask consisting of random spans of noise tokens.
+        The number of noise tokens and the number of noise spans and non-noise spans
+        are determined deterministically as follows:
+        num_noise_tokens = round(length * noise_density)
+        num_nonnoise_spans = num_noise_spans = round(num_noise_tokens / mean_noise_span_length)
+        Spans alternate between non-noise and noise, beginning with non-noise.
+        Subject to the above restrictions, all masks are equally likely.
+
+        Args:
+            length: an int32 scalar (length of the incoming token sequence)
+            noise_density: a float - approximate density of output mask
+            mean_noise_span_length: a number
+
+        Returns:
+            a boolean tensor with shape [length]
+        """
+
+        orig_length = length
+
+        num_noise_tokens = int(np.round(length * self.noise_density))
+        # avoid degeneracy by ensuring positive numbers of noise and nonnoise tokens.
+        num_noise_tokens = min(max(num_noise_tokens, 1), length - 1)
+        num_noise_spans = int(np.round(num_noise_tokens / self.mean_noise_span_length))
+
+        # avoid degeneracy by ensuring positive number of noise spans
+        num_noise_spans = max(num_noise_spans, 1)
+        num_nonnoise_tokens = length - num_noise_tokens
+
+        # pick the lengths of the noise spans and the non-noise spans
+        def _random_segmentation(num_items, num_segments):
+            """Partition a sequence of items randomly into non-empty segments.
+            Args:
+                num_items: an integer scalar > 0
+                num_segments: an integer scalar in [1, num_items]
+            Returns:
+                a Tensor with shape [num_segments] containing positive integers that add
+                up to num_items
+            """
+            mask_indices = np.arange(num_items - 1) < (num_segments - 1)
+            np.random.shuffle(mask_indices)
+            first_in_segment = np.pad(mask_indices, [[1, 0]])
+            segment_id = np.cumsum(first_in_segment)
+            segment_length = np.asarray(jax.ops.segment_sum(np.ones_like(segment_id), segment_id))
+            return segment_length
+
+        noise_span_lengths = _random_segmentation(num_noise_tokens, num_noise_spans)
+        nonnoise_span_lengths = _random_segmentation(num_nonnoise_tokens, num_noise_spans)
+
+        interleaved_span_lengths = np.reshape(
+            np.stack([nonnoise_span_lengths, noise_span_lengths], axis=1), [num_noise_spans * 2]
+        )
+        span_starts = np.cumsum(interleaved_span_lengths)[:-1]
+        span_start_indicator = np.zeros((length,), dtype=np.int8)
+        span_start_indicator[span_starts] = True
+        span_num = np.cumsum(span_start_indicator)
+        is_noise = np.equal(span_num % 2, 1)
+
+        return is_noise[:orig_length]
+
+
+def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
+    num_samples = len(samples_idx)
+    samples_to_remove = num_samples % batch_size
+
+    if samples_to_remove != 0:
+        samples_idx = samples_idx[:-samples_to_remove]
+    sections_split = num_samples // batch_size
+    batch_idx = np.split(samples_idx, sections_split)
+    return batch_idx
+
+
+def write_train_metric(summary_writer, train_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+
+def write_eval_metric(summary_writer, eval_metrics, step):
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+if __name__ == "__main__":
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        level="NOTSET",
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+
+        if "validation" not in datasets.keys():
+            datasets["validation"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[:{data_args.validation_split_percentage}%]",
+                cache_dir=model_args.cache_dir,
+            )
+            datasets["train"] = load_dataset(
+                extension,
+                data_files=data_files,
+                split=f"train[{data_args.validation_split_percentage}%:]",
+                cache_dir=model_args.cache_dir,
+            )
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    if model_args.config_name:
+        config = T5Config.from_pretrained(
+            model_args.config_name, cache_dir=model_args.cache_dir, vocab_size=len(tokenizer)
+        )
+    elif model_args.model_name_or_path:
+        config = T5Config.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, vocab_size=len(tokenizer)
+        )
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+    # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+    # Since we make sure that all sequences are of the same length, no attention_mask is needed.
+    def tokenize_function(examples):
+        return tokenizer(examples[text_column_name], return_attention_mask=False)
+
+    tokenized_datasets = datasets.map(
+        tokenize_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        remove_columns=column_names,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    # T5-like span masked language modeling will fuse consecutively masked tokens to a single sentinel token.
+    # To ensure that the input length is `max_seq_length`, we need to increase the maximum length
+    # according to `mlm_probability` and `mean_noise_span_length`. We can also define the label length accordingly.
+    expanded_inputs_length, targets_length = compute_input_and_target_lengths(
+        inputs_length=max_seq_length,
+        noise_density=data_args.mlm_probability,
+        mean_noise_span_length=data_args.mean_noise_span_length,
+    )
+
+    # Main data processing function that will concatenate all texts from our dataset and generate chunks of expanded_inputs_length.
+    def group_texts(examples):
+        # Concatenate all texts.
+        concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
+        total_length = len(concatenated_examples[list(examples.keys())[0]])
+        # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+        # customize this part to your needs.
+        if total_length >= expanded_inputs_length:
+            total_length = (total_length // expanded_inputs_length) * expanded_inputs_length
+        # Split by chunks of max_len.
+        result = {
+            k: [t[i : i + expanded_inputs_length] for i in range(0, total_length, expanded_inputs_length)]
+            for k, t in concatenated_examples.items()
+        }
+        return result
+
+    # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
+    # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
+    # might be slower to preprocess.
+    #
+    # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+    # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+    tokenized_datasets = tokenized_datasets.map(
+        group_texts,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+
+            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+    else:
+        logger.warning(
+            "Unable to display metrics through TensorBoard because the package is not installed: "
+            "Please run pip install tensorboard to enable."
+        )
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    dropout_rngs = jax.random.split(rng, jax.local_device_count())
+
+    if model_args.model_name_or_path:
+        model = FlaxT5ForConditionalGeneration.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxT5ForConditionalGeneration(config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype))
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = FlaxDataCollatorForT5MLM(
+        tokenizer=tokenizer,
+        noise_density=data_args.mlm_probability,
+        mean_noise_span_length=data_args.mean_noise_span_length,
+        input_length=max_seq_length,
+        target_length=targets_length,
+        pad_token_id=model.config.pad_token_id,
+        decoder_start_token_id=model.config.decoder_start_token_id,
+    )
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+
+    num_train_steps = len(tokenized_datasets["train"]) // train_batch_size * num_epochs
+
+    # Create learning rate schedule
+    warmup_fn = optax.linear_schedule(
+        init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
+    )
+    decay_fn = optax.linear_schedule(
+        init_value=training_args.learning_rate,
+        end_value=0,
+        transition_steps=num_train_steps - training_args.warmup_steps,
+    )
+    linear_decay_lr_schedule_fn = optax.join_schedules(
+        schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
+    )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {
+            path: (path[-1] != "bias" and path[-2:] not in [("layer_norm", "scale"), ("final_layer_norm", "scale")])
+            for path in flat_params
+        }
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
+
+    # Setup train state
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
+
+    # Define gradient update step fn
+    def train_step(state, batch, dropout_rng):
+        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+        def loss_fn(params):
+            labels = batch.pop("labels")
+
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+
+            # compute loss
+            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])).mean()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(loss_fn)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+        new_state = state.apply_gradients(grads=grad)
+
+        metrics = jax.lax.pmean(
+            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
+        )
+
+        return new_state, metrics, new_dropout_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+
+        logits = model(**batch, params=params, train=False)[0]
+
+        # compute loss
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1]))
+
+        # compute accuracy
+        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels)
+
+        # summarize metrics
+        metrics = {"loss": loss.mean(), "accuracy": accuracy.mean()}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return metrics
+
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+
+    train_time = 0
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+        train_metrics = []
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+
+        # 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))
+        train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size)
+
+        # Gather the indexes for creating the batch and do a training step
+        for step, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1)):
+            samples = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples)
+
+            # Model forward
+            model_inputs = shard(model_inputs.data)
+            state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+            train_metrics.append(train_metric)
+
+            cur_step = epoch * (num_train_samples // train_batch_size) + step
+
+            if cur_step % training_args.logging_steps == 0 and cur_step > 0:
+                # Save metrics
+                train_metric = jax_utils.unreplicate(train_metric)
+                train_time += time.time() - train_start
+                if has_tensorboard and jax.process_index() == 0:
+                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
+
+                epochs.write(
+                    f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
+                )
+
+                train_metrics = []
+
+            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)
+                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+                eval_metrics = []
+                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+                    samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+                    model_inputs = data_collator(samples)
+
+                    # Model forward
+                    model_inputs = shard(model_inputs.data)
+                    metrics = p_eval_step(state.params, model_inputs)
+                    eval_metrics.append(metrics)
+
+                # get eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
+
+                # Update progress bar
+                epochs.write(f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})")
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and cur_step > 0:
+                # save checkpoint after each epoch and push checkpoint to the hub
+                if jax.process_index() == 0:
+                    params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+                    model.save_pretrained(
+                        training_args.output_dir,
+                        params=params,
+                        push_to_hub=training_args.push_to_hub,
+                        commit_message=f"Saving weights and logs of step {cur_step}",
+                    )

examples/flax/language-modeling/t5_tokenizer_model.py

@@ -0,0 +1,112 @@
+#!/usr/bin/env python3
+import json
+from typing import Iterator, List, Union
+
+from tokenizers import AddedToken, Regex, Tokenizer, decoders, normalizers, pre_tokenizers, trainers
+from tokenizers.implementations.base_tokenizer import BaseTokenizer
+from tokenizers.models import Unigram
+from tokenizers.processors import TemplateProcessing
+
+
+class SentencePieceUnigramTokenizer(BaseTokenizer):
+    """
+    This class is a copy of `DeDLOC's tokenizer implementation <https://github.com/yandex-research/DeDLOC/blob/main/sahajbert/tokenizer/tokenizer_model.py>`__ .
+
+    Custom SentencePiece Unigram Tokenizer with NMT, NKFC, spaces and lower-casing characters normalization
+    Represents the Unigram algorithm, with the pretokenization used by SentencePiece
+    """
+
+    def __init__(
+        self,
+        replacement: str = "▁",
+        add_prefix_space: bool = True,
+        unk_token: Union[str, AddedToken] = "<unk>",
+        eos_token: Union[str, AddedToken] = "</s>",
+        pad_token: Union[str, AddedToken] = "<pad>",
+    ):
+        self.special_tokens = {
+            "pad": {"id": 0, "token": pad_token},
+            "eos": {"id": 1, "token": eos_token},
+            "unk": {"id": 2, "token": unk_token},
+        }
+
+        self.special_tokens_list = [None] * len(self.special_tokens)
+        for token_dict in self.special_tokens.values():
+            self.special_tokens_list[token_dict["id"]] = token_dict["token"]
+
+        tokenizer = Tokenizer(Unigram())
+
+        tokenizer.normalizer = normalizers.Sequence(
+            [
+                normalizers.Nmt(),
+                normalizers.NFKC(),
+                normalizers.Replace(Regex(" {2,}"), " "),
+                normalizers.Lowercase(),
+            ]
+        )
+        tokenizer.pre_tokenizer = pre_tokenizers.Sequence(
+            [
+                pre_tokenizers.Metaspace(replacement=replacement, add_prefix_space=add_prefix_space),
+                pre_tokenizers.Digits(individual_digits=True),
+                pre_tokenizers.Punctuation(),
+            ]
+        )
+        tokenizer.decoder = decoders.Metaspace(replacement=replacement, add_prefix_space=add_prefix_space)
+
+        tokenizer.post_processor = TemplateProcessing(
+            single=f"$A {self.special_tokens['eos']['token']}",
+            special_tokens=[(self.special_tokens["eos"]["token"], self.special_tokens["eos"]["id"])],
+        )
+
+        parameters = {
+            "model": "SentencePieceUnigram",
+            "replacement": replacement,
+            "add_prefix_space": add_prefix_space,
+        }
+
+        super().__init__(tokenizer, parameters)
+
+    def train(
+        self,
+        files: Union[str, List[str]],
+        vocab_size: int = 8000,
+        show_progress: bool = True,
+    ):
+        """Train the model using the given files"""
+
+        trainer = trainers.UnigramTrainer(
+            vocab_size=vocab_size,
+            special_tokens=self.special_tokens_list,
+            show_progress=show_progress,
+        )
+
+        if isinstance(files, str):
+            files = [files]
+        self._tokenizer.train(files, trainer=trainer)
+
+        self.add_unk_id()
+
+    def train_from_iterator(
+        self,
+        iterator: Union[Iterator[str], Iterator[Iterator[str]]],
+        vocab_size: int = 8000,
+        show_progress: bool = True,
+    ):
+        """Train the model using the given iterator"""
+
+        trainer = trainers.UnigramTrainer(
+            vocab_size=vocab_size,
+            special_tokens=self.special_tokens_list,
+            show_progress=show_progress,
+        )
+
+        self._tokenizer.train_from_iterator(iterator, trainer=trainer)
+
+        self.add_unk_id()
+
+    def add_unk_id(self):
+        tokenizer_json = json.loads(self._tokenizer.to_str())
+
+        tokenizer_json["model"]["unk_id"] = self.special_tokens["unk"]["id"]
+
+        self._tokenizer = Tokenizer.from_str(json.dumps(tokenizer_json))

examples/flax/summarization/README.md

@@ -0,0 +1,66 @@
+# Summarization (Seq2Seq model) training examples
+
+The following example showcases how to finetune a sequence-to-sequence model for summarization
+using the JAX/Flax backend.
+
+JAX/Flax allows you to trace pure functions and compile them into efficient, fused accelerator code on both GPU and TPU.
+Models written in JAX/Flax are **immutable** and updated in a purely functional
+way which enables simple and efficient model parallelism.
+
+`run_summarization_flax.py` is a lightweight example of how to download and preprocess a dataset from the 🤗 Datasets library or use your own files (jsonlines or csv), then fine-tune one of the architectures above on it.
+
+For custom datasets in `jsonlines` format please see: https://huggingface.co/docs/datasets/loading_datasets.html#json-files and you also will find examples of these below.
+
+Let's start by creating a model repository to save the trained model and logs.
+Here we call the model `"bart-base-xsum"`, but you can change the model name as you like.
+
+You can do this either directly on [huggingface.co](https://huggingface.co/new) (assuming that
+you are logged in) or via the command line:
+
+```
+huggingface-cli repo create bart-base-xsum
+```
+Next we clone the model repository to add the tokenizer and model files.
+```
+git clone https://huggingface.co/<your-username>/bart-base-xsum
+```
+To ensure that all tensorboard traces will be uploaded correctly, we need to 
+track them. You can run the following command inside your model repo to do so.
+
+```
+cd bart-base-xsum
+git lfs track "*tfevents*"
+```
+
+Great, we have set up our model repository. During training, we will automatically
+push the training logs and model weights to the repo.
+
+Next, let's add a symbolic link to the `run_summarization_flax.py`.
+
+```bash
+export MODEL_DIR="./bart-base-xsum"
+ln -s ~/transformers/examples/flax/summarization/run_summarization_flax.py run_summarization_flax.py
+```
+
+### Train the model
+Next we can run the example script to train the model:
+
+```bash
+python run_summarization_flax.py \
+	--output_dir ${MODEL_DIR} \
+	--model_name_or_path facebook/bart-base \
+	--tokenizer_name facebook/bart-base \
+	--dataset_name="xsum" \
+	--do_train --do_eval --do_predict --predict_with_generate \
+	--num_train_epochs 6 \
+	--learning_rate 5e-5 --warmup_steps 0 \
+	--per_device_train_batch_size 64 \
+	--per_device_eval_batch_size 64 \
+	--overwrite_output_dir \
+	--max_source_length 512 --max_target_length 64 \
+	--push_to_hub
+```
+
+This should finish in 37min, with validation loss and ROUGE2 score of 1.7785 and 17.01 respectively after 6 epochs. training statistics can be accessed on [tfhub.de](https://tensorboard.dev/experiment/OcPfOIgXRMSJqYB4RdK2tA/#scalars).
+
+> Note that here we used default `generate` arguments, using arguments specific for `xsum` dataset should give better ROUGE scores.  

examples/flax/summarization/requirements.txt

@@ -0,0 +1,5 @@
+datasets >= 1.1.3
+jax>=0.2.8
+jaxlib>=0.1.59
+flax>=0.3.4
+optax>=0.0.8

examples/flax/summarization/run_summarization_flax.py

@@ -0,0 +1,812 @@
+#!/usr/bin/env python
+# coding=utf-8
+# 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.
+"""
+Fine-tuning the library models for summarization.
+"""
+# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
+
+import logging
+import os
+import sys
+import time
+from dataclasses import dataclass, field
+from functools import partial
+from pathlib import Path
+from typing import Callable, Optional
+
+import datasets
+import nltk  # Here to have a nice missing dependency error message early on
+import numpy as np
+from datasets import Dataset, load_dataset, load_metric
+from tqdm import tqdm
+
+import jax
+import jax.numpy as jnp
+import optax
+import transformers
+from filelock import FileLock
+from flax import jax_utils, traverse_util
+from flax.jax_utils import unreplicate
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxAutoModelForSeq2SeqLM,
+    HfArgumentParser,
+    TrainingArguments,
+    is_tensorboard_available,
+)
+from transformers.file_utils import is_offline_mode
+
+
+logger = logging.getLogger(__name__)
+
+try:
+    nltk.data.find("tokenizers/punkt")
+except (LookupError, OSError):
+    if is_offline_mode():
+        raise LookupError(
+            "Offline mode: run this script without TRANSFORMERS_OFFLINE first to download nltk data files"
+        )
+    with FileLock(".lock") as lock:
+        nltk.download("punkt", quiet=True)
+
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    text_column: Optional[str] = field(
+        default=None,
+        metadata={"help": "The name of the column in the datasets containing the full texts (for summarization)."},
+    )
+    summary_column: Optional[str] = field(
+        default=None,
+        metadata={"help": "The name of the column in the datasets containing the summaries (for summarization)."},
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    test_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input predict data file to do prediction on (a text file)."},
+    )
+    max_source_length: Optional[int] = field(
+        default=1024,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated, sequences shorter will be padded."
+        },
+    )
+    max_target_length: Optional[int] = field(
+        default=128,
+        metadata={
+            "help": "The maximum total sequence length for target text after tokenization. Sequences longer "
+            "than this will be truncated, sequences shorter will be padded."
+        },
+    )
+    val_max_target_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total sequence length for validation target text after tokenization. Sequences longer "
+            "than this will be truncated, sequences shorter will be padded. Will default to `max_target_length`."
+            "This argument is also used to override the `max_length` param of `model.generate`, which is used "
+            "during evaluation."
+        },
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+            "value if set."
+        },
+    )
+    max_predict_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of prediction examples to this "
+            "value if set."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    source_prefix: Optional[str] = field(
+        default=None, metadata={"help": "A prefix to add before every source text (useful for T5 models)."}
+    )
+    predict_with_generate: bool = field(
+        default=False, metadata={"help": "Whether to use generate to calculate generative metrics (ROUGE, BLEU)."}
+    )
+    num_beams: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Number of beams to use for evaluation. This argument will be passed to `model.generate`, "
+            "which is used during evaluation."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."
+        if self.val_max_target_length is None:
+            self.val_max_target_length = self.max_target_length
+
+
+summarization_name_mapping = {
+    "amazon_reviews_multi": ("review_body", "review_title"),
+    "big_patent": ("description", "abstract"),
+    "cnn_dailymail": ("article", "highlights"),
+    "orange_sum": ("text", "summary"),
+    "pn_summary": ("article", "summary"),
+    "psc": ("extract_text", "summary_text"),
+    "samsum": ("dialogue", "summary"),
+    "thaisum": ("body", "summary"),
+    "xglue": ("news_body", "news_title"),
+    "xsum": ("document", "summary"),
+    "wiki_summary": ("article", "highlights"),
+}
+
+
+class TrainState(train_state.TrainState):
+    dropout_rng: jnp.ndarray
+
+    def replicate(self):
+        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
+
+
+def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False):
+    """
+    Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
+    Shuffle batches if `shuffle` is `True`.
+    """
+    steps_per_epoch = len(dataset) // batch_size
+
+    if shuffle:
+        batch_idx = jax.random.permutation(rng, len(dataset))
+    else:
+        batch_idx = jnp.arange(len(dataset))
+
+    batch_idx = batch_idx[: steps_per_epoch * batch_size]  # Skip incomplete batch.
+    batch_idx = batch_idx.reshape((steps_per_epoch, batch_size))
+
+    for idx in batch_idx:
+        batch = dataset[idx]
+        batch = {k: jnp.array(v) for k, v in batch.items()}
+
+        batch = shard(batch)
+
+        yield batch
+
+
+def write_metric(summary_writer, train_metrics, eval_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+def create_learning_rate_fn(
+    train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, linear_decay learning rate function."""
+    steps_per_epoch = train_ds_size // train_batch_size
+    num_train_steps = steps_per_epoch * num_train_epochs
+    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
+    decay_fn = optax.linear_schedule(
+        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
+    )
+    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
+    return schedule_fn
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files this script will use the first column for the full texts and the second column for the
+    # summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments).
+    #
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        dataset = load_dataset(
+            data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir, keep_in_memory=False
+        )
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+            extension = data_args.train_file.split(".")[-1]
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+            extension = data_args.validation_file.split(".")[-1]
+        if data_args.test_file is not None:
+            data_files["test"] = data_args.test_file
+            extension = data_args.test_file.split(".")[-1]
+        dataset = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForSeq2SeqLM.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForSeq2SeqLM.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    if model.config.decoder_start_token_id is None:
+        raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
+
+    prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
+
+    # Preprocessing the datasets.
+    # We need to tokenize inputs and targets.
+    if training_args.do_train:
+        column_names = dataset["train"].column_names
+    elif training_args.do_eval:
+        column_names = dataset["validation"].column_names
+    elif training_args.do_predict:
+        column_names = dataset["test"].column_names
+    else:
+        logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.")
+        return
+
+    # Get the column names for input/target.
+    dataset_columns = summarization_name_mapping.get(data_args.dataset_name, None)
+    if data_args.text_column is None:
+        text_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
+    else:
+        text_column = data_args.text_column
+        if text_column not in column_names:
+            raise ValueError(
+                f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}"
+            )
+    if data_args.summary_column is None:
+        summary_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
+    else:
+        summary_column = data_args.summary_column
+        if summary_column not in column_names:
+            raise ValueError(
+                f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}"
+            )
+
+    # Temporarily set max_target_length for training.
+    max_target_length = data_args.max_target_length
+
+    # In Flax, for seq2seq models we need to pass `decoder_input_ids`
+    # as the Flax models don't accept `labels`, we need to prepare the decoder_input_ids here
+    # for that dynamically import the `shift_tokens_right` function from the model file
+    model_module = __import__(model.__module__, fromlist=["shift_tokens_tight"])
+    shift_tokens_right_fn = getattr(model_module, "shift_tokens_right")
+
+    # Setting padding="max_length" as we need fixed length inputs for jitted functions
+    def preprocess_function(examples):
+        inputs = examples[text_column]
+        targets = examples[summary_column]
+        inputs = [prefix + inp for inp in inputs]
+        model_inputs = tokenizer(
+            inputs, max_length=data_args.max_source_length, padding="max_length", truncation=True, return_tensors="np"
+        )
+
+        # Setup the tokenizer for targets
+        with tokenizer.as_target_tokenizer():
+            labels = tokenizer(
+                targets, max_length=max_target_length, padding="max_length", truncation=True, return_tensors="np"
+            )
+
+        model_inputs["labels"] = labels["input_ids"]
+        decoder_input_ids = shift_tokens_right_fn(
+            jnp.array(labels["input_ids"]), config.pad_token_id, config.decoder_start_token_id
+        )
+        model_inputs["decoder_input_ids"] = np.asarray(decoder_input_ids)
+
+        # We need decoder_attention_mask so we can ignore pad tokens from loss
+        model_inputs["decoder_attention_mask"] = labels["attention_mask"]
+
+        return model_inputs
+
+    if training_args.do_train:
+        if "train" not in dataset:
+            raise ValueError("--do_train requires a train dataset")
+        train_dataset = dataset["train"]
+        if data_args.max_train_samples is not None:
+            train_dataset = train_dataset.select(range(data_args.max_train_samples))
+        train_dataset = train_dataset.map(
+            preprocess_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+            desc="Running tokenizer on train dataset",
+        )
+
+    if training_args.do_eval:
+        max_target_length = data_args.val_max_target_length
+        if "validation" not in dataset:
+            raise ValueError("--do_eval requires a validation dataset")
+        eval_dataset = dataset["validation"]
+        if data_args.max_eval_samples is not None:
+            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
+        eval_dataset = eval_dataset.map(
+            preprocess_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+            desc="Running tokenizer on validation dataset",
+        )
+
+    if training_args.do_predict:
+        max_target_length = data_args.val_max_target_length
+        if "test" not in dataset:
+            raise ValueError("--do_predict requires a test dataset")
+        predict_dataset = dataset["test"]
+        if data_args.max_predict_samples is not None:
+            predict_dataset = predict_dataset.select(range(data_args.max_predict_samples))
+        predict_dataset = predict_dataset.map(
+            preprocess_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+            desc="Running tokenizer on prediction dataset",
+        )
+
+    # Metric
+    metric = load_metric("rouge")
+
+    def postprocess_text(preds, labels):
+        preds = [pred.strip() for pred in preds]
+        labels = [label.strip() for label in labels]
+
+        # rougeLSum expects newline after each sentence
+        preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
+        labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
+
+        return preds, labels
+
+    def compute_metrics(preds, labels):
+        decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
+        decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
+
+        # Some simple post-processing
+        decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
+
+        result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)
+        # Extract a few results from ROUGE
+        result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
+
+        prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
+        result["gen_len"] = np.mean(prediction_lens)
+        result = {k: round(v, 4) for k, v in result.items()}
+        return result
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+
+            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+    else:
+        logger.warning(
+            "Unable to display metrics through TensorBoard because the package is not installed: "
+            "Please run pip install tensorboard to enable."
+        )
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    rng, dropout_rng = jax.random.split(rng)
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+    steps_per_epoch = len(train_dataset) // train_batch_size
+    total_train_steps = steps_per_epoch * num_epochs
+
+    # Create learning rate schedule
+    linear_decay_lr_schedule_fn = create_learning_rate_fn(
+        len(train_dataset),
+        train_batch_size,
+        training_args.num_train_epochs,
+        training_args.warmup_steps,
+        training_args.learning_rate,
+    )
+
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    # Note that this mask is specifically adapted for FlaxBart.
+    # For FlaxT5, one should correct the layer norm parameter naming
+    # accordingly - see `run_t5_mlm_flax.py` e.g.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        layer_norm_params = [
+            (name, "scale") for name in ["self_attn_layer_norm", "layernorm_embedding", "final_layer_norm"]
+        ]
+        flat_mask = {path: (path[-1] != "bias" and path[-2:] not in layer_norm_params) for path in flat_params}
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    adamw = optax.adamw(
+        learning_rate=linear_decay_lr_schedule_fn,
+        b1=training_args.adam_beta1,
+        b2=training_args.adam_beta2,
+        eps=training_args.adam_epsilon,
+        weight_decay=training_args.weight_decay,
+        mask=decay_mask_fn,
+    )
+
+    # Setup train state
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw, dropout_rng=dropout_rng)
+
+    # label smoothed cross entropy
+    def loss_fn(logits, labels, padding_mask, label_smoothing_factor=0.0):
+        """
+        The label smoothing implementation is adapted from Flax's official example:
+        https://github.com/google/flax/blob/87a211135c6a377c8f29048a1cac3840e38b9da4/examples/wmt/train.py#L104
+        """
+        vocab_size = logits.shape[-1]
+        confidence = 1.0 - label_smoothing_factor
+        low_confidence = (1.0 - confidence) / (vocab_size - 1)
+        normalizing_constant = -(
+            confidence * jnp.log(confidence) + (vocab_size - 1) * low_confidence * jnp.log(low_confidence + 1e-20)
+        )
+        soft_labels = onehot(labels, vocab_size, on_value=confidence, off_value=low_confidence)
+
+        loss = optax.softmax_cross_entropy(logits, soft_labels)
+        loss = loss - normalizing_constant
+
+        # ignore padded tokens from loss
+        loss = loss * padding_mask
+        loss = loss.sum() / padding_mask.sum()
+        return loss
+
+    # Define gradient update step fn
+    def train_step(state, batch, label_smoothing_factor=0.0):
+        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+
+        def compute_loss(params):
+            labels = batch.pop("labels")
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+            loss = loss_fn(logits, labels, batch["decoder_attention_mask"], label_smoothing_factor)
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+
+        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
+
+        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_state, metrics
+
+    # Define eval fn
+    def eval_step(params, batch, label_smoothing_factor=0.0):
+        labels = batch.pop("labels")
+        logits = model(**batch, params=params, train=False)[0]
+        loss = loss_fn(logits, labels, batch["decoder_attention_mask"], label_smoothing_factor)
+
+        # summarize metrics
+        metrics = {"loss": loss}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+        return metrics
+
+    # Define generation function
+    max_length = (
+        data_args.val_max_target_length if data_args.val_max_target_length is not None else model.config.max_length
+    )
+    num_beams = data_args.num_beams if data_args.num_beams is not None else model.config.num_beams
+    gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
+
+    def generate_step(params, batch):
+        model.params = params
+        output_ids = model.generate(batch["input_ids"], attention_mask=batch["attention_mask"], **gen_kwargs)
+        return output_ids.sequences
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(
+        partial(train_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch", donate_argnums=(0,)
+    )
+    p_eval_step = jax.pmap(partial(eval_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch")
+    p_generate_step = jax.pmap(generate_step, "batch")
+
+    # Replicate the train state on each device
+    state = state.replicate()
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {num_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    train_time = 0
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+        train_metrics = []
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True)
+        steps_per_epoch = len(train_dataset) // train_batch_size
+        # train
+        for _ in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False):
+            batch = next(train_loader)
+            state, train_metric = p_train_step(state, batch)
+            train_metrics.append(train_metric)
+
+        train_time += time.time() - train_start
+
+        train_metric = unreplicate(train_metric)
+
+        epochs.write(
+            f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+        )
+
+        # ======================== Evaluating ==============================
+        eval_metrics = []
+        eval_preds = []
+        eval_labels = []
+
+        eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
+        eval_steps = len(eval_dataset) // eval_batch_size
+        for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
+            # Model forward
+            batch = next(eval_loader)
+            labels = batch["labels"]
+
+            metrics = p_eval_step(state.params, batch)
+            eval_metrics.append(metrics)
+
+            # generation
+            if data_args.predict_with_generate:
+                generated_ids = p_generate_step(state.params, batch)
+                eval_preds.extend(jax.device_get(generated_ids.reshape(-1, gen_kwargs["max_length"])))
+                eval_labels.extend(jax.device_get(labels.reshape(-1, labels.shape[-1])))
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
+
+        # compute ROUGE metrics
+        rouge_desc = ""
+        if data_args.predict_with_generate:
+            rouge_metrics = compute_metrics(eval_preds, eval_labels)
+            eval_metrics.update(rouge_metrics)
+            rouge_desc = " ".join([f"Eval {key}: {value} |" for key, value in rouge_metrics.items()])
+
+        # Print metrics and update progress bar
+        desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']} | {rouge_desc})"
+        epochs.write(desc)
+        epochs.desc = desc
+
+        # Save metrics
+        if has_tensorboard and jax.process_index() == 0:
+            cur_step = epoch * (len(train_dataset) // train_batch_size)
+            write_metric(summary_writer, train_metrics, eval_metrics, train_time, cur_step)
+
+    # ======================== Prediction loop ==============================
+    if training_args.do_predict:
+        logger.info("*** Predict ***")
+
+        pred_metrics = []
+        pred_generations = []
+        pred_labels = []
+
+        pred_loader = data_loader(input_rng, predict_dataset, eval_batch_size)
+        pred_steps = len(predict_dataset) // eval_batch_size
+        for _ in tqdm(range(pred_steps), desc="Predicting...", position=2, leave=False):
+            # Model forward
+            batch = next(pred_loader)
+            labels = batch["labels"]
+
+            metrics = p_eval_step(state.params, batch)
+            pred_metrics.append(metrics)
+
+            # generation
+            if data_args.predict_with_generate:
+                generated_ids = p_generate_step(state.params, batch)
+                pred_generations.extend(jax.device_get(generated_ids.reshape(-1, gen_kwargs["max_length"])))
+                pred_labels.extend(jax.device_get(labels.reshape(-1, labels.shape[-1])))
+
+        # normalize prediction metrics
+        pred_metrics = get_metrics(pred_metrics)
+        pred_metrics = jax.tree_map(jnp.mean, pred_metrics)
+
+        # compute ROUGE metrics
+        rouge_desc = ""
+        if data_args.predict_with_generate:
+            rouge_metrics = compute_metrics(pred_generations, pred_labels)
+            pred_metrics.update(rouge_metrics)
+            rouge_desc = " ".join([f"Predict {key}: {value} |" for key, value in rouge_metrics.items()])
+
+        # Print metrics
+        desc = f"Predict Loss: {pred_metrics['loss']} | {rouge_desc})"
+        logger.info(desc)
+
+        # save checkpoint after each epoch and push checkpoint to the hub
+        if jax.process_index() == 0:
+            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+            model.save_pretrained(
+                training_args.output_dir,
+                params=params,
+                push_to_hub=training_args.push_to_hub,
+                commit_message=f"Saving weights and logs of epoch {epoch+1}",
+            )
+
+
+if __name__ == "__main__":
+    main()

examples/flax/text-classification/README.md

@@ -23,31 +23,68 @@ Based on the script [`run_flax_glue.py`](https://github.com/huggingface/transfor
 Fine-tuning the library models for sequence classification on the GLUE benchmark: [General Language Understanding
 Evaluation](https://gluebenchmark.com/). This script can fine-tune any of the models on the [hub](https://huggingface.co/models).
 
-GLUE is made up of a total of 9 different tasks. Here is how to run the script on one of them:
+To begin with it is recommended to create a model repository to save the trained model and logs.
+Here we call the model `"bert-glue-mrpc-test"`, but you can change the model name as you like.
+
+You can do this either directly on [huggingface.co](https://huggingface.co/new) (assuming that
+you are logged in) or via the command line:
+
+```
+huggingface-cli repo create bert-glue-mrpc-test
+```
+
+Next we clone the model repository to add the tokenizer and model files.
+
+```
+git clone https://huggingface.co/<your-username>/bert-glue-mrpc-test
+```
+
+To ensure that all tensorboard traces will be uploaded correctly, we need to 
+track them. You can run the following command inside your model repo to do so.
+
+```
+cd bert-glue-mrpc-test
+git lfs track "*tfevents*"
+```
+
+Great, we have set up our model repository. During training, we will automatically
+push the training logs and model weights to the repo.
+
+Next, let's add a symbolic link to the `run_flax_glue.py`.
 
 ```bash
 export TASK_NAME=mrpc
+export MODEL_DIR="./bert-glue-mrpc-test"
+ln -s ~/transformers/examples/flax/text-classification/run_flax_glue.py run_flax_glue.py
+```
 
+
+GLUE is made up of a total of 9 different tasks. Here is how to run the script on one of them:
+
+```bash
 python run_flax_glue.py \
   --model_name_or_path bert-base-cased \
-  --task_name $TASK_NAME \
+  --task_name ${TASK_NAME} \
   --max_length 128 \
   --learning_rate 2e-5 \
   --num_train_epochs 3 \
   --per_device_train_batch_size 4 \
-  --output_dir /tmp/$TASK_NAME/
+  --output_dir ${MODEL_DIR} \
+  --push_to_hub
 ```
 
 where task name can be one of cola, mnli, mnli-mm, mrpc, qnli, qqp, rte, sst2, stsb, wnli.
 
 Using the command above, the script will train for 3 epochs and run eval after each epoch. 
-Metrics and hyperparameters are stored in Tensorflow event files in `---output_dir`.
+Metrics and hyperparameters are stored in Tensorflow event files in `--output_dir`.
 You can see the results by running `tensorboard` in that directory:
 
 ```bash
 $ tensorboard --logdir .
 ```
 
+or directly on the hub under *Training metrics*.
+
 ### Accuracy Evaluation
 
 We train five replicas and report mean accuracy and stdev on the dev set below.
@@ -59,20 +96,19 @@ On the task other than MRPC and WNLI we train for 3 these epochs because this is
 but looking at the training curves of some of them (e.g., SST-2, STS-b), it appears the models
 are undertrained and we could get better results when training longer.
 
-In the Tensorboard results linked below, the random seed of each model is equal to the ID of the run. So in order to reproduce run 1, run the command above with `--seed=1`. The best run used random seed 2, which is the default in the script. The results of all runs are in [this Google Sheet](https://docs.google.com/spreadsheets/d/1zKL_xn32HwbxkFMxB3ftca-soTHAuBFgIhYhOhCnZ4E/edit?usp=sharing).
-
+In the Tensorboard results linked below, the random seed of each model is equal to the ID of the run. So in order to reproduce run 1, run the command above with `--seed=1`. The best run used random seed 3, which is the default in the script. The results of all runs are in [this Google Sheet](https://docs.google.com/spreadsheets/d/1p3XzReMO75m_XdEJvPue-PIq_PN-96J2IJpJW1yS-10/edit?usp=sharing).
 
 | Task  | Metric                       | Acc (best run) | Acc (avg/5runs) | Stdev     | Metrics                                                                  |
 |-------|------------------------------|----------------|-----------------|-----------|--------------------------------------------------------------------------|
-| CoLA  | Matthew's corr               | 59.57          | 58.04           | 1.81      | [tfhub.dev](https://tensorboard.dev/experiment/f4OvQpWtRq6CvddpxGBd0A/)  |
-| SST-2 | Accuracy                     | 92.43          | 91.79           | 0.59      | [tfhub.dev](https://tensorboard.dev/experiment/BYFwa49MRTaLIn93DgAEtA/)  |
-| MRPC  | F1/Accuracy                  | 89.50/84.8     | 88.70/84.02     | 0.56/0.48 | [tfhub.dev](https://tensorboard.dev/experiment/9ZWH5xwXRS6zEEUE4RaBhQ/)  |
-| STS-B | Pearson/Spearman corr.       | 90.00/88.71    | 89.09/88.61     | 0.51/0.07 | [tfhub.dev](https://tensorboard.dev/experiment/mUlI5B9QQ0WGEJip7p3Tng/)  |
-| QQP   | Accuracy/F1                  | 90.88/87.64    | 90.75/87.53     | 0.11/0.13 | [tfhub.dev](https://tensorboard.dev/experiment/pO6h75L3SvSXSWRcgljXKA/)  |
-| MNLI  | Matched acc.                 | 84.06          | 83.88           | 0.16      | [tfhub.dev](https://tensorboard.dev/experiment/LKwaOH18RMuo7nJkESrpKg/)  |
-| QNLI  | Accuracy                     | 91.01          | 90.86           | 0.18      | [tfhub.dev](https://tensorboard.dev/experiment/qesXxNcaQhmKxPmbw1sOoA/)  |
-| RTE   | Accuracy                     | 66.80          | 65.27           | 1.07      | [tfhub.dev](https://tensorboard.dev/experiment/Z84xC0r6RjyzT4SLqiAbzQ/)  |
-| WNLI  | Accuracy                     | 39.44          | 32.96           | 5.85      | [tfhub.dev](https://tensorboard.dev/experiment/gV73w9v0RIKrqVw32PZbAQ/)  |
+| CoLA  | Matthew's corr               | 60.57          | 59.04           | 1.06      | [tfhub.dev](https://tensorboard.dev/experiment/lfr2adVpRtmLDALKrElkzg/)  |
+| SST-2 | Accuracy                     | 92.66          | 92.23           | 0.57      | [tfhub.dev](https://tensorboard.dev/experiment/jYvfv2trRHKMjoWnXVwrZA/)  |
+| MRPC  | F1/Accuracy                  | 89.90/85.78    | 88.97/84.36     | 0.72/1.09 | [tfhub.dev](https://tensorboard.dev/experiment/bo3W3DEoRw2Q7YXjWrJkfg/)  |
+| STS-B | Pearson/Spearman corr.       | 89.04/88.70    | 88.94/88.63     | 0.07/0.07 | [tfhub.dev](https://tensorboard.dev/experiment/fxVwbLD7QpKhbot0r9rn2w/)  |
+| QQP   | Accuracy/F1                  | 90.81/87.58    | 90.76/87.51     | 0.05/0.06 | [tfhub.dev](https://tensorboard.dev/experiment/di089Rc9TZmsnKRMrYNLsA/)  |
+| MNLI  | Matched acc.                 | 84.10          | 83.80           | 0.16      | [tfhub.dev](https://tensorboard.dev/experiment/JgNCGHDJSRaW6HBx6YQFYQ/)  |
+| QNLI  | Accuracy                     | 91.01          | 90.82           | 0.17      | [tfhub.dev](https://tensorboard.dev/experiment/Bq7cMGJnQMSggYgL8qNGeQ/)  |
+| RTE   | Accuracy                     | 66.06          | 64.76           | 1.04      | [tfhub.dev](https://tensorboard.dev/experiment/66Eq24bhRjqN6CEhgDSGqQ/)  |
+| WNLI  | Accuracy                     | 46.48          | 37.01           | 6.83      | [tfhub.dev](https://tensorboard.dev/experiment/TAqcnddqTkWvVEeGaWwIdQ/)  |
 
 Some of these results are significantly different from the ones reported on the test set of GLUE benchmark on the
 website. For QQP and WNLI, please refer to [FAQ #12](https://gluebenchmark.com/faq) on the website.
@@ -83,24 +119,21 @@ We also ran each task once on a single V100 GPU, 8 V100 GPUs, and 8 Cloud v3 TPU
 overall training time below. For comparison we ran Pytorch's [run_glue.py](https://github.com/huggingface/transformers/blob/master/examples/pytorch/text-classification/run_glue.py) on a single GPU (last column).
 
 
-| Task  | TPU v3-8  | 8 GPU      | 1 GPU      | 1 GPU (Pytorch) |
+| Task  | TPU v3-8  | 8 GPU      | [1 GPU](https://tensorboard.dev/experiment/mkPS4Zh8TnGe1HB6Yzwj4Q)  | 1 GPU (Pytorch) |
 |-------|-----------|------------|------------|-----------------|
-| CoLA  |  1m 46s   |  1m 26s    | 3m 6s      | 4m 6s           |
-| SST-2 |  5m 30s   |  6m 28s    | 22m 6s     | 34m 37s         |
-| MRPC  |  1m 32s   |  1m 14s    | 2m 17s     | 2m 56s          |
-| STS-B |  1m 33s   |  1m 12s    | 2m 11s     | 2m 48s          |
-| QQP   | 24m 40s   | 31m 48s    | 1h 20m 15s | 2h 54m          |
-| MNLI  | 26m 30s   | 33m 55s    | 2h 7m 30s  | 3h 7m 6s        |
-| QNLI  |  8m       |  9m 40s    | 34m 20s    | 49m 8s          |
-| RTE   |  1m 21s   |     55s    | 1m 8s      | 1m 16s          |
-| WNLI  |  1m 12s   |     48s    | 38s        | 36s             |
+| CoLA  |  1m 42s   |  1m 26s    | 3m 9s      | 4m 6s           |
+| SST-2 |  5m 12s   |  6m 28s    | 22m 33s    | 34m 37s         |
+| MRPC  |  1m 29s   |  1m 14s    | 2m 20s     | 2m 56s          |
+| STS-B |  1m 30s   |  1m 12s    | 2m 16s     | 2m 48s          |
+| QQP   | 22m 50s   | 31m 48s    | 1h 59m 41s | 2h 54m          |
+| MNLI  | 25m 03s   | 33m 55s    | 2h 9m 37s  | 3h 7m 6s        |
+| QNLI  |  7m30s    |  9m 40s    | 34m 40s    | 49m 8s          |
+| RTE   |  1m 20s   |     55s    | 1m 10s     | 1m 16s          |
+| WNLI  |  1m 11s   |     48s    | 39s        | 36s             |
 |-------|
-| **TOTAL** | 1h 13m | 1h 28m | 4h 34m | 6h 37m      |
-| **COST*** | $9.60     | $29.10     | $11.33     | $16.41          |
+| **TOTAL** | 1h 03m | 1h 28m | 5h 16m | 6h 37m      |
 
-
-*All experiments are ran on Google Cloud Platform. Prices are on-demand prices
-(not preemptible), obtained from the following tables:
-[TPU pricing table](https://cloud.google.com/tpu/pricing),
-[GPU pricing table](https://cloud.google.com/compute/gpus-pricing). GPU
-experiments are ran without further optimizations besides JAX transformations.
+*All experiments are ran on Google Cloud Platform. 
+GPU experiments are ran without further optimizations besides JAX
+transformations. GPU experiments are ran with full precision (fp32). "TPU v3-8"
+are 8 TPU cores on 4 chips (each chips has 2 cores), while "8 GPU" are 8 GPU chips.

examples/flax/text-classification/requirements.txt

@@ -1,5 +1,5 @@
 datasets >= 1.1.3
 jax>=0.2.8
 jaxlib>=0.1.59
-git+https://github.com/google/flax.git
-git+https://github.com/deepmind/optax.git
+flax>=0.3.4
+optax>=0.0.8

examples/flax/text-classification/run_flax_glue.py

@@ -29,12 +29,11 @@ import jax
 import jax.numpy as jnp
 import optax
 import transformers
-from flax import linen as nn
 from flax import struct, traverse_util
 from flax.jax_utils import replicate, unreplicate
 from flax.metrics import tensorboard
 from flax.training import train_state
-from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from flax.training.common_utils import get_metrics, onehot, shard
 from transformers import AutoConfig, AutoTokenizer, FlaxAutoModelForSequenceClassification, PretrainedConfig
 
 
@@ -119,17 +118,16 @@ def parse_args():
         default=None,
         help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
     )
-    parser.add_argument(
-        "--gradient_accumulation_steps",
-        type=int,
-        default=1,
-        help="Number of updates steps to accumulate before performing a backward/update pass.",
-    )
     parser.add_argument(
         "--num_warmup_steps", type=int, default=0, help="Number of steps for the warmup in the lr scheduler."
     )
     parser.add_argument("--output_dir", type=str, default=None, help="Where to store the final model.")
-    parser.add_argument("--seed", type=int, default=2, help="A seed for reproducible training.")
+    parser.add_argument("--seed", type=int, default=3, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--push_to_hub",
+        action="store_true",
+        help="If passed, model checkpoints and tensorboard logs will be pushed to the hub",
+    )
     args = parser.parse_args()
 
     # Sanity checks
@@ -154,6 +152,7 @@ def create_train_state(
     learning_rate_fn: Callable[[int], float],
     is_regression: bool,
     num_labels: int,
+    weight_decay: float,
 ) -> train_state.TrainState:
     """Create initial training state."""
 
@@ -171,25 +170,17 @@ def create_train_state(
         logits_fn: Callable = struct.field(pytree_node=False)
         loss_fn: Callable = struct.field(pytree_node=False)
 
-    # Creates a multi-optimizer consisting of two "Adam with weight decay" optimizers.
-    def adamw(weight_decay):
-        return optax.adamw(learning_rate=learning_rate_fn, b1=0.9, b2=0.999, eps=1e-6, weight_decay=weight_decay)
+    # We use Optax's "masking" functionality to not apply weight decay
+    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
+    # mask boolean with the same structure as the parameters.
+    # The mask is True for parameters that should be decayed.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params}
+        return traverse_util.unflatten_dict(flat_mask)
 
-    def traverse(fn):
-        def mask(data):
-            flat = traverse_util.flatten_dict(data)
-            return traverse_util.unflatten_dict({k: fn(k, v) for k, v in flat.items()})
-
-        return mask
-
-    # We use Optax's "masking" functionality to create a multi-optimizer, one
-    # with weight decay and the other without. Note masking means the optimizer
-    # will ignore these paths.
-    decay_path = lambda p: not any(x in p for x in ["bias", "LayerNorm.weight"])  # noqa: E731
-
-    tx = optax.chain(
-        optax.masked(adamw(0.0), mask=traverse(lambda path, _: decay_path(path))),
-        optax.masked(adamw(0.01), mask=traverse(lambda path, _: not decay_path(path))),
+    tx = optax.adamw(
+        learning_rate=learning_rate_fn, b1=0.9, b2=0.999, eps=1e-6, weight_decay=weight_decay, mask=decay_mask_fn
     )
 
     if is_regression:
@@ -207,7 +198,6 @@ def create_train_state(
     else:  # Classification.
 
         def cross_entropy_loss(logits, labels):
-            logits = nn.log_softmax(logits)
             xentropy = optax.softmax_cross_entropy(logits, onehot(labels, num_classes=num_labels))
             return jnp.mean(xentropy)
 
@@ -264,7 +254,7 @@ def main():
 
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
         datefmt="%m/%d/%Y %H:%M:%S",
         level=logging.INFO,
     )
@@ -412,6 +402,7 @@ def main():
 
     num_epochs = int(args.num_train_epochs)
     rng = jax.random.PRNGKey(args.seed)
+    dropout_rngs = jax.random.split(rng, jax.local_device_count())
 
     train_batch_size = args.per_device_train_batch_size * jax.local_device_count()
     eval_batch_size = args.per_device_eval_batch_size * jax.local_device_count()
@@ -420,26 +411,29 @@ def main():
         len(train_dataset), train_batch_size, args.num_train_epochs, args.num_warmup_steps, args.learning_rate
     )
 
-    state = create_train_state(model, learning_rate_fn, is_regression, num_labels=num_labels)
+    state = create_train_state(
+        model, learning_rate_fn, is_regression, num_labels=num_labels, weight_decay=args.weight_decay
+    )
 
     # define step functions
     def train_step(
         state: train_state.TrainState, batch: Dict[str, Array], dropout_rng: PRNGKey
     ) -> Tuple[train_state.TrainState, float]:
         """Trains model with an optimizer (both in `state`) on `batch`, returning a pair `(new_state, loss)`."""
+        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
         targets = batch.pop("labels")
 
         def loss_fn(params):
             logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
             loss = state.loss_fn(logits, targets)
-            return loss, logits
+            return loss
 
-        grad_fn = jax.value_and_grad(loss_fn, has_aux=True)
-        (loss, logits), grad = grad_fn(state.params)
+        grad_fn = jax.value_and_grad(loss_fn)
+        loss, grad = grad_fn(state.params)
         grad = jax.lax.pmean(grad, "batch")
         new_state = state.apply_gradients(grads=grad)
         metrics = jax.lax.pmean({"loss": loss, "learning_rate": learning_rate_fn(state.step)}, axis_name="batch")
-        return new_state, metrics
+        return new_state, metrics, new_dropout_rng
 
     p_train_step = jax.pmap(train_step, axis_name="batch", donate_argnums=(0,))
 
@@ -457,27 +451,25 @@ def main():
     logger.info(f"===== Starting training ({num_epochs} epochs) =====")
     train_time = 0
 
+    # make sure weights are replicated on each device
+    state = replicate(state)
+
     for epoch in range(1, num_epochs + 1):
         logger.info(f"Epoch {epoch}")
         logger.info("  Training...")
 
-        # make sure weights are replicated on each device
-        state = replicate(state)
-
         train_start = time.time()
         train_metrics = []
-        rng, input_rng, dropout_rng = jax.random.split(rng, 3)
+        rng, input_rng = jax.random.split(rng)
 
         # train
         for batch in glue_train_data_collator(input_rng, train_dataset, train_batch_size):
-            dropout_rngs = shard_prng_key(dropout_rng)
-            state, metrics = p_train_step(state, batch, dropout_rngs)
+            state, metrics, dropout_rngs = p_train_step(state, batch, dropout_rngs)
             train_metrics.append(metrics)
         train_time += time.time() - train_start
         logger.info(f"    Done! Training metrics: {unreplicate(metrics)}")
 
         logger.info("  Evaluating...")
-        rng, input_rng = jax.random.split(rng)
 
         # evaluate
         for batch in glue_eval_data_collator(eval_dataset, eval_batch_size):
@@ -490,15 +482,12 @@ def main():
 
         # make sure leftover batch is evaluated on one device
         if num_leftover_samples > 0 and jax.process_index() == 0:
-            # put weights on single device
-            state = unreplicate(state)
-
             # take leftover samples
             batch = eval_dataset[-num_leftover_samples:]
             batch = {k: jnp.array(v) for k, v in batch.items()}
 
             labels = batch.pop("labels")
-            predictions = eval_step(state, batch)
+            predictions = eval_step(unreplicate(state), batch)
             metric.add_batch(predictions=predictions, references=labels)
 
         eval_metric = metric.compute()
@@ -507,10 +496,15 @@ def main():
         cur_step = epoch * (len(train_dataset) // train_batch_size)
         write_metric(train_metrics, eval_metric, train_time, cur_step)
 
-    # save last checkpoint
-    if jax.process_index() == 0:
-        params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
-        model.save_pretrained(args.output_dir, params=params)
+        # save checkpoint after each epoch and push checkpoint to the hub
+        if jax.process_index() == 0:
+            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+            model.save_pretrained(
+                args.output_dir,
+                params=params,
+                push_to_hub=args.push_to_hub,
+                commit_message=f"Saving weights and logs of epoch {epoch}",
+            )
 
 
 if __name__ == "__main__":

examples/flax/vision/README.md

@@ -0,0 +1,101 @@
+<!---
+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.
+-->
+
+# Image Classification training examples
+
+The following example showcases how to train/fine-tune `ViT` for image-classification using the JAX/Flax backend.
+
+JAX/Flax allows you to trace pure functions and compile them into efficient, fused accelerator code on both GPU and TPU.
+Models written in JAX/Flax are **immutable** and updated in a purely functional
+way which enables simple and efficient model parallelism.
+
+
+In this example we will train/fine-tune the model on the [imagenette](https://github.com/fastai/imagenette) dataset.
+
+Let's start by creating a model repository to save the trained model and logs.
+Here we call the model `"vit-base-patch16-imagenette"`, but you can change the model name as you like.
+
+You can do this either directly on [huggingface.co](https://huggingface.co/new) (assuming that
+you are logged in) or via the command line:
+
+```
+huggingface-cli repo create vit-base-patch16-imagenette
+```
+Next we clone the model repository to add the tokenizer and model files.
+```
+git clone https://huggingface.co/<your-username>/vit-base-patch16-imagenette
+```
+To ensure that all tensorboard traces will be uploaded correctly, we need to 
+track them. You can run the following command inside your model repo to do so.
+
+```
+cd vit-base-patch16-imagenette
+git lfs track "*tfevents*"
+```
+
+Great, we have set up our model repository. During training, we will automatically
+push the training logs and model weights to the repo.
+
+Next, let's add a symbolic link to the `run_image_classification_flax.py`.
+
+```bash
+export MODEL_DIR="./vit-base-patch16-imagenette
+ln -s ~/transformers/examples/flax/summarization/run_image_classification_flax.py run_image_classification_flax.py
+```
+
+## Prepare the dataset
+
+We will use the [imagenette](https://github.com/fastai/imagenette) dataset to train/fine-tune our model. Imagenette is a subset of 10 easily classified classes from Imagenet (tench, English springer, cassette player, chain saw, church, French horn, garbage truck, gas pump, golf ball, parachute).
+
+
+### Download and extract the data.
+
+```bash
+wget https://s3.amazonaws.com/fast-ai-imageclas/imagenette2.tgz
+tar -xvzf imagenette2.tgz
+```
+
+This will create a `imagenette2` dir with two subdirectories `train` and `val` each with multiple subdirectories per class. The training script expects the following directory structure
+
+```bash
+root/dog/xxx.png
+root/dog/xxy.png
+root/dog/[...]/xxz.png
+
+root/cat/123.png
+root/cat/nsdf3.png
+root/cat/[...]/asd932_.png
+```
+
+## Train the model
+
+Next we can run the example script to fine-tune the model:
+
+```bash
+python run_image_classification.py \
+    --output_dir ${MODEL_DIR} \
+    --model_name_or_path google/vit-base-patch16-224-in21k \
+    --train_dir="imagenette2/train" \
+    --validation_dir="imagenette2/val" \
+    --num_train_epochs 5 \
+    --learning_rate 1e-3 \
+    --per_device_train_batch_size 128 --per_device_eval_batch_size 128 \
+    --overwrite_output_dir \
+    --preprocessing_num_workers 32 \
+    --push_to_hub
+```
+
+This should finish in ~7mins with 99% validation accuracy.

examples/flax/vision/requirements.txt

@@ -0,0 +1,8 @@
+jax>=0.2.8
+jaxlib>=0.1.59
+flax>=0.3.4
+optax>=0.0.8
+-f https://download.pytorch.org/whl/torch_stable.html
+torch==1.9.0+cpu 
+-f https://download.pytorch.org/whl/torch_stable.html
+torchvision==0.10.0+cpu

examples/flax/vision/run_image_classification.py

@@ -0,0 +1,467 @@
+#!/usr/bin/env python
+# coding=utf-8
+# 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.
+"""
+Pre-training/Fine-tuning ViT for image classification .
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=vit
+"""
+
+import logging
+import os
+import sys
+import time
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable, Optional
+
+# for dataset and preprocessing
+import torch
+import torchvision
+import torchvision.transforms as transforms
+from tqdm import tqdm
+
+import jax
+import jax.numpy as jnp
+import optax
+import transformers
+from flax import jax_utils
+from flax.jax_utils import unreplicate
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
+    AutoConfig,
+    FlaxAutoModelForImageClassification,
+    HfArgumentParser,
+    TrainingArguments,
+    is_tensorboard_available,
+    set_seed,
+)
+
+
+logger = logging.getLogger(__name__)
+
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    dtype: Optional[str] = field(
+        default="float32",
+        metadata={
+            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    train_dir: str = field(
+        metadata={"help": "Path to the root training directory which contains one subdirectory per class."}
+    )
+    validation_dir: str = field(
+        metadata={"help": "Path to the root validation directory which contains one subdirectory per class."},
+    )
+    image_size: Optional[int] = field(default=224, metadata={"help": " The size (resolution) of each image."})
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+            "value if set."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+
+
+class TrainState(train_state.TrainState):
+    dropout_rng: jnp.ndarray
+
+    def replicate(self):
+        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
+
+
+def write_metric(summary_writer, train_metrics, eval_metrics, train_time, step):
+    summary_writer.scalar("train_time", train_time, step)
+
+    train_metrics = get_metrics(train_metrics)
+    for key, vals in train_metrics.items():
+        tag = f"train_{key}"
+        for i, val in enumerate(vals):
+            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
+
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+def create_learning_rate_fn(
+    train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
+) -> Callable[[int], jnp.array]:
+    """Returns a linear warmup, linear_decay learning rate function."""
+    steps_per_epoch = train_ds_size // train_batch_size
+    num_train_steps = steps_per_epoch * num_train_epochs
+    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
+    decay_fn = optax.linear_schedule(
+        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
+    )
+    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
+    return schedule_fn
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    # Setup logging, we only want one process per machine to log things on the screen.
+    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
+    if jax.process_index() == 0:
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        transformers.utils.logging.set_verbosity_error()
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # set seed for random transforms and torch dataloaders
+    set_seed(training_args.seed)
+
+    # Initialize datasets and pre-processing transforms
+    # We use torchvision here for faster pre-processing
+    # Note that here we are using some default pre-processing, for maximum accuray
+    # one should tune this part and carefully select what transformations to use.
+    normalize = transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+    train_dataset = torchvision.datasets.ImageFolder(
+        data_args.train_dir,
+        transforms.Compose(
+            [
+                transforms.RandomResizedCrop(data_args.image_size),
+                transforms.RandomHorizontalFlip(),
+                transforms.ToTensor(),
+                normalize,
+            ]
+        ),
+    )
+
+    eval_dataset = torchvision.datasets.ImageFolder(
+        data_args.validation_dir,
+        transforms.Compose(
+            [
+                transforms.Resize(data_args.image_size),
+                transforms.CenterCrop(data_args.image_size),
+                transforms.ToTensor(),
+                normalize,
+            ]
+        ),
+    )
+
+    # Load pretrained model and tokenizer
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(
+            model_args.config_name,
+            num_labels=len(train_dataset.classes),
+            image_size=data_args.image_size,
+            cache_dir=model_args.cache_dir,
+        )
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(
+            model_args.model_name_or_path,
+            num_labels=len(train_dataset.classes),
+            image_size=data_args.image_size,
+            cache_dir=model_args.cache_dir,
+        )
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.model_name_or_path:
+        model = FlaxAutoModelForImageClassification.from_pretrained(
+            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+    else:
+        model = FlaxAutoModelForImageClassification.from_config(
+            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
+        )
+
+    # Store some constant
+    num_epochs = int(training_args.num_train_epochs)
+    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
+    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
+    steps_per_epoch = len(train_dataset) // train_batch_size
+    total_train_steps = steps_per_epoch * num_epochs
+
+    def collate_fn(examples):
+        pixel_values = torch.stack([example[0] for example in examples])
+        labels = torch.tensor([example[1] for example in examples])
+
+        batch = {"pixel_values": pixel_values, "labels": labels}
+        batch = {k: v.numpy() for k, v in batch.items()}
+
+        return batch
+
+    # Create data loaders
+    train_loader = torch.utils.data.DataLoader(
+        train_dataset,
+        batch_size=train_batch_size,
+        shuffle=True,
+        num_workers=data_args.preprocessing_num_workers,
+        persistent_workers=True,
+        drop_last=True,
+        collate_fn=collate_fn,
+    )
+
+    eval_loader = torch.utils.data.DataLoader(
+        eval_dataset,
+        batch_size=eval_batch_size,
+        shuffle=False,
+        num_workers=data_args.preprocessing_num_workers,
+        persistent_workers=True,
+        drop_last=True,
+        collate_fn=collate_fn,
+    )
+
+    # Enable tensorboard only on the master node
+    has_tensorboard = is_tensorboard_available()
+    if has_tensorboard and jax.process_index() == 0:
+        try:
+            from flax.metrics.tensorboard import SummaryWriter
+
+            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
+        except ImportError as ie:
+            has_tensorboard = False
+            logger.warning(
+                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
+            )
+    else:
+        logger.warning(
+            "Unable to display metrics through TensorBoard because the package is not installed: "
+            "Please run pip install tensorboard to enable."
+        )
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    rng, dropout_rng = jax.random.split(rng)
+
+    # Create learning rate schedule
+    linear_decay_lr_schedule_fn = create_learning_rate_fn(
+        len(train_dataset),
+        train_batch_size,
+        training_args.num_train_epochs,
+        training_args.warmup_steps,
+        training_args.learning_rate,
+    )
+
+    # create adam optimizer
+    adamw = optax.adamw(
+        learning_rate=linear_decay_lr_schedule_fn,
+        b1=training_args.adam_beta1,
+        b2=training_args.adam_beta2,
+        eps=training_args.adam_epsilon,
+        weight_decay=training_args.weight_decay,
+    )
+
+    # Setup train state
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw, dropout_rng=dropout_rng)
+
+    def loss_fn(logits, labels):
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1]))
+        return loss.mean()
+
+    # Define gradient update step fn
+    def train_step(state, batch):
+        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+
+        def compute_loss(params):
+            labels = batch.pop("labels")
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+            loss = loss_fn(logits, labels)
+            return loss
+
+        grad_fn = jax.value_and_grad(compute_loss)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+
+        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
+
+        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return new_state, metrics
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+        logits = model(**batch, params=params, train=False)[0]
+        loss = loss_fn(logits, labels)
+
+        # summarize metrics
+        accuracy = (jnp.argmax(logits, axis=-1) == labels).mean()
+        metrics = {"loss": loss, "accuracy": accuracy}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+        return metrics
+
+    # Create parallel version of the train and eval step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+    p_eval_step = jax.pmap(eval_step, "batch")
+
+    # Replicate the train state on each device
+    state = state.replicate()
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {num_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
+    logger.info(f"  Total optimization steps = {total_train_steps}")
+
+    train_time = 0
+    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
+    for epoch in epochs:
+        # ======================== Training ================================
+        train_start = time.time()
+
+        # Create sampling rng
+        rng, input_rng = jax.random.split(rng)
+        train_metrics = []
+
+        steps_per_epoch = len(train_dataset) // train_batch_size
+        train_step_progress_bar = tqdm(total=steps_per_epoch, desc="Training...", position=1, leave=False)
+        # train
+        for batch in train_loader:
+            batch = shard(batch)
+            state, train_metric = p_train_step(state, batch)
+            train_metrics.append(train_metric)
+
+            train_step_progress_bar.update(1)
+
+        train_time += time.time() - train_start
+
+        train_metric = unreplicate(train_metric)
+
+        train_step_progress_bar.close()
+        epochs.write(
+            f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
+        )
+
+        # ======================== Evaluating ==============================
+        eval_metrics = []
+        eval_steps = len(eval_dataset) // eval_batch_size
+        eval_step_progress_bar = tqdm(total=eval_steps, desc="Evaluating...", position=2, leave=False)
+        for batch in eval_loader:
+            # Model forward
+            batch = shard(batch)
+            metrics = p_eval_step(state.params, batch)
+            eval_metrics.append(metrics)
+
+            eval_step_progress_bar.update(1)
+
+        # normalize eval metrics
+        eval_metrics = get_metrics(eval_metrics)
+        eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
+
+        # Print metrics and update progress bar
+        eval_step_progress_bar.close()
+        desc = (
+            f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {round(eval_metrics['loss'].item(), 4)} | "
+            f"Eval Accuracy: {round(eval_metrics['accuracy'].item(), 4)})"
+        )
+        epochs.write(desc)
+        epochs.desc = desc
+
+        # Save metrics
+        if has_tensorboard and jax.process_index() == 0:
+            cur_step = epoch * (len(train_dataset) // train_batch_size)
+            write_metric(summary_writer, train_metrics, eval_metrics, train_time, cur_step)
+
+        # save checkpoint after each epoch and push checkpoint to the hub
+        if jax.process_index() == 0:
+            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+            model.save_pretrained(
+                training_args.output_dir,
+                params=params,
+                push_to_hub=training_args.push_to_hub,
+                commit_message=f"Saving weights and logs of epoch {epoch+1}",
+            )
+
+
+if __name__ == "__main__":
+    main()

examples/legacy/multiple_choice/run_multiple_choice.py

@@ -107,7 +107,7 @@ def main():
 
     # Setup logging
     logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
         datefmt="%m/%d/%Y %H:%M:%S",
         level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
     )

examples/legacy/pytorch-lightning/lightning_base.py

@@ -28,12 +28,12 @@ from transformers.optimization import (
     get_linear_schedule_with_warmup,
     get_polynomial_decay_schedule_with_warmup,
 )
-from transformers.utils.versions import require_version_examples
+from transformers.utils.versions import require_version
 
 
 logger = logging.getLogger(__name__)
 
-require_version_examples("pytorch_lightning>=1.0.4")
+require_version("pytorch_lightning>=1.0.4")
 
 MODEL_MODES = {
     "base": AutoModel,

examples/legacy/question-answering/run_squad.py

@@ -702,7 +702,7 @@ def main():
 
     # Setup logging
     logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
         datefmt="%m/%d/%Y %H:%M:%S",
         level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
     )

examples/legacy/question-answering/run_squad_trainer.py

@@ -89,7 +89,7 @@ def main():
 
     # Setup logging
     logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
         datefmt="%m/%d/%Y %H:%M:%S",
         level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
     )

examples/legacy/run_language_modeling.py

@@ -211,7 +211,7 @@ def main():
 
     # Setup logging
     logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
         datefmt="%m/%d/%Y %H:%M:%S",
         level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
     )

examples/legacy/run_openai_gpt.py

@@ -50,7 +50,7 @@ from transformers import (
 
 
 logging.basicConfig(
-    format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO
+    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO
 )
 logger = logging.getLogger(__name__)
 

examples/legacy/run_swag.py

@@ -617,7 +617,7 @@ def main():
 
     # Setup logging
     logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
         datefmt="%m/%d/%Y %H:%M:%S",
         level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
     )

examples/legacy/run_transfo_xl.py

@@ -33,7 +33,7 @@ from transformers import TransfoXLCorpus, TransfoXLLMHeadModel
 
 
 logging.basicConfig(
-    format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO
+    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO
 )
 logger = logging.getLogger(__name__)
 

examples/legacy/seq2seq/finetune_trainer.py

@@ -163,7 +163,7 @@ def main():
 
     # Setup logging
     logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
         datefmt="%m/%d/%Y %H:%M:%S",
         level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
     )

examples/legacy/text-classification/run_tf_text_classification.py

@@ -220,7 +220,7 @@ def main():
 
     # Setup logging
     logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
         datefmt="%m/%d/%Y %H:%M:%S",
         level=logging.INFO,
     )

examples/legacy/token-classification/run_ner.py

@@ -131,7 +131,7 @@ def main():
 
     # Setup logging
     logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
         datefmt="%m/%d/%Y %H:%M:%S",
         level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
     )