Release: v3.5.0

Model sharing rst (#8439 )
* Update RST * Finer details * Re-organize * Style
2020-11-10 08:50:43 -05:00 · 2020-11-10 08:35:11 -05:00 · 2020-11-10 14:28:30 +01:00 · 2020-11-10 14:03:29 +01:00 · 2020-11-10 07:29:34 -05:00 · 2020-11-10 07:11:02 -05:00
517 changed files with 16357 additions and 6698 deletions
@@ -74,21 +74,20 @@ jobs:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.3-torch_and_tf-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
+                      - v0.4-torch_and_tf-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
            - run: pip install .[sklearn,tf-cpu,torch,testing]
-            - run: pip install codecov pytest-cov
            - save_cache:
-                key: v0.3-{{ checksum "setup.py" }}
+                key: v0.4-{{ checksum "setup.py" }}
                paths:
                    - '~/.cache/pip'
-            - run: RUN_PT_TF_CROSS_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ -m is_pt_tf_cross_test --cov --durations=0 | tee output.txt
-            - run: codecov
+            - run: RUN_PT_TF_CROSS_TESTS=1 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
            - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
    run_tests_torch:
        working_directory: ~/transformers
        docker:
@@ -101,19 +100,20 @@ jobs:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.3-torch-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
+                      - v0.4-torch-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
            - run: pip install .[sklearn,torch,testing]
            - save_cache:
-                  key: v0.3-torch-{{ checksum "setup.py" }}
+                  key: v0.4-torch-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ | tee output.txt
+            - run: python -m pytest -n 8 --dist=loadfile -s --make-reports=tests_torch ./tests/ | tee tests_output.txt
            - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
    run_tests_tf:
        working_directory: ~/transformers
        docker:
@@ -126,19 +126,20 @@ jobs:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.3-tf-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
+                      - v0.4-tf-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
            - run: pip install .[sklearn,tf-cpu,testing]
            - save_cache:
-                  key: v0.3-tf-{{ checksum "setup.py" }}
+                  key: v0.4-tf-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ | tee output.txt
+            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_tf ./tests/ | tee tests_output.txt
            - store_artifacts:
-               path: ~/transformers/output.txt
-               destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
    run_tests_flax:
        working_directory: ~/transformers
        docker:
@@ -151,19 +152,20 @@ jobs:
            - checkout
            - restore_cache:
                keys:
-                    - v0.3-flax-{{ checksum "setup.py" }}
-                    - v0.3-{{ checksum "setup.py" }}
+                    - v0.4-flax-{{ checksum "setup.py" }}
+                    - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
            - run: sudo pip install .[flax,sklearn,torch,testing]
            - save_cache:
-                  key: v0.3-flax-{{ checksum "setup.py" }}
+                  key: v0.4-flax-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ | tee output.txt
+            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_flax ./tests/ | tee tests_output.txt
            - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
    run_tests_pipelines_torch:
        working_directory: ~/transformers
        docker:
@@ -176,19 +178,20 @@ jobs:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.3-torch-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
+                      - v0.4-torch-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
            - run: pip install .[sklearn,torch,testing]
            - save_cache:
-                  key: v0.3-torch-{{ checksum "setup.py" }}
+                  key: v0.4-torch-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
-            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ -m is_pipeline_test | tee output.txt
+            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_pipelines_torch -m is_pipeline_test ./tests/ | tee tests_output.txt
            - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
    run_tests_pipelines_tf:
        working_directory: ~/transformers
        docker:
@@ -201,19 +204,20 @@ jobs:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.3-tf-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
+                      - v0.4-tf-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
            - run: pip install .[sklearn,tf-cpu,testing]
            - save_cache:
-                  key: v0.3-tf-{{ checksum "setup.py" }}
+                  key: v0.4-tf-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
-            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ -m is_pipeline_test | tee output.txt
+            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_pipelines_tf ./tests/ -m is_pipeline_test | tee tests_output.txt
            - store_artifacts:
-               path: ~/transformers/output.txt
-               destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
    run_tests_custom_tokenizers:
        working_directory: ~/transformers
        docker:
@@ -224,19 +228,21 @@ jobs:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.3-custom_tokenizers-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
+                      - v0.4-custom_tokenizers-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[ja,testing]
            - run: python -m unidic download
            - save_cache:
-                  key: v0.3-custom_tokenizers-{{ checksum "setup.py" }}
+                  key: v0.4-custom_tokenizers-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
-            - run: python -m pytest -s ./tests/test_tokenization_bert_japanese.py | tee output.txt
+            - run: python -m pytest -s --make-reports=tests_custom_tokenizers ./tests/test_tokenization_bert_japanese.py | tee tests_output.txt
            - store_artifacts:
-                path: ~/transformers/output.txt
-                destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
    run_examples_torch:
        working_directory: ~/transformers
        docker:
@@ -249,19 +255,21 @@ jobs:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.3-torch_examples-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
+                      - v0.4-torch_examples-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[sklearn,torch,testing]
            - run: pip install -r examples/requirements.txt
            - save_cache:
-                  key: v0.3-torch_examples-{{ checksum "setup.py" }}
+                  key: v0.4-torch_examples-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s ./examples/ | tee output.txt
+            - run: python -m pytest -n 8 --dist=loadfile -s --make-reports=examples_torch ./examples/ | tee examples_output.txt
            - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+                  path: ~/transformers/examples_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
    build_doc:
        working_directory: ~/transformers
        docker:
@@ -270,17 +278,18 @@ jobs:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.3-build_doc-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
+                      - v0.4-build_doc-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install .[tf,torch,sentencepiece,docs]
            - save_cache:
-                  key: v0.3-build_doc-{{ checksum "setup.py" }}
+                  key: v0.4-build_doc-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: cd docs && make html SPHINXOPTS="-W"
            - store_artifacts:
                path: ./docs/_build
+
    deploy_doc:
        working_directory: ~/transformers
        docker:
@@ -292,14 +301,15 @@ jobs:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.3-deploy_doc-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
+                      - v0.4-deploy_doc-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install .[tf,torch,sentencepiece,docs]
            - save_cache:
-                  key: v0.3-deploy_doc-{{ checksum "setup.py" }}
+                  key: v0.4-deploy_doc-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
            - run: ./.circleci/deploy.sh
+
    check_code_quality:
        working_directory: ~/transformers
        docker:
@@ -310,22 +320,23 @@ jobs:
            - checkout
            - restore_cache:
                  keys:
-                      - v0.3-code_quality-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
+                      - v0.4-code_quality-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
            - run: pip install isort
            - run: pip install .[tf,torch,flax,quality]
            - save_cache:
-                  key: v0.3-code_quality-{{ checksum "setup.py" }}
+                  key: v0.4-code_quality-{{ checksum "setup.py" }}
                  paths:
                      - '~/.cache/pip'
-            - run: black --check examples templates tests src utils
-            - run: isort --check-only examples templates tests src utils
-            - run: flake8 examples templates tests src utils
+            - run: black --check examples tests src utils
+            - run: isort --check-only examples tests src utils
+            - run: flake8 examples tests src utils
            - run: python utils/style_doc.py src/transformers docs/source --max_len 119 --check_only
            - run: python utils/check_copies.py
            - run: python utils/check_dummies.py
            - run: python utils/check_repo.py
+
    check_repository_consistency:
        working_directory: ~/transformers
        docker:
@@ -356,6 +367,7 @@ jobs:
            - setup_remote_docker
            - *build_push_docker
            - *deploy_cluster
+
    cleanup-gke-jobs:
        docker:
            - image: circleci/python:3.6
@@ -365,6 +377,7 @@ jobs:
                  cluster: $GKE_CLUSTER
                  perform-login: true
            - *delete_gke_jobs
+
 workflow_filters: &workflow_filters
    filters:
        branches:
@@ -30,20 +30,22 @@ assignees: ''
 Trainer: @sgugger
 Speed and Memory Benchmarks: @patrickvonplaten
 Model Cards: @julien-c
- Translation: @sshleifer
- Summarization: @sshleifer
 TextGeneration: @TevenLeScao 
 examples/distillation: @VictorSanh
 nlp datasets: [different repo](https://github.com/huggingface/nlp)
 rust tokenizers: [different repo](https://github.com/huggingface/tokenizers)
- Text Generation: @TevenLeScao
- blenderbot: @mariamabarham
- Bart: @sshleifer
- Marian: @sshleifer
+ Text Generation: @patrickvonplaten @TevenLeScao
+ Blenderbot: @patrickvonplaten
+ Bart: @patrickvonplaten
+ Marian: @patrickvonplaten
+ Pegasus: @patrickvonplaten
+ mBART: @patrickvonplaten
 T5: @patrickvonplaten
 Longformer/Reformer: @patrickvonplaten
- TransfoXL/XLNet: @TevenLeScao 
- examples/seq2seq: @sshleifer
+ TransfoXL/XLNet: @TevenLeScao
+ RAG: @patrickvonplaten, @lhoestq
+ FSMT: @stas00
+ examples/seq2seq: @patil-suraj
 examples/bert-loses-patience: @JetRunner
 tensorflow: @jplu
 examples/token-classification: @stefan-it
@@ -43,21 +43,20 @@ members/contributors which may be interested in your PR.
 Trainer: @sgugger
 Benchmarks: @patrickvonplaten
 Model Cards: @julien-c
- Translation: @sshleifer
- Summarization: @sshleifer
 examples/distillation: @VictorSanh
 nlp datasets: [different repo](https://github.com/huggingface/nlp)
 rust tokenizers: [different repo](https://github.com/huggingface/tokenizers)
 Text Generation: @patrickvonplaten, @TevenLeScao
- Blenderbot, Bart, Marian, Pegasus: @sshleifer
+ Blenderbot, Bart, Marian, Pegasus: @patrickvonplaten
 T5: @patrickvonplaten
 Rag: @patrickvonplaten, @lhoestq
 EncoderDecoder: @patrickvonplaten
 Longformer, Reformer: @patrickvonplaten
 TransfoXL, XLNet: @TevenLeScao, @patrickvonplaten
- examples/seq2seq: @sshleifer
+ examples/seq2seq: @patil-suraj
 examples/bert-loses-patience: @JetRunner
 tensorflow: @jplu
 examples/token-classification: @stefan-it
 documentation: @sgugger
+ FSTM: @stas00
 -->
@@ -1,10 +1,10 @@
 name: Self-hosted runner (push)

-on: 
+on:
  push:
    branches:
      - master
-    paths: 
+    paths:
      - "src/**"
      - "tests/**"
      - ".github/**"
@@ -13,7 +13,69 @@ on:


 jobs:
-  run_tests_torch_and_tf_gpu:
+  run_tests_torch_gpu:
+    runs-on: [self-hosted, single-gpu]
+    steps:
+      - uses: actions/checkout@v2
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v1.1-tests_torch_gpu-${{ hashFiles('setup.py') }}
+
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install .[torch,sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
+
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
+          python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
+
+      - name: Run all non-slow tests on GPU
+        env:
+          OMP_NUM_THREADS: 1
+          CUDA_VISIBLE_DEVICES: 0
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 2 --dist=loadfile -s --make-reports=tests_torch_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_torch_gpu_failures_short.txt
+        
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_torch_gpu_test_reports
+          path: reports
+                  
+
+  run_tests_tf_gpu:
    runs-on: [self-hosted, single-gpu]
    steps:
      - uses: actions/checkout@v2
@@ -31,7 +93,7 @@ jobs:
        id: cache
        with:
          path: .env
-          key: v0-tests_tf_torch_gpu-${{ hashFiles('setup.py') }}
+          key: v1.1-tests_tf_gpu-${{ hashFiles('setup.py') }}

      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
        run: |
@@ -40,30 +102,40 @@ jobs:
          which python
          python --version
          pip --version
+
      - name: Install dependencies
        run: |
          source .env/bin/activate
          pip install --upgrade pip
-          pip install torch!=1.6.0
-          pip install .[sklearn,testing,onnxruntime]
+          pip install .[tf,sklearn,testing,onnxruntime]
          pip install git+https://github.com/huggingface/datasets

      - name: Are GPUs recognized by our DL frameworks
        run: |
          source .env/bin/activate
-          python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
-          python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
+          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_FORCE_GPU_ALLOW_GROWTH: "true"
-          # TF_GPU_MEMORY_LIMIT: 4096
          OMP_NUM_THREADS: 1
+          CUDA_VISIBLE_DEVICES: 0
        run: |
          source .env/bin/activate
-          python -m pytest -n 2 --dist=loadfile -s ./tests/
+          python -m pytest -n 2 --dist=loadfile -s --make-reports=tests_tf_gpu tests

-  run_tests_torch_and_tf_multiple_gpu:
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_tf_gpu_failures_short.txt
+        
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_tf_gpu_test_reports
+          path: reports
+
+  run_tests_torch_multiple_gpu:
    runs-on: [self-hosted, multi-gpu]
    steps:
      - uses: actions/checkout@v2
@@ -72,6 +144,7 @@ jobs:
          which python
          python --version
          pip --version
+
      - name: Current dir
        run: pwd
      - run: nvidia-smi
@@ -81,7 +154,7 @@ jobs:
        id: cache
        with:
          path: .env
-          key: v0-tests_tf_torch_multiple_gpu-${{ hashFiles('setup.py') }}
+          key: v1.1-tests_torch_multiple_gpu-${{ hashFiles('setup.py') }}

      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
        run: |
@@ -94,8 +167,7 @@ jobs:
        run: |
          source .env/bin/activate
          pip install --upgrade pip
-          pip install torch!=1.6.0
-          pip install .[sklearn,testing,onnxruntime]
+          pip install .[torch,sklearn,testing,onnxruntime]
          pip install git+https://github.com/huggingface/datasets

      - name: Are GPUs recognized by our DL frameworks
@@ -106,9 +178,78 @@ jobs:

      - name: Run all non-slow tests on GPU
        env:
-          TF_FORCE_GPU_ALLOW_GROWTH: "true"
-          # TF_GPU_MEMORY_LIMIT: 4096
          OMP_NUM_THREADS: 1
        run: |
          source .env/bin/activate
-          python -m pytest -n 2 --dist=loadfile -s ./tests/
+          python -m pytest -n 2 --dist=loadfile -s --make-reports=tests_torch_multiple_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_torch_multiple_gpu_failures_short.txt          
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_torch_multi_gpu_test_reports
+          path: reports
+
+  run_tests_tf_multiple_gpu:
+    runs-on: [self-hosted, multi-gpu]
+    steps:
+      - uses: actions/checkout@v2
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v1.1-tests_tf_multiple_gpu-${{ hashFiles('setup.py') }}
+
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install .[tf,sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
+
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          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:
+          OMP_NUM_THREADS: 1
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 2 --dist=loadfile -s --make-reports=tests_tf_multiple_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_tf_multiple_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
+          
@@ -1,3 +1,8 @@
+# configuration notes:
+#
+# - `source .env/bin/activate` is currently needed to be run first thing first in each step. Otherwise
+# the step uses the system-wide python interpreter.
+
 name: Self-hosted runner (scheduled)

 on:
@@ -9,7 +14,7 @@ on:
    - cron: "0 0 * * *"

 jobs:
-  run_all_tests_torch_and_tf_gpu:
+  run_all_tests_torch_gpu:
    runs-on: [self-hosted, single-gpu]
    steps:
      - uses: actions/checkout@v2
@@ -19,16 +24,18 @@ jobs:
        id: cache
        with:
          path: .env
-          key: v0-slow_tests_tf_torch_gpu-${{ hashFiles('setup.py') }}
+          key: v  1.1-slow_tests_torch_gpu-${{ hashFiles('setup.py') }}

      - name: Python version
        run: |
          which python
          python --version
          pip --version
+
      - name: Current dir
        run: pwd
      - run: nvidia-smi
+
      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
        if: steps.cache.outputs.cache-hit != 'true'
        run: |
@@ -37,13 +44,14 @@ jobs:
          which python
          python --version
          pip --version
+
      - name: Install dependencies
        run: |
          source .env/bin/activate
          pip install --upgrade pip
-          pip install torch!=1.6.0
-          pip install .[sklearn,testing,onnxruntime]
+          pip install .[torch,sklearn,testing,onnxruntime]
          pip install git+https://github.com/huggingface/datasets
+          pip list

      - name: Are GPUs recognized by our DL frameworks
        run: |
@@ -51,27 +59,34 @@ jobs:
          python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
          python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"

-
      - name: Run all tests on GPU
        env:
-          TF_FORCE_GPU_ALLOW_GROWTH: "true"
          OMP_NUM_THREADS: 1
          RUN_SLOW: yes
        run: |
          source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s ./tests/ --durations=50
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=tests_torch_gpu tests

+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_torch_gpu_failures_short.txt
+        
      - name: Run examples tests on GPU
+        if: ${{ always() }}
        env:
-          TF_FORCE_GPU_ALLOW_GROWTH: "true"
          OMP_NUM_THREADS: 1
          RUN_SLOW: yes
        run: |
          source .env/bin/activate
          pip install -r examples/requirements.txt
-          python -m pytest -n 1 --dist=loadfile -s examples --durations=50
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=examples_torch_gpu examples
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/examples_torch_gpu_failures_short.txt

      - name: Run all pipeline tests on GPU
+        if: ${{ always() }}
        env:
          TF_FORCE_GPU_ALLOW_GROWTH: "true"
          OMP_NUM_THREADS: 1
@@ -79,10 +94,100 @@ jobs:
          RUN_PIPELINE_TESTS: yes
        run: |
          source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s ./tests/ -m is_pipeline_test --durations=50
+          python -m pytest -n 1 --dist=loadfile -s -m is_pipeline_test --make-reports=tests_torch_pipeline_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_torch_pipeline_gpu_failures_short.txt
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_torch_gpu_test_reports
+          path: reports


-  run_all_tests_torch_and_tf_multiple_gpu:
+  run_all_tests_tf_gpu:
+    runs-on: [self-hosted, single-gpu]
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v1.1-slow_tests_tf_gpu-${{ hashFiles('setup.py') }}
+
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
+
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        if: steps.cache.outputs.cache-hit != 'true'
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install .[tf,sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
+          pip list
+
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          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 tests on GPU
+        env:
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=tests_tf_gpu tests
+          
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_tf_gpu_failures_short.txt
+
+      - name: Run all pipeline tests on GPU
+        if: ${{ always() }}
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+          RUN_PIPELINE_TESTS: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --dist=loadfile -s -m is_pipeline_test --make-reports=tests_tf_pipelines_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_tf_pipelines_gpu_failures_short.txt
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_tf_gpu_test_reports
+          path: reports
+          
+  run_all_tests_torch_multiple_gpu:
    runs-on: [self-hosted, multi-gpu]
    steps:
      - uses: actions/checkout@v2
@@ -92,16 +197,18 @@ jobs:
        id: cache
        with:
          path: .env
-          key: v0-slow_tests_tf_torch_multi_gpu-${{ hashFiles('setup.py') }}
+          key: v1.1-slow_tests_torch_multi_gpu-${{ hashFiles('setup.py') }}

      - name: Python version
        run: |
          which python
          python --version
          pip --version
+
      - name: Current dir
        run: pwd
      - run: nvidia-smi
+
      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
        if: steps.cache.outputs.cache-hit != 'true'
        run: |
@@ -110,13 +217,14 @@ jobs:
          which python
          python --version
          pip --version
+
      - name: Install dependencies
        run: |
          source .env/bin/activate
          pip install --upgrade pip
-          pip install torch!=1.6.0
-          pip install .[sklearn,testing,onnxruntime]
+          pip install .[torch,sklearn,testing,onnxruntime]
          pip install git+https://github.com/huggingface/datasets
+          pip list

      - name: Are GPUs recognized by our DL frameworks
        run: |
@@ -124,26 +232,32 @@ jobs:
          python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
          python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"

-      - name: Run all tests on GPU
+      - name: Run all tests on multi-GPU
        env:
-          TF_FORCE_GPU_ALLOW_GROWTH: "true"
          OMP_NUM_THREADS: 1
          RUN_SLOW: yes
        run: |
          source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s ./tests/ --durations=50
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=tests_torch_multiple_gpu tests

-      - name: Run examples tests on GPU
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_torch_multiple_gpu_failures_short.txt
+
+      - name: Run examples tests on multi-GPU
        env:
-          TF_FORCE_GPU_ALLOW_GROWTH: "true"
          OMP_NUM_THREADS: 1
          RUN_SLOW: yes
        run: |
          source .env/bin/activate
-          pip install -r examples/requirements.txt
-          python -m pytest -n 1 --dist=loadfile -s examples --durations=50
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=examples_torch_multiple_gpu examples

-      - name: Run all pipeline tests on GPU
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/examples_torch_multiple_gpu_failures_short.txt
+
+      - name: Run all pipeline tests on multi-GPU
+        if: ${{ always() }}
        env:
          TF_FORCE_GPU_ALLOW_GROWTH: "true"
          OMP_NUM_THREADS: 1
@@ -151,4 +265,95 @@ jobs:
          RUN_PIPELINE_TESTS: yes
        run: |
          source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s ./tests/ -m is_pipeline_test --durations=50
+          python -m pytest -n 1 --dist=loadfile -s -m is_pipeline_test --make-reports=tests_torch_pipeline_multiple_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_torch_pipeline_multiple_gpu_failures_short.txt
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_torch_multi_gpu_test_reports
+          path: reports
+
+  run_all_tests_tf_multiple_gpu:
+    runs-on: [self-hosted, multi-gpu]
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v1.1-slow_tests_tf_multi_gpu-${{ hashFiles('setup.py') }}
+
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
+
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        if: steps.cache.outputs.cache-hit != 'true'
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install .[tf,sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
+          pip list
+
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          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 tests on multi-GPU
+        env:
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=tests_tf_multiple_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_tf_multiple_gpu_failures_short.txt
+
+      - name: Run all pipeline tests on multi-GPU
+        if: ${{ always() }}
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+          RUN_PIPELINE_TESTS: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --dist=loadfile -s -m is_pipeline_test --make-reports=tests_tf_pipelines_multiple_gpu tests
+          
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_tf_multiple_gpu_pipelines_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
+          
@@ -96,7 +96,7 @@ folder.

 ## Start contributing! (Pull Requests)

-Before writing code, we strongly advise you to search through the exising PRs or
+Before writing code, we strongly advise you to search through the existing PRs or
 issues to make sure that nobody is already working on the same thing. If you are
 unsure, it is always a good idea to open an issue to get some feedback.

@@ -235,7 +235,7 @@ Follow these steps to start contributing:
 ### Checklist

 1. The title of your pull request should be a summary of its contribution;
-2. If your pull request adresses an issue, please mention the issue number in
+2. If your pull request addresses an issue, please mention the issue number in
   the pull request description to make sure they are linked (and people
   consulting the issue know you are working on it);
 3. To indicate a work in progress please prefix the title with `[WIP]`. These
@@ -1,21 +1,15 @@
 .PHONY: modified_only_fixup extra_quality_checks quality style fixup fix-copies test test-examples docs


-check_dirs := examples templates tests src utils
-
-# get modified files since the branch was made
-fork_point_sha := $(shell git merge-base --fork-point master)
-joined_dirs := $(shell echo $(check_dirs) | tr " " "|")
-modified_py_files := $(shell git diff --name-only $(fork_point_sha) | egrep '^($(joined_dirs))' | egrep '\.py$$')
-#$(info modified files are: $(modified_py_files))
+check_dirs := examples tests src utils

 modified_only_fixup:
-	@if [ -n "$(modified_py_files)" ]; then \
+	$(eval modified_py_files := $(shell python utils/get_modified_files.py $(check_dirs)))
+	@if test -n "$(modified_py_files)"; then \
 		echo "Checking/fixing $(modified_py_files)"; \
 		black $(modified_py_files); \
 		isort $(modified_py_files); \
 		flake8 $(modified_py_files); \
-		python utils/style_doc.py $(modified_py_files) --max_len 119; \
 	else \
 		echo "No library .py files were modified"; \
 	fi
@@ -26,6 +20,7 @@ extra_quality_checks:
 	python utils/check_copies.py
 	python utils/check_dummies.py
 	python utils/check_repo.py
+	python utils/style_doc.py src/transformers docs/source --max_len 119

 # this target runs checks on all files
 quality:
@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'3.4.0'
+release = u'3.5.0'


 # -- General configuration ---------------------------------------------------
@@ -272,3 +272,4 @@ conversion utilities for the following models:
    internal/pipelines_utils
    internal/tokenization_utils
    internal/trainer_utils
+    internal/generation_utils
@@ -80,9 +80,9 @@ cache home followed by ``/transformers/`` (even if you don't have PyTorch instal
 So if you don't have any specific environment variable set, the cache directory will be at
 ``~/.cache/torch/transformers/``.

-**Note:** If you have set a shell enviromnent variable for one of the predecessors of this library
+**Note:** If you have set a shell environment variable for one of the predecessors of this library
 (``PYTORCH_TRANSFORMERS_CACHE`` or ``PYTORCH_PRETRAINED_BERT_CACHE``), those will be used if there is no shell
-enviromnent variable for ``TRANSFORMERS_CACHE``.
+environment variable for ``TRANSFORMERS_CACHE``.

 ### Note on model downloads (Continuous Integration or large-scale deployments)

@@ -0,0 +1,50 @@
+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`, and :meth:`~transformers.PretrainedModel.beam_sample`.
+
+Most of those are only useful if you are studying the code of the generate methods in the library.
+
+LogitsProcessor
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+A :class:`~transformers.LogitsProcessor` can be used to modify the prediction scores of a language model head for
+generation.
+
+.. autoclass:: transformers.LogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.LogitsProcessorList
+    :members: __call__
+
+.. autoclass:: transformers.MinLengthLogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.TemperatureLogitsWarper
+    :members: __call__
+
+.. autoclass:: transformers.RepetitionPenaltyLogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.TopPLogitsWarper
+    :members: __call__
+
+.. autoclass:: transformers.TopKLogitsWarper
+    :members: __call__
+
+.. autoclass:: transformers.NoRepeatNGramLogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.NoBadWordsLogitsProcessor
+    :members: __call__
+
+BeamSearch
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BeamScorer
+    :members: process, finalize
+
+.. autoclass:: transformers.BeamSearchScorer
+    :members: process, finalize
@@ -13,7 +13,7 @@ subclass :class:`~transformers.Trainer` and override the methods you need (see :
 By default a :class:`~transformers.Trainer` will use the following callbacks:

 - :class:`~transformers.DefaultFlowCallback` which handles the default behavior for logging, saving and evaluation.
- :class:`~transformers.PrinterCallback` or :class:`~transformers.ProrgressCallback` to display progress and print the
+- :class:`~transformers.PrinterCallback` or :class:`~transformers.ProgressCallback` to display progress and print the
  logs (the first one is used if you deactivate tqdm through the :class:`~transformers.TrainingArguments`, otherwise
  it's the second one).
 - :class:`~transformers.integrations.TensorBoardCallback` if tensorboard is accessible (either through PyTorch >= 1.4
@@ -21,6 +21,8 @@ By default a :class:`~transformers.Trainer` will use the following callbacks:
 - :class:`~transformers.integrations.WandbCallback` if `wandb <https://www.wandb.com/>`__ is installed.
 - :class:`~transformers.integrations.CometCallback` if `comet_ml <https://www.comet.ml/site/>`__ is installed.
 - :class:`~transformers.integrations.MLflowCallback` if `mlflow <https://www.mlflow.org/>`__ is installed.
+- :class:`~transformers.integrations.AzureMLCallback` if `azureml-sdk <https://pypi.org/project/azureml-sdk/>`__ is
+  installed.

 The main class that implements callbacks is :class:`~transformers.TrainerCallback`. It gets the
 :class:`~transformers.TrainingArguments` used to instantiate the :class:`~transformers.Trainer`, can access that
@@ -50,6 +52,7 @@ Here is the list of the available :class:`~transformers.TrainerCallback` in the
 .. autoclass:: transformers.integrations.MLflowCallback
    :members: setup

+.. autoclass:: transformers.integrations.AzureMLCallback

 TrainerCallback
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -45,7 +45,7 @@ TFModelUtilsMixin
    :members:


-Generative models
+Generation
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.generation_utils.GenerationMixin
@@ -65,12 +65,34 @@ BaseModelOutputWithPooling
    :members:


+BaseModelOutputWithCrossAttentions
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_outputs.BaseModelOutputWithCrossAttentions
+    :members:
+
+
+BaseModelOutputWithPoolingAndCrossAttentions
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_outputs.BaseModelOutputWithPoolingAndCrossAttentions
+    :members:
+
+
 BaseModelOutputWithPast
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.modeling_outputs.BaseModelOutputWithPast
    :members:

+
+BaseModelOutputWithPastAndCrossAttentions
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_outputs.BaseModelOutputWithPastAndCrossAttentions
+    :members:
+
+
 Seq2SeqModelOutput
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -85,6 +107,20 @@ CausalLMOutput
    :members:


+CausalLMOutputWithCrossAttentions
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_outputs.CausalLMOutputWithCrossAttentions
+    :members:
+
+
+CausalLMOutputWithPastAndCrossAttentions
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_outputs.CausalLMOutputWithPastAndCrossAttentions
+    :members:
+
+
 CausalLMOutputWithPast
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -20,7 +20,7 @@ Here is a quick summary of what you should take care of when migrating from `pyt

 The main breaking change when migrating from `pytorch-pretrained-bert` to 🤗 Transformers is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.

-The exact content of the tuples for each model are detailled in the models' docstrings and the [documentation](https://huggingface.co/transformers/).
+The exact content of the tuples for each model are detailed in the models' docstrings and the [documentation](https://huggingface.co/transformers/).

 In pretty much every case, you will be fine by taking the first element of the output as the output you previously used in `pytorch-pretrained-bert`.

@@ -109,7 +109,7 @@ for batch in train_data:
    loss.backward()
    optimizer.step()

-### In 🤗 Transformers, optimizer and schedules are splitted and instantiated like this:
+### In 🤗 Transformers, optimizer and schedules are split and instantiated like this:
 optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False)  # To reproduce BertAdam specific behavior set correct_bias=False
 scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)  # PyTorch scheduler
 ### and used like this:
@@ -1,4 +1,4 @@
-AutoClasses
+Auto Classes
 -----------------------------------------------------------------------------------------------------------------------

 In many cases, the architecture you want to use can be guessed from the name or the path of the pretrained model you
@@ -46,10 +46,24 @@ AutoModelForPreTraining
    :members:


-AutoModelWithLMHead
+AutoModelForCausalLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.AutoModelWithLMHead
+.. autoclass:: transformers.AutoModelForCausalLM
+    :members:
+
+
+AutoModelForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AutoModelForMaskedLM
+    :members:
+
+
+AutoModelForSeq2SeqLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AutoModelForSeq2SeqLM
    :members:


@@ -67,6 +81,13 @@ AutoModelForMultipleChoice
    :members:


+AutoModelForNextSentencePrediction
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AutoModelForNextSentencePrediction
+    :members:
+
+
 AutoModelForTokenClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -95,10 +116,24 @@ TFAutoModelForPreTraining
    :members:


-TFAutoModelWithLMHead
+TFAutoModelForCausalLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-.. autoclass:: transformers.TFAutoModelWithLMHead
+.. autoclass:: transformers.TFAutoModelForCausalLM
+    :members:
+
+
+TFAutoModelForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFAutoModelForMaskedLM
+    :members:
+
+
+TFAutoModelForSeq2SeqLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFAutoModelForSeq2SeqLM
    :members:


@@ -3,7 +3,7 @@ BART

 **DISCLAIMER:** If you see something strange, file a `Github Issue
 <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
-@sshleifer
+@patrickvonplaten

 Overview
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -26,6 +26,16 @@ According to the abstract,
 The Authors' code can be found `here <https://github.com/pytorch/fairseq/tree/master/examples/bart>`__.


+Examples
+_______________________________________________________________________________________________________________________
+
+- Examples and scripts for fine-tuning BART and other models for sequence to sequence tasks can be found in
+  `examples/seq2seq/ <https://github.com/huggingface/transformers/blob/master/examples/seq2seq/README.md>`__.
+- An example of how to train :class:`~transformers.BartForConditionalGeneration` with a Hugging Face :obj:`datasets`
+  object can be found in this `forum discussion
+  <https://discuss.huggingface.co/t/train-bart-for-conditional-generation-e-g-summarization/1904>`__.
+
+
 Implementation Notes
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -95,3 +95,12 @@ See :obj:`transformers.BartForConditionalGeneration` for arguments to `forward`

 .. autoclass:: transformers.BlenderbotForConditionalGeneration
    :members:
+
+
+TFBlenderbotForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+See :obj:`transformers.TFBartForConditionalGeneration` for arguments to `forward` and `generate`
+
+.. autoclass:: transformers.TFBlenderbotForConditionalGeneration
+    :members:
@@ -90,6 +90,32 @@ LongformerTokenizerFast
 .. autoclass:: transformers.LongformerTokenizerFast
    :members: 

+Longformer specific outputs
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_longformer.LongformerBaseModelOutput
+    :members: 
+
+.. autoclass:: transformers.modeling_longformer.LongformerBaseModelOutputWithPooling
+    :members: 
+
+.. autoclass:: transformers.modeling_longformer.LongformerMultipleChoiceModelOutput
+    :members: 
+
+.. autoclass:: transformers.modeling_longformer.LongformerQuestionAnsweringModelOutput
+    :members: 
+
+.. autoclass:: transformers.modeling_tf_longformer.TFLongformerBaseModelOutput
+    :members: 
+
+.. autoclass:: transformers.modeling_tf_longformer.TFLongformerBaseModelOutputWithPooling
+    :members: 
+
+.. autoclass:: transformers.modeling_tf_longformer.TFLongformerQuestionAnsweringModelOutput
+    :members: 
+
+LongformerModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 LongformerModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -3,7 +3,7 @@ MarianMT

 **Bugs:** If you see something strange, file a `Github Issue
 <https://github.com/huggingface/transformers/issues/new?assignees=sshleifer&labels=&template=bug-report.md&title>`__
-and assign @sshleifer.
+and assign @patrickvonplaten.

 Translations should be similar, but not identical to, output in the test set linked to in each model card.

@@ -129,3 +129,9 @@ MarianMTModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.MarianMTModel
+
+
+TFMarianMTModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFMarianMTModel
@@ -3,7 +3,7 @@ MBart

 **DISCLAIMER:** If you see something strange, file a `Github Issue
 <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
-@sshleifer
+@patrickvonplaten

 Overview
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -79,4 +79,11 @@ MBartForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.MBartForConditionalGeneration
-    :members: forward
+    :members:
+
+
+TFMBartForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFMBartForConditionalGeneration
+    :members:
@@ -3,7 +3,7 @@ Pegasus

 **DISCLAIMER:** If you see something strange, file a `Github Issue
 <https://github.com/huggingface/transformers/issues/new?assignees=sshleifer&labels=&template=bug-report.md&title>`__
-and assign @sshleifer.
+and assign @patrickvonplaten.


 Overview
@@ -95,3 +95,9 @@ PegasusForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.PegasusForConditionalGeneration
+
+
+TFPegasusForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFPegasusForConditionalGeneration
@@ -18,39 +18,65 @@ done something similar on your task, either using the model directly in your own
 :class:`~.transformers.Trainer`/:class:`~.transformers.TFTrainer` class. Let's see how you can share the result on the
 `model hub <https://huggingface.co/models>`__.

+Model versioning
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Since version v3.5.0, the model hub has built-in model versioning based on git and git-lfs. It is based on the paradigm
+that one model *is* one repo.
+
+This allows:
+
+- built-in versioning
+- access control
+- scalability
+
+This is built around *revisions*, which is a way to pin a specific version of a model, using a commit hash, tag or
+branch.
+
+For instance:
+
+.. code-block::
+
+    >>> tokenizer = AutoTokenizer.from_pretrained(
+    >>>   "julien-c/EsperBERTo-small",
+    >>>   revision="v2.0.1" # tag name, or branch name, or commit hash
+    >>> )
+
 Basic steps
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

-.. 
-    When #5258 is merged, we can remove the need to create the directory.
+In order to upload a model, you'll need to first create a git repo. This repo will live on the model hub, allowing
+users to clone it and you (and your organization members) to push to it. First, you should ensure you are logged in the
+``transformers-cli``:

-First, pick a directory with the name you want your model to have on the model hub (its full name will then be
-`username/awesome-name-you-picked` or `organization/awesome-name-you-picked`) and create it with either
+Go in a terminal and run the following command. It should be in the virtual environment where you installed 🤗
+Transformers, since that command :obj:`transformers-cli` comes from the library.

 .. code-block::

-    mkdir path/to/awesome-name-you-picked
+    transformers-cli login

-or in python
+
+Once you are logged in with your model hub credentials, you can start building your repositories. To create a repo:

 .. code-block::

-    import os
-    os.makedirs("path/to/awesome-name-you-picked")
+    transformers-cli repo create your-model-name

-then you can save your model and tokenizer with:
+This creates a repo on the model hub, which can be cloned. You can then add/remove from that repo as you would with any
+other git repo.

 .. code-block::

-    model.save_pretrained("path/to/awesome-name-you-picked")
-    tokenizer.save_pretrained("path/to/awesome-name-you-picked")
+    git clone https://huggingface.co/username/your-model-name

-Or, if you're using the Trainer API
+    # Then commit as usual
+    cd your-model-name
+    echo "hello" >> README.md
+    git add . && git commit -m "Update from $USER"

-.. code-block::
+We are intentionally not wrapping git too much, so as to stay intuitive and easy-to-use.

-    trainer.save_model("path/to/awesome-name-you-picked")
-    tokenizer.save_pretrained("path/to/awesome-name-you-picked")

 Make your model work on all frameworks
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -71,13 +97,13 @@ or removing TF. For instance, if you trained a :class:`~transformers.DistilBertF

 .. code-block::

-    from transformers import TFDistilBertForSequenceClassification
+    >>> from transformers import TFDistilBertForSequenceClassification

 and if you trained a :class:`~transformers.TFDistilBertForSequenceClassification`, try to type

 .. code-block::

-    from transformers import DistilBertForSequenceClassification
+    >>> from transformers import DistilBertForSequenceClassification

 This will give back an error if your model does not exist in the other framework (something that should be pretty rare
 since we're aiming for full parity between the two frameworks). In this case, skip this and go to the next step.
@@ -87,20 +113,20 @@ model class:

 .. code-block::

-    tf_model = TFDistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_pt=True)
-    tf_model.save_pretrained("path/to/awesome-name-you-picked")
+    >>> tf_model = TFDistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_pt=True)
+    >>> tf_model.save_pretrained("path/to/awesome-name-you-picked")

 and if you trained your model in TensorFlow and have to create a PyTorch version, adapt the following code to your
 model class:

 .. code-block::

-    pt_model = DistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_tf=True)
-    pt_model.save_pretrained("path/to/awesome-name-you-picked")
+    >>> pt_model = DistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_tf=True)
+    >>> pt_model.save_pretrained("path/to/awesome-name-you-picked")

 That's all there is to it!

-Check the directory before uploading
+Check the directory before pushing to the model hub.
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

 Make sure there are no garbage files in the directory you'll upload. It should only have:
@@ -116,62 +142,46 @@ Make sure there are no garbage files in the directory you'll upload. It should o

 Other files can safely be deleted.

-Upload your model with the CLI
+
+Uploading your files
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-Now go in a terminal and run the following command. It should be in the virtual enviromnent where you installed 🤗
-Transformers, since that command :obj:`transformers-cli` comes from the library.
+Once the repo is cloned, you can add the model, configuration and tokenizer files. For instance, saving the model and
+tokenizer files:

 .. code-block::

-    transformers-cli login
+    >>> model.save_pretrained("path/to/repo/clone/your-model-name")
+    >>> tokenizer.save_pretrained("path/to/repo/clone/your-model-name")

-Then log in using the same credentials as on huggingface.co. To upload your model, just type
+Or, if you're using the Trainer API

 .. code-block::

-    transformers-cli upload path/to/awesome-name-you-picked/
+    >>> trainer.save_model("path/to/awesome-name-you-picked")

-This will upload the folder containing the weights, tokenizer and configuration we prepared in the previous section.
-
-By default you will be prompted to confirm that you want these files to be uploaded. If you are uploading multiple
-models and need to script that process, you can add `-y` to bypass the prompt. For example:
+You can then add these files to the staging environment and verify that they have been correctly staged with the ``git
+status`` command:

 .. code-block::

-    transformers-cli upload -y path/to/awesome-name-you-picked/
+    git add --all
+    git status

-
-If you want to upload a single file (a new version of your model, or the other framework checkpoint you want to add),
-just type:
+Finally, the files should be comitted:

 .. code-block::

-    transformers-cli upload path/to/awesome-name-you-picked/that-file 
+    git commit -m "First version of the your-model-name model and tokenizer."

-or
+And pushed to the remote:

 .. code-block::

-   transformers-cli upload path/to/awesome-name-you-picked/that-file --filename awesome-name-you-picked/new_name
+    git push

-if you want to change its filename.
+This will upload the folder containing the weights, tokenizer and configuration we have just prepared.

-This uploads the model to your personal account. If you want your model to be namespaced by your organization name
-rather than your username, add the following flag to any command:
-
-.. code-block::
-
-    --organization organization_name
-
-so for instance:
-
-.. code-block::
-
-    transformers-cli upload path/to/awesome-name-you-picked/ --organization organization_name
-
-Your model will then be accessible through its identifier, which is, as we saw above,
-`username/awesome-name-you-picked` or `organization/awesome-name-you-picked`.

 Add a model card
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -203,20 +213,15 @@ Anyone can load it from code:

 .. code-block::

-    tokenizer = AutoTokenizer.from_pretrained("namespace/awesome-name-you-picked")
-    model = AutoModel.from_pretrained("namespace/awesome-name-you-picked")
+    >>> tokenizer = AutoTokenizer.from_pretrained("namespace/awesome-name-you-picked")
+    >>> model = AutoModel.from_pretrained("namespace/awesome-name-you-picked")

-Additional commands
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

-You can list all the files you uploaded on the hub like this:
+You may specify a revision by using the ``revision`` flag in the ``from_pretrained`` method:

 .. code-block::

-    transformers-cli s3 ls
-
-You can also delete unneeded files with
-
-.. code-block::
-
-    transformers-cli s3 rm awesome-name-you-picked/filename
+    >>> tokenizer = AutoTokenizer.from_pretrained(
+    >>>   "julien-c/EsperBERTo-small",
+    >>>   revision="v2.0.1" # tag name, or branch name, or commit hash
+    >>> )
@@ -510,8 +510,8 @@ As a default all models apply *Top-K* sampling when used in pipelines, as config


 Here, the model generates a random text with a total maximal length of *50* tokens from context *"As far as I am
-concerned, I will"*. The default arguments of ``PreTrainedModel.generate()`` can be directly overriden in the pipeline,
-as is shown above for the argument ``max_length``.
+concerned, I will"*. The default arguments of ``PreTrainedModel.generate()`` can be directly overridden in the
+pipeline, as is shown above for the argument ``max_length``.

 Here is an example of text generation using ``XLNet`` and its tokenzier.

@@ -409,6 +409,22 @@ decorators are used to set the requirements of tests CPU/GPU/TPU-wise:
 * ``require_torch_non_multigpu`` - as ``require_torch`` plus requires 0 or 1 GPUs
 * ``require_torch_tpu`` - as ``require_torch`` plus requires at least 1 TPU

+Let's depict the GPU requirements in the following table:
+
+
+----------+---------------------------------+
+| n gpus   |  decorator                      |
+==========+=================================+
+| ``>= 0`` | ``@require_torch``              |
+----------+---------------------------------+
+| ``>= 1`` | ``@require_torch_gpu``          |
+----------+---------------------------------+
+| ``>= 2`` | ``@require_torch_multigpu``     |
+----------+---------------------------------+
+| ``< 2``  | ``@require_torch_non_multigpu`` |
+----------+---------------------------------+
+
+
 For example, here is a test that must be run only when there are 2 or more GPUs available and pytorch is installed:

 .. code-block:: python
@@ -450,7 +466,8 @@ Inside tests:

 .. code-block:: bash

-   torch.cuda.device_count()
+   from transformers.testing_utils import get_gpu_count
+   n_gpu = get_gpu_count() # works with torch and tf



@@ -470,7 +487,7 @@ This is still under development but you can study 2 different tests that perform
  <https://github.com/huggingface/transformers/blob/master/examples/seq2seq/test_finetune_trainer.py>`__ - a normal
  (non-PL) test

-To jump right into the execution point, search for the ``execute_async_std`` function in those tests.
+To jump right into the execution point, search for the ``execute_subprocess_async`` function in those tests.

 You will need at least 2 GPUs to see these tests in action:

@@ -646,6 +663,55 @@ as in the previous example.



+Files and directories
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+In tests often we need to know where things are relative to the current test file, and it's not trivial since the test
+could be invoked from more than one directory or could reside in sub-directories with different depths. A helper class
+:obj:`transformers.test_utils.TestCasePlus` solves this problem by sorting out all the basic paths and provides easy
+accessors to them:
+
+* ``pathlib`` objects (all fully resolved):
+
+   - ``test_file_path`` - the current test file path, i.e. ``__file__``
+   - ``test_file_dir`` - the directory containing the current test file
+   - ``tests_dir`` - the directory of the ``tests`` test suite
+   - ``examples_dir`` - the directory of the ``examples`` test suite
+   - ``repo_root_dir`` - the directory of the repository
+   - ``src_dir`` - the directory of ``src`` (i.e. where the ``transformers`` sub-dir resides)
+
+* stringified paths---same as above but these return paths as strings, rather than ``pathlib`` objects:
+
+   - ``test_file_path_str``
+   - ``test_file_dir_str``
+   - ``tests_dir_str``
+   - ``examples_dir_str``
+   - ``repo_root_dir_str``
+   - ``src_dir_str``
+
+To start using those all you need is to make sure that the test resides in a subclass of
+:obj:`transformers.test_utils.TestCasePlus`. For example:
+
+.. code-block:: python
+
+    from transformers.testing_utils import TestCasePlus
+    class PathExampleTest(TestCasePlus):
+        def test_something_involving_local_locations(self):
+            data_dir = self.examples_dir / "seq2seq/test_data/wmt_en_ro"
+
+If you don't need to manipulated paths via ``pathlib`` or you just need a path as a string, you can always invoked
+``str()`` on the ``pathlib`` oboject or use the accessors ending with ``_str``. For example:
+
+.. code-block:: python
+
+    from transformers.testing_utils import TestCasePlus
+    class PathExampleTest(TestCasePlus):
+        def test_something_involving_stringified_locations(self):
+            examples_dir = self.examples_dir_str
+
+
+
+
 Temporary files and directories
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -665,8 +731,8 @@ Here is an example of its usage:

    from transformers.testing_utils import TestCasePlus
    class ExamplesTests(TestCasePlus):
-    def test_whatever(self):
-        tmp_dir = self.get_auto_remove_tmp_dir()
+        def test_whatever(self):
+            tmp_dir = self.get_auto_remove_tmp_dir()

 This code creates a unique temporary directory, and sets :obj:`tmp_dir` to its location.

@@ -1008,6 +1074,24 @@ If you want to test the impact of environment variables for a specific test you
        def test_env_override(self):
            env_level_str = os.getenv("TRANSFORMERS_VERBOSITY", None)

+At times an external program needs to be called, which requires setting ``PYTHONPATH`` in ``os.environ`` to include
+multiple local paths. A helper class :obj:`transformers.test_utils.TestCasePlus` comes to help:
+
+.. code-block:: python
+
+    from transformers.testing_utils import TestCasePlus
+    class EnvExampleTest(TestCasePlus):
+        def test_external_prog(self):
+            env = self.get_env()
+            # now call the external program, passing ``env`` to it
+
+Depending on whether the test file was under the ``tests`` test suite or ``examples`` it'll correctly set up
+``env[PYTHONPATH]`` to include one of these two directories, and also the ``src`` directory to ensure the testing is
+done against the current repo, and finally with whatever ``env[PYTHONPATH]`` was already set to before the test was
+called if anything.
+
+This helper method creates a copy of the ``os.environ`` object, so the original remains intact.
+

 Getting reproducible results
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -1,41 +1,22 @@
 # Examples

-Version 2.9 of 🤗 Transformers introduces a new [`Trainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py) class for PyTorch, and its equivalent [`TFTrainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer_tf.py) for TF 2.
+Version 2.9 of 🤗 Transformers introduced a new [`Trainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py) class for PyTorch, and its equivalent [`TFTrainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer_tf.py) for TF 2.
 Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.2+.

 Here is the list of all our examples:
 - **grouped by task** (all official examples work for multiple models)
- with information on whether they are **built on top of `Trainer`/`TFTrainer`** (if not, they still work, they might just lack some features),
- whether they also include examples for **`pytorch-lightning`**, which is a great fully-featured, general-purpose training library for PyTorch,
+- with information on whether they are **built on top of `Trainer`/`TFTrainer`** (if not, they still work, they might
+  just lack some features),
+- 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,
 - links to **Cloud deployments** to be able to deploy large-scale trainings in the Cloud with little to no setup.

-This is still a work-in-progress – in particular documentation is still sparse – so please **contribute improvements/pull requests.**
-
-
-## The Big Table of Tasks
-
-| Task | Example datasets | Trainer support | TFTrainer support | pytorch-lightning | Colab
-|---|---|:---:|:---:|:---:|:---:|
-| [**`language-modeling`**](https://github.com/huggingface/transformers/tree/master/examples/language-modeling)       | Raw text        | ✅ | -  | -  | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/01_how_to_train.ipynb)
-| [**`text-classification`**](https://github.com/huggingface/transformers/tree/master/examples/text-classification)   | GLUE, XNLI      | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/trainer/01_text_classification.ipynb)
-| [**`token-classification`**](https://github.com/huggingface/transformers/tree/master/examples/token-classification) | CoNLL NER       | ✅ | ✅ | ✅ | -
-| [**`multiple-choice`**](https://github.com/huggingface/transformers/tree/master/examples/multiple-choice)           | SWAG, RACE, ARC | ✅ | ✅ | -  | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ViktorAlm/notebooks/blob/master/MPC_GPU_Demo_for_TF_and_PT.ipynb)
-| [**`question-answering`**](https://github.com/huggingface/transformers/tree/master/examples/question-answering)     | SQuAD           | ✅ | ✅ | -  | -
-| [**`text-generation`**](https://github.com/huggingface/transformers/tree/master/examples/text-generation)           | -               | n/a | n/a | n/a | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/02_how_to_generate.ipynb)
-| [**`distillation`**](https://github.com/huggingface/transformers/tree/master/examples/distillation)                 | All             | - | -  | -  | -
-| [**`summarization`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)                     | CNN/Daily Mail  | ✅  | -  | ✅  | -
-| [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)                       | WMT             | ✅  | -  | ✅  | -
-| [**`bertology`**](https://github.com/huggingface/transformers/tree/master/examples/bertology)                       | -               | - | -  | -  | -
-| [**`adversarial`**](https://github.com/huggingface/transformers/tree/master/examples/adversarial)                   | HANS            | ✅ | -  | -  | -
-
-
-<br>

 ## Important note

 **Important**
-To make sure you can successfully run the latest versions of the example scripts, you have to install the library from source and install some example-specific requirements.
+
+To make sure you can successfully run the latest versions of the example scripts, you have to **install the library from source** and install some example-specific requirements.
 Execute the following steps in a new virtual environment:

 ```bash
@@ -45,6 +26,30 @@ pip install .
 pip install -r ./examples/requirements.txt
 ```

+Alternatively, you can run the version of the examples as they were for your current version of Transformers via (for instance with v3.4.0):
+```bash
+git checkout tags/v3.4.0
+```
+
+## The Big Table of Tasks
+
+| Task | Example datasets | Trainer support | TFTrainer support | 🤗 Datasets | Colab
+|---|---|:---:|:---:|:---:|:---:|
+| [**`language-modeling`**](https://github.com/huggingface/transformers/tree/master/examples/language-modeling)       | Raw text        | ✅ | -  | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/01_how_to_train.ipynb)
+| [**`text-classification`**](https://github.com/huggingface/transformers/tree/master/examples/text-classification)   | GLUE, XNLI      | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://github.com/huggingface/notebooks/blob/master/examples/text_classification.ipynb)
+| [**`token-classification`**](https://github.com/huggingface/transformers/tree/master/examples/token-classification) | CoNLL NER       | ✅ | ✅ | ✅ | -
+| [**`multiple-choice`**](https://github.com/huggingface/transformers/tree/master/examples/multiple-choice)           | SWAG, RACE, ARC | ✅ | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ViktorAlm/notebooks/blob/master/MPC_GPU_Demo_for_TF_and_PT.ipynb)
+| [**`question-answering`**](https://github.com/huggingface/transformers/tree/master/examples/question-answering)     | SQuAD           | ✅ | ✅ | - | -
+| [**`text-generation`**](https://github.com/huggingface/transformers/tree/master/examples/text-generation)           | -               | n/a | n/a | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/02_how_to_generate.ipynb)
+| [**`distillation`**](https://github.com/huggingface/transformers/tree/master/examples/distillation)                 | All             | - | -  | - | -
+| [**`summarization`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)                     | CNN/Daily Mail  | ✅  | - | - | -
+| [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)                       | WMT             | ✅  | - | - | -
+| [**`bertology`**](https://github.com/huggingface/transformers/tree/master/examples/bertology)                       | -               | - | - | - | -
+| [**`adversarial`**](https://github.com/huggingface/transformers/tree/master/examples/adversarial)                   | HANS            | ✅ | - | - | -
+
+
+<br>
+
 ## One-click Deploy to Cloud (wip)

 **Coming soon!**
@@ -57,13 +62,14 @@ When using PyTorch, we support TPUs thanks to `pytorch/xla`. For more context an
 very detailed [pytorch/xla README](https://github.com/pytorch/xla/blob/master/README.md).

 In this repo, we provide a very simple launcher script named [xla_spawn.py](https://github.com/huggingface/transformers/tree/master/examples/xla_spawn.py) that lets you run our example scripts on multiple TPU cores without any boilerplate.
-Just pass a `--num_cores` flag to this script, then your regular training script with its arguments (this is similar to the `torch.distributed.launch` helper for torch.distributed).
+Just pass a `--num_cores` flag to this script, then your regular training script with its arguments (this is similar to the `torch.distributed.launch` helper for torch.distributed). 
+Note that this approach does not work for examples that use `pytorch-lightning`.

 For example for `run_glue`:

 ```bash
 python examples/xla_spawn.py --num_cores 8 \
-	examples/text-classification/run_glue.py
+	examples/text-classification/run_glue.py \
 	--model_name_or_path bert-base-cased \
 	--task_name mnli \
 	--data_dir ./data/glue_data/MNLI \
@@ -291,10 +291,9 @@ def hans_convert_examples_to_features(

    Args:
        examples: List of ``InputExamples`` containing the examples.
-        tokenizer: Instance of a tokenizer that will tokenize the examples.
-        max_length: Maximum example length.
        label_list: List of labels. Can be obtained from the processor using the ``processor.get_labels()`` method.
-        output_mode: String indicating the output mode. Either ``regression`` or ``classification``.
+        max_length: Maximum example length.
+        tokenizer: Instance of a tokenizer that will tokenize the examples.

    Returns:
        A list of task-specific ``InputFeatures`` which can be fed to the model.
@@ -20,7 +20,7 @@ import torch
 import torch.nn as nn
 from torch.nn import CrossEntropyLoss, MSELoss

-from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
+from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
 from transformers.modeling_albert import (
    ALBERT_INPUTS_DOCSTRING,
    ALBERT_START_DOCSTRING,
@@ -87,7 +87,7 @@ class AlbertModelWithPabee(AlbertModel):
        message = f"*** Patience = {self.patience} Avg. Inference Layers = {avg_inf_layers:.2f} Speed Up = {1 - avg_inf_layers / self.config.num_hidden_layers:.2f} ***"
        print(message)

-    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -230,7 +230,7 @@ class AlbertForSequenceClassificationWithPabee(AlbertPreTrainedModel):

        self.init_weights()

-    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -22,7 +22,7 @@ import torch
 from torch import nn
 from torch.nn import CrossEntropyLoss, MSELoss

-from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
+from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
 from transformers.modeling_bert import (
    BERT_INPUTS_DOCSTRING,
    BERT_START_DOCSTRING,
@@ -92,7 +92,7 @@ class BertModelWithPabee(BertModel):
        message = f"*** Patience = {self.patience} Avg. Inference Layers = {avg_inf_layers:.2f} Speed Up = {1 - avg_inf_layers / self.config.num_hidden_layers:.2f} ***"
        print(message)

-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -155,7 +155,7 @@ class BertModelWithPabee(BertModel):
        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)

        # If a 2D ou 3D attention mask is provided for the cross-attention
-        # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
        if self.config.is_decoder and encoder_hidden_states is not None:
            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
@@ -254,7 +254,7 @@ class BertForSequenceClassificationWithPabee(BertPreTrainedModel):

        self.init_weights()

-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -4,7 +4,7 @@ import sys
 from unittest.mock import patch

 import run_glue_with_pabee
-from transformers.testing_utils import TestCasePlus
+from transformers.testing_utils import TestCasePlus, require_torch_non_multigpu_but_fix_me


 logging.basicConfig(level=logging.DEBUG)
@@ -20,6 +20,7 @@ def get_setup_file():


 class PabeeTests(TestCasePlus):
+    @require_torch_non_multigpu_but_fix_me
    def test_run_glue(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)
@@ -14,3 +14,17 @@ sys.path.insert(1, git_repo_path)
 # silence FutureWarning warnings in tests since often we can't act on them until
 # they become normal warnings - i.e. the tests still need to test the current functionality
 warnings.simplefilter(action="ignore", category=FutureWarning)
+
+
+def pytest_addoption(parser):
+    from transformers.testing_utils import pytest_addoption_shared
+
+    pytest_addoption_shared(parser)
+
+
+def pytest_terminal_summary(terminalreporter):
+    from transformers.testing_utils import pytest_terminal_summary_main
+
+    make_reports = terminalreporter.config.getoption("--make-reports")
+    if make_reports:
+        pytest_terminal_summary_main(terminalreporter, id=make_reports)
@@ -103,9 +103,13 @@ class DataTrainingArguments:
        default=None,
        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
    )
-    chinese_ref_file: Optional[str] = field(
+    train_ref_file: Optional[str] = field(
        default=None,
-        metadata={"help": "An optional input ref data file for whole word mask in Chinees."},
+        metadata={"help": "An optional input train ref data file for whole word mask in Chinese."},
+    )
+    eval_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input eval ref data file for whole word mask in Chinese."},
    )
    line_by_line: bool = field(
        default=False,
@@ -148,16 +152,16 @@ def get_dataset(
    evaluate: bool = False,
    cache_dir: Optional[str] = None,
 ):
-    def _dataset(file_path):
+    def _dataset(file_path, ref_path=None):
        if args.line_by_line:
-            if args.chinese_ref_file is not None:
+            if ref_path is not None:
                if not args.whole_word_mask or not args.mlm:
                    raise ValueError("You need to set world whole masking and mlm to True for Chinese Whole Word Mask")
                return LineByLineWithRefDataset(
                    tokenizer=tokenizer,
                    file_path=file_path,
                    block_size=args.block_size,
-                    ref_path=args.chinese_ref_file,
+                    ref_path=ref_path,
                )

            return LineByLineTextDataset(tokenizer=tokenizer, file_path=file_path, block_size=args.block_size)
@@ -171,11 +175,11 @@ def get_dataset(
            )

    if evaluate:
-        return _dataset(args.eval_data_file)
+        return _dataset(args.eval_data_file, args.eval_ref_file)
    elif args.train_data_files:
        return ConcatDataset([_dataset(f) for f in glob(args.train_data_files)])
    else:
-        return _dataset(args.train_data_file)
+        return _dataset(args.train_data_file, args.train_ref_file)


 def main():
@@ -118,7 +118,7 @@ def main(args):
    # If we want to fine-tune these model, we have to use same tokenizer : LTP (https://github.com/HIT-SCIR/ltp)
    with open(args.file_name, "r", encoding="utf-8") as f:
        data = f.readlines()
-
+    data = [line.strip() for line in data if len(line) > 0 and not line.isspace()]  # avoid delimiter like '\u2029'
    ltp_tokenizer = LTP(args.ltp)  # faster in GPU device
    bert_tokenizer = BertTokenizer.from_pretrained(args.bert)

@@ -2,7 +2,7 @@ import torch
 from torch import nn
 from torch.nn import CrossEntropyLoss, MSELoss

-from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
+from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
 from transformers.modeling_bert import (
    BERT_INPUTS_DOCSTRING,
    BERT_START_DOCSTRING,
@@ -134,7 +134,7 @@ class DeeBertModel(BertPreTrainedModel):
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -198,7 +198,7 @@ class DeeBertModel(BertPreTrainedModel):
        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)

        # If a 2D ou 3D attention mask is provided for the cross-attention
-        # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
        if encoder_attention_mask.dim() == 3:
            encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
        if encoder_attention_mask.dim() == 2:
@@ -260,7 +260,7 @@ class BertHighway(nn.Module):

        # BertModel
        bmodel_output = (pooler_input, pooler_output) + encoder_outputs[1:]
-        # "return" bodel_output
+        # "return" bmodel_output

        # Dropout and classification
        pooled_output = bmodel_output[1]
@@ -288,7 +288,7 @@ class DeeBertForSequenceClassification(BertPreTrainedModel):

        self.init_weights()

-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -4,7 +4,7 @@ import torch.nn as nn
 from torch.nn import CrossEntropyLoss, MSELoss

 from transformers.configuration_roberta import RobertaConfig
-from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
+from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
 from transformers.modeling_roberta import ROBERTA_INPUTS_DOCSTRING, ROBERTA_START_DOCSTRING, RobertaEmbeddings

 from .modeling_highway_bert import BertPreTrainedModel, DeeBertModel, HighwayException, entropy
@@ -45,7 +45,7 @@ class DeeRobertaForSequenceClassification(BertPreTrainedModel):
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)

-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -5,7 +5,7 @@ import unittest
 from unittest.mock import patch

 import run_glue_deebert
-from transformers.testing_utils import slow
+from transformers.testing_utils import require_torch_non_multigpu_but_fix_me, slow


 logging.basicConfig(level=logging.DEBUG)
@@ -26,6 +26,7 @@ class DeeBertTests(unittest.TestCase):
        logger.addHandler(stream_handler)

    @slow
+    @require_torch_non_multigpu_but_fix_me
    def test_glue_deebert_train(self):

        train_args = """
@@ -12,7 +12,7 @@ This folder contains the original code used to train Distil* as well as examples

 **October 3, 2019 - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2. **The paper supersedes our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances. Please use the paper as a reference when comparing/reporting results on DistilBERT.**

-**September 19, 2019 - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 99% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
+**September 19, 2019 - Update:** We fixed bugs in the code and released an updated version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 99% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!


 ## What is Distil*
@@ -265,7 +265,7 @@ class Distiller:
        -------
            token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
            attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
-            clm_labels: `torch.tensor(bs, seq_length)` - The causal languge modeling labels. There is a -100 where there is nothing to predict.
+            clm_labels: `torch.tensor(bs, seq_length)` - The causal language modeling labels. There is a -100 where there is nothing to predict.
        """
        token_ids, lengths = batch
        token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
@@ -401,9 +401,9 @@ class Distiller:
        # https://github.com/peterliht/knowledge-distillation-pytorch/blob/master/model/net.py#L100
        # https://github.com/peterliht/knowledge-distillation-pytorch/issues/2
        if self.params.restrict_ce_to_mask:
-            mask = (lm_labels > -1).unsqueeze(-1).expand_as(s_logits)  # (bs, seq_lenth, voc_size)
+            mask = (lm_labels > -1).unsqueeze(-1).expand_as(s_logits)  # (bs, seq_length, voc_size)
        else:
-            mask = attention_mask.unsqueeze(-1).expand_as(s_logits)  # (bs, seq_lenth, voc_size)
+            mask = attention_mask.unsqueeze(-1).expand_as(s_logits)  # (bs, seq_length, voc_size)
        s_logits_slct = torch.masked_select(s_logits, mask)  # (bs * seq_length * voc_size) modulo the 1s in mask
        s_logits_slct = s_logits_slct.view(-1, s_logits.size(-1))  # (bs * seq_length, voc_size) modulo the 1s in mask
        t_logits_slct = torch.masked_select(t_logits, mask)  # (bs * seq_length * voc_size) modulo the 1s in mask
@@ -61,7 +61,7 @@ class LmSeqsDataset(Dataset):

    def remove_long_sequences(self):
        """
-        Sequences that are too long are splitted by chunk of max_model_input_size.
+        Sequences that are too long are split by chunk of max_model_input_size.
        """
        max_len = self.params.max_model_input_size
        indices = self.lengths > max_len
@@ -138,8 +138,8 @@ class LmSeqsDataset(Dataset):
        # logger.info(f'{data_len} tokens ({nb_unique_tokens} unique)')

        # unk_idx = self.params.special_tok_ids['unk_token']
-        # nb_unkown = sum([(t==unk_idx).sum() for t in self.token_ids])
-        # logger.info(f'{nb_unkown} unknown tokens (covering {100*nb_unkown/data_len:.2f}% of the data)')
+        # nb_unknown = sum([(t==unk_idx).sum() for t in self.token_ids])
+        # logger.info(f'{nb_unknown} unknown tokens (covering {100*nb_unknown/data_len:.2f}% of the data)')

    def batch_sequences(self, batch):
        """
@@ -96,7 +96,7 @@ if __name__ == "__main__":
        compressed_sd["lm_head.weight"] = state_dict["lm_head.weight"]

    print(f"N layers selected for distillation: {std_idx}")
-    print(f"Number of params transfered for distillation: {len(compressed_sd.keys())}")
+    print(f"Number of params transferred for distillation: {len(compressed_sd.keys())}")

-    print(f"Save transfered checkpoint to {args.dump_checkpoint}.")
+    print(f"Save transferred checkpoint to {args.dump_checkpoint}.")
    torch.save(compressed_sd, args.dump_checkpoint)
@@ -1,16 +1,19 @@
-
 ## Language model training

-Based on the script [`run_language_modeling.py`](https://github.com/huggingface/transformers/blob/master/examples/language-modeling/run_language_modeling.py).
+Fine-tuning (or training from scratch) the library models for language modeling on a text dataset for GPT, GPT-2,
+ALBERT, BERT, DistilBERT, RoBERTa, XLNet... GPT and GPT-2 are trained or fine-tuned using a causal language modeling
+(CLM) loss while ALBERT, BERT, DistilBERT and RoBERTa are trained or fine-tuned using a masked language modeling (MLM)
+loss. XLNet uses permutation language modeling (PLM), you can find more information about the differences between those
+objectives in our [model summary](https://huggingface.co/transformers/model_summary.html).

-Fine-tuning (or training from scratch) the library models for language modeling on a text dataset for GPT, GPT-2, BERT, DistilBERT and RoBERTa. GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT, DistilBERT and RoBERTa
-are fine-tuned using a masked language modeling (MLM) loss.
+These scripts leverage the 🤗 Datasets library and the Trainer API. You can easily customize them to your needs if you
+need extra processing on your datasets.

-Before running the following example, you should get a file that contains text on which the language model will be
-trained or fine-tuned. A good example of such text is the [WikiText-2 dataset](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/).
+**Note:** The old script `run_language_modeling.py` is still available
+[here](https://github.com/huggingface/transformers/blob/master/examples/contrib/legacy/run_language_modeling.py).

-We will refer to two different files: `$TRAIN_FILE`, which contains text for training, and `$TEST_FILE`, which contains
-text that will be used for evaluation.
+The following examples, will run on a datasets hosted on our [hub](https://huggingface.co/datasets) or with your own
+text files for training and validation. We give examples of both below.

 ### GPT-2/GPT and causal language modeling

@@ -18,63 +21,106 @@ The following example fine-tunes GPT-2 on WikiText-2. We're using the raw WikiTe
 the tokenization). The loss here is that of causal language modeling.

 ```bash
-export TRAIN_FILE=/path/to/dataset/wiki.train.raw
-export TEST_FILE=/path/to/dataset/wiki.test.raw
-
-python run_language_modeling.py \
-    --output_dir=output \
-    --model_type=gpt2 \
-    --model_name_or_path=gpt2 \
+python run_clm.py \
+    --model_name_or_path gpt2 \
+    --dataset_name wikitext \
+    --dataset_config_name wikitext-2-raw-v1 \
    --do_train \
-    --train_data_file=$TRAIN_FILE \
    --do_eval \
-    --eval_data_file=$TEST_FILE
+    --output_dir /tmp/test-clm
 ```

 This takes about half an hour to train on a single K80 GPU and about one minute for the evaluation to run. It reaches
 a score of ~20 perplexity once fine-tuned on the dataset.

+To run on your own training and validation files, use the following command:
+
+```bash
+python run_clm.py \
+    --model_name_or_path gpt2 \
+    --train_file path_to_train_file \
+    --validation_file path_to_validation_file \
+    --do_train \
+    --do_eval \
+    --output_dir /tmp/test-clm
+```
+
+
 ### RoBERTa/BERT/DistilBERT and masked language modeling

 The following example fine-tunes RoBERTa on WikiText-2. Here too, we're using the raw WikiText-2. The loss is different
 as BERT/RoBERTa have a bidirectional mechanism; we're therefore using the same loss that was used during their
 pre-training: masked language modeling.

-In accordance to the RoBERTa paper, we use dynamic masking rather than static masking. The model may, therefore, converge
-slightly slower (over-fitting takes more epochs).
-
-We use the `--mlm` flag so that the script may change its loss function.
-
-If using whole-word masking, use both the`--mlm` and `--wwm` flags.
+In accordance to the RoBERTa paper, we use dynamic masking rather than static masking. The model may, therefore,
+converge slightly slower (over-fitting takes more epochs).

 ```bash
-export TRAIN_FILE=/path/to/dataset/wiki.train.raw
-export TEST_FILE=/path/to/dataset/wiki.test.raw
-
-python run_language_modeling.py \
-    --output_dir=output \
-    --model_type=roberta \
-    --model_name_or_path=roberta-base \
+python run_mlm.py \
+    --model_name_or_path roberta-base \
+    --dataset_name wikitext \
+    --dataset_config_name wikitext-2-raw-v1 \
    --do_train \
-    --train_data_file=$TRAIN_FILE \
    --do_eval \
-    --eval_data_file=$TEST_FILE \
-    --mlm \
-    --whole_word_mask
+    --output_dir /tmp/test-mlm
 ```

-For Chinese models, it's same with English model with only --mlm`. If using whole-word masking, we need to generate a reference files, case it's char level.
+To run on your own training and validation files, use the following command:

-**Q :** Why ref file ?
+```bash
+python run_mlm.py \
+    --model_name_or_path roberta-base \
+    --train_file path_to_train_file \
+    --validation_file path_to_validation_file \
+    --do_train \
+    --do_eval \
+    --output_dir /tmp/test-mlm
+```

-**A :** Suppose we have a Chinese sentence like : `我喜欢你` The original Chinese-BERT will tokenize it as `['我','喜','欢','你']` in char level.
-Actually, `喜欢` is a whole word. For whole word mask proxy, We need res like `['我','喜','##欢','你']`.
-So we need a ref file to tell model which pos of BERT original token should be added `##`.
+If your dataset is organized with one sample per line, you can use the `--line_by_line` flag (otherwise the script
+concatenates all texts and then splits them in blocks of the same length).
+
+**Note:** On TPU, you should use the flag `--pad_to_max_length` in conjunction with the `--line_by_line` flag to make
+sure all your batches have the same length.
+
+### Whole word masking
+
+The BERT authors released a new version of BERT using Whole Word Masking in May 2019. Instead of masking randomly
+selected tokens (which may be part of words), they mask randomly selected words (masking all the tokens corresponding
+to that word). This technique has been refined for Chinese in [this paper](https://arxiv.org/abs/1906.08101).
+
+To fine-tune a model using whole word masking, use the following script:
+
+python run_mlm_wwm.py \
+    --model_name_or_path roberta-base \
+    --dataset_name wikitext \
+    --dataset_config_name wikitext-2-raw-v1 \
+    --do_train \
+    --do_eval \
+    --output_dir /tmp/test-mlm-wwm
+```
+
+For Chinese models, we need to generate a reference files (which requires the ltp library), because it's tokenized at
+the character level.
+
+**Q :** Why a reference file?
+
+**A :** Suppose we have a Chinese sentence like: `我喜欢你` The original Chinese-BERT will tokenize it as
+`['我','喜','欢','你']` (character level). But `喜欢` is a whole word. For whole word masking proxy, we need a result
+like `['我','喜','##欢','你']`, so we need a reference file to tell the model which position of the BERT original token
+should be added `##`.

 **Q :** Why LTP ?

-**A :** Cause the best known Chinese WWM BERT is [Chinese-BERT-wwm](https://github.com/ymcui/Chinese-BERT-wwm) by HIT. It works well on so many Chines Task like CLUE (Chinese GLUE).
-They use LTP, so if we want to fine-tune their model, we need LTP.
+**A :** Cause the best known Chinese WWM BERT is [Chinese-BERT-wwm](https://github.com/ymcui/Chinese-BERT-wwm) by HIT.
+It works well on so many Chines Task like CLUE (Chinese GLUE). They use LTP, so if we want to fine-tune their model,
+we need LTP.
+
+Now LTP only only works well on `transformers==3.2.0`. So we don't add it to requirements.txt.
+You need to create a separate environment with this version of Transformers to run the `run_chinese_ref.py` script that
+will create the reference files. The script is in `examples/contrib`. Once in the proper environment, run the
+following:
+

 ```bash
 export TRAIN_FILE=/path/to/dataset/wiki.train.raw
@@ -82,34 +128,30 @@ export LTP_RESOURCE=/path/to/ltp/tokenizer
 export BERT_RESOURCE=/path/to/bert/tokenizer
 export SAVE_PATH=/path/to/data/ref.txt

-python chinese_ref.py \
-    --file_name=$TRAIN_FILE \
-    --ltp=$LTP_RESOURCE
-    --bert=$BERT_RESOURCE \
-    --save_path=$SAVE_PATH 
+python examples/contrib/run_chinese_ref.py \
+    --file_name=path_to_train_or_eval_file \
+    --ltp=path_to_ltp_tokenizer \
+    --bert=path_to_bert_tokenizer \
+    --save_path=path_to_reference_file
 ```
-Now Chinese Ref is only supported by `LineByLineWithRefDataset` Class, so we need add `line_by_line` flag: 
+
+Then you can run the script like this: 


 ```bash
-export TRAIN_FILE=/path/to/dataset/wiki.train.raw
-export TEST_FILE=/path/to/dataset/wiki.test.raw
-export REF_FILE=/path/to/ref.txt
-
-python run_language_modeling.py \
-    --output_dir=output \
-    --model_type=roberta \
-    --model_name_or_path=roberta-base \
+python run_mlm_wwm.py \
+    --model_name_or_path roberta-base \
+    --train_file path_to_train_file \
+    --validation_file path_to_validation_file \
+    --train_ref_file path_to_train_chinese_ref_file \
+    --validation_ref_file path_to_validation_chinese_ref_file \
    --do_train \
-    --train_data_file=$TRAIN_FILE \
-    --chinese_ref_file=$REF_FILE \
    --do_eval \
-    --eval_data_file=$TEST_FILE \
-    --mlm \
-    --line_by_line \
-    --whole_word_mask
+    --output_dir /tmp/test-mlm-wwm
 ```

+**Note:** On TPU, you should the flag `--pad_to_max_length` to make sure all your batches have the same length.
+
 ### XLNet and permutation language modeling

 XLNet uses a different training objective, which is permutation language modeling. It is an autoregressive method 
@@ -122,15 +164,32 @@ context length for permutation language modeling.
 The `--max_span_length` flag may also be used to limit the length of a span of masked tokens used 
 for permutation language modeling.

-```bash
-export TRAIN_FILE=/path/to/dataset/wiki.train.raw
-export TEST_FILE=/path/to/dataset/wiki.test.raw
+Here is how to fine-tun XLNet on wikitext-2:

-python run_language_modeling.py \
-    --output_dir=output \
+```bash
+python run_plm.py \
    --model_name_or_path=xlnet-base-cased \
+    --dataset_name wikitext \
+    --dataset_config_name wikitext-2-raw-v1 \
    --do_train \
-    --train_data_file=$TRAIN_FILE \
    --do_eval \
-    --eval_data_file=$TEST_FILE \
+    --output_dir /tmp/test-plm
 ```
+
+To fine-tune it on your own training and validation file, run:
+
+```bash
+python run_plm.py \
+    --model_name_or_path=xlnet-base-cased \
+    --train_file path_to_train_file \
+    --validation_file path_to_validation_file \
+    --do_train \
+    --do_eval \
+    --output_dir /tmp/test-plm
+```
+
+If your dataset is organized with one sample per line, you can use the `--line_by_line` flag (otherwise the script
+concatenates all texts and then splits them in blocks of the same length).
+
+**Note:** On TPU, you should use the flag `--pad_to_max_length` in conjunction with the `--line_by_line` flag to make
+sure all your batches have the same length.
@@ -0,0 +1,345 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the 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
+from dataclasses import dataclass, field
+from typing import Optional
+
+from datasets import load_dataset
+
+import transformers
+from transformers import (
+    CONFIG_MAPPING,
+    MODEL_FOR_CAUSAL_LM_MAPPING,
+    AutoConfig,
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    default_data_collator,
+    set_seed,
+)
+from transformers.trainer_utils import is_main_process
+
+
+logger = logging.getLogger(__name__)
+
+
+MODEL_CONFIG_CLASSES = list(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."},
+    )
+
+
+@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)."},
+    )
+    block_size: int = field(
+        default=-1,
+        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."
+
+
+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."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
+    )
+
+    # Log on each process the small summary:
+    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):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", 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).
+    #
+    # In distributed training, the load_dataset function guarantee 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.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+    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)
+    # 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 = AutoModelForCausalLM.from_pretrained(
+            model_args.model_name_or_path,
+            from_tf=bool(".ckpt" in model_args.model_name_or_path),
+            config=config,
+            cache_dir=model_args.cache_dir,
+        )
+    else:
+        logger.info("Training new model from scratch")
+        model = AutoModelForCausalLM.from_config(config)
+
+    model.resize_token_embeddings(len(tokenizer))
+
+    # 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]
+
+    def tokenize_function(examples):
+        return tokenizer(examples[text_column_name])
+
+    tokenized_datasets = datasets.map(
+        tokenize_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        remove_columns=[text_column_name],
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    if data_args.block_size <= 0:
+        block_size = tokenizer.model_max_length
+    else:
+        if data_args.block_size > tokenizer.model_max_length:
+            logger.warn(
+                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.
+        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,
+    )
+
+    # Initialize our Trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=lm_datasets["train"] if training_args.do_train else None,
+        eval_dataset=lm_datasets["validation"] if training_args.do_eval else None,
+        tokenizer=tokenizer,
+        # Data collator will default to DataCollatorWithPadding, so we change it.
+        data_collator=default_data_collator,
+    )
+
+    # Training
+    if training_args.do_train:
+        trainer.train(
+            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
+        )
+        trainer.save_model()  # Saves the tokenizer too for easy upload
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+
+        eval_output = trainer.evaluate()
+
+        perplexity = math.exp(eval_output["eval_loss"])
+        results["perplexity"] = perplexity
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results_clm.txt")
+        if trainer.is_world_process_zero():
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Eval results *****")
+                for key, value in results.items():
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,386 @@
+# coding=utf-8
+# 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.
+"""
+Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) 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=masked-lm
+"""
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+
+import logging
+import math
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Optional
+
+from datasets import load_dataset
+
+import transformers
+from transformers import (
+    CONFIG_MAPPING,
+    MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoModelForMaskedLM,
+    AutoTokenizer,
+    DataCollatorForLanguageModeling,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    set_seed,
+)
+from transformers.trainer_utils import is_main_process
+
+
+logger = logging.getLogger(__name__)
+MODEL_CONFIG_CLASSES = list(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."},
+    )
+
+
+@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)."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated."
+        },
+    )
+    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 masked language modeling loss"}
+    )
+    line_by_line: bool = field(
+        default=False,
+        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+
+    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 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",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
+    )
+
+    # Log on each process the small summary:
+    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):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", 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. You can easily tweak this
+    # behavior (see below)
+    #
+    # In distributed training, the load_dataset function guarantee 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.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+    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)
+    # 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 = AutoModelForMaskedLM.from_pretrained(
+            model_args.model_name_or_path,
+            from_tf=bool(".ckpt" in model_args.model_name_or_path),
+            config=config,
+            cache_dir=model_args.cache_dir,
+        )
+    else:
+        logger.info("Training new model from scratch")
+        model = AutoModelForMaskedLM.from_config(config)
+
+    model.resize_token_embeddings(len(tokenizer))
+
+    # 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]
+
+    if data_args.line_by_line:
+        # When using line_by_line, we just tokenize each nonempty line.
+        padding = "max_length" if data_args.pad_to_max_length else False
+
+        def tokenize_function(examples):
+            # Remove empty lines
+            examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
+            return tokenizer(
+                examples["text"],
+                padding=padding,
+                truncation=True,
+                max_length=data_args.max_seq_length,
+                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
+                # receives the `special_tokens_mask`.
+                return_special_tokens_mask=True,
+            )
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=[text_column_name],
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+    else:
+        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
+        # efficient when it receives the `special_tokens_mask`.
+        def tokenize_function(examples):
+            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=[text_column_name],
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+        if data_args.max_seq_length is None:
+            max_seq_length = tokenizer.model_max_length
+        else:
+            if data_args.max_seq_length > tokenizer.model_max_length:
+                logger.warn(
+                    f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
+                    f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
+                )
+            max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
+        # max_seq_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.
+            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)]
+                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,
+        )
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
+
+    # Initialize our Trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+    )
+
+    # Training
+    if training_args.do_train:
+        trainer.train(
+            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
+        )
+        trainer.save_model()  # Saves the tokenizer too for easy upload
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+
+        eval_output = trainer.evaluate()
+
+        perplexity = math.exp(eval_output["eval_loss"])
+        results["perplexity"] = perplexity
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results_mlm.txt")
+        if trainer.is_world_process_zero():
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Eval results *****")
+                for key, value in results.items():
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,334 @@
+# coding=utf-8
+# 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.
+"""
+Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking 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=masked-lm
+"""
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+
+import json
+import logging
+import math
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Optional
+
+from datasets import Dataset, load_dataset
+
+import transformers
+from transformers import (
+    CONFIG_MAPPING,
+    MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoModelForMaskedLM,
+    AutoTokenizer,
+    DataCollatorForWholeWordMask,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    set_seed,
+)
+from transformers.trainer_utils import is_main_process
+
+
+logger = logging.getLogger(__name__)
+MODEL_CONFIG_CLASSES = list(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."},
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    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"}
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. 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 masked language modeling loss"}
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+
+    def __post_init__(self):
+        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 add_chinese_references(dataset, ref_file):
+    with open(ref_file, "r", encoding="utf-8") as f:
+        refs = [json.loads(line) for line in f.read().splitlines() if (len(line) > 0 and not line.isspace())]
+    assert len(dataset) == len(refs)
+
+    dataset_dict = {c: dataset[c] for c in dataset.column_names}
+    dataset_dict["chinese_ref"] = refs
+    return Dataset.from_dict(dataset_dict)
+
+
+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."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
+    )
+
+    # Log on each process the small summary:
+    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):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", 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).
+    #
+    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
+    # download the dataset.
+    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)
+    # 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 = AutoModelForMaskedLM.from_pretrained(
+            model_args.model_name_or_path,
+            from_tf=bool(".ckpt" in model_args.model_name_or_path),
+            config=config,
+            cache_dir=model_args.cache_dir,
+        )
+    else:
+        logger.info("Training new model from scratch")
+        model = AutoModelForMaskedLM.from_config(config)
+
+    model.resize_token_embeddings(len(tokenizer))
+
+    # 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]
+
+    padding = "max_length" if data_args.pad_to_max_length else False
+
+    def tokenize_function(examples):
+        # Remove empty lines
+        examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
+        return tokenizer(examples["text"], padding=padding, truncation=True, max_length=data_args.max_seq_length)
+
+    tokenized_datasets = datasets.map(
+        tokenize_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        remove_columns=[text_column_name],
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    # Add the chinese references if provided
+    if data_args.train_ref_file is not None:
+        tokenized_datasets["train"] = add_chinese_references(tokenized_datasets["train"], data_args.train_ref_file)
+    if data_args.valid_ref_file is not None:
+        tokenized_datasets["validation"] = add_chinese_references(
+            tokenized_datasets["validation"], data_args.validation_ref_file
+        )
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = DataCollatorForWholeWordMask(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
+
+    # Initialize our Trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+    )
+
+    # Training
+    if training_args.do_train:
+        trainer.train(
+            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
+        )
+        trainer.save_model()  # Saves the tokenizer too for easy upload
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+
+        eval_output = trainer.evaluate()
+
+        perplexity = math.exp(eval_output["eval_loss"])
+        results["perplexity"] = perplexity
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results_mlm_wwm.txt")
+        if trainer.is_world_process_zero():
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Eval results *****")
+                for key, value in results.items():
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,376 @@
+# coding=utf-8
+# 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.
+"""
+Fine-tuning the library models for permutation language modeling.
+"""
+# You can also adapt this script on your own permutation language modeling task. Pointers for this are left as comments.
+
+import logging
+import math
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Optional
+
+from datasets import load_dataset
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    DataCollatorForPermutationLanguageModeling,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    XLNetConfig,
+    XLNetLMHeadModel,
+    set_seed,
+)
+from transformers.trainer_utils import is_main_process
+
+
+logger = logging.getLogger(__name__)
+
+
+@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."
+        },
+    )
+    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."},
+    )
+
+
+@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)."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. 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."},
+    )
+    plm_probability: float = field(
+        default=1 / 6,
+        metadata={
+            "help": "Ratio of length of a span of masked tokens to surrounding context length for "
+            "permutation language modeling."
+        },
+    )
+    max_span_length: int = field(
+        default=5, metadata={"help": "Maximum length of a span of masked tokens for permutation language modeling."}
+    )
+    line_by_line: bool = field(
+        default=False,
+        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+
+    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 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",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
+    )
+
+    # Log on each process the small summary:
+    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):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", 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).
+    #
+    # In distributed training, the load_dataset function guarantee 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.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+    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)
+    # 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 = XLNetConfig()
+        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 = XLNetLMHeadModel.from_pretrained(
+            model_args.model_name_or_path,
+            from_tf=bool(".ckpt" in model_args.model_name_or_path),
+            config=config,
+            cache_dir=model_args.cache_dir,
+        )
+    else:
+        logger.info("Training new model from scratch")
+        model = XLNetLMHeadModel.from_config(config)
+
+    model.resize_token_embeddings(len(tokenizer))
+
+    # 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]
+
+    if data_args.line_by_line:
+        # When using line_by_line, we just tokenize each nonempty line.
+        padding = "max_length" if data_args.pad_to_max_length else False
+
+        def tokenize_function(examples):
+            # Remove empty lines
+            examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
+            return tokenizer(examples["text"], padding=padding, truncation=True, max_length=data_args.max_seq_length)
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=[text_column_name],
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+    else:
+        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        def tokenize_function(examples):
+            return tokenizer(examples[text_column_name])
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=[text_column_name],
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+        if data_args.max_seq_length is None:
+            max_seq_length = tokenizer.model_max_length
+        else:
+            if data_args.max_seq_length > tokenizer.model_max_length:
+                logger.warn(
+                    f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
+                    f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
+                )
+            max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
+        # max_seq_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.
+            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)]
+                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,
+        )
+
+    # Data collator
+    data_collator = DataCollatorForPermutationLanguageModeling(
+        tokenizer=tokenizer,
+        plm_probability=data_args.plm_probability,
+        max_span_length=data_args.max_span_length,
+    )
+
+    # Initialize our Trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+    )
+
+    # Training
+    if training_args.do_train:
+        trainer.train(
+            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
+        )
+        trainer.save_model()  # Saves the tokenizer too for easy upload
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+
+        eval_output = trainer.evaluate()
+
+        perplexity = math.exp(eval_output["eval_loss"])
+        results["perplexity"] = perplexity
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results_plm.txt")
+        if trainer.is_world_process_zero():
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Eval results *****")
+                for key, value in results.items():
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()
@@ -7,6 +7,7 @@ from typing import Any, Dict
 import pytorch_lightning as pl
 from pytorch_lightning.utilities import rank_zero_info

+import pkg_resources
 from transformers import (
    AdamW,
    AutoConfig,
@@ -32,6 +33,15 @@ from transformers.optimization import (

 logger = logging.getLogger(__name__)

+try:
+    pkg = "pytorch_lightning"
+    min_ver = "1.0.4"
+    pkg_resources.require(f"{pkg}>={min_ver}")
+except pkg_resources.VersionConflict:
+    logger.warning(
+        f"{pkg}>={min_ver} is required for a normal functioning of this module, but found {pkg}=={pkg_resources.get_distribution(pkg).version}. Try pip install -r examples/requirements.txt"
+    )
+

 MODEL_MODES = {
    "base": AutoModel,
@@ -337,7 +347,7 @@ def add_generic_args(parser, root_dir) -> None:
 def generic_train(
    model: BaseTransformer,
    args: argparse.Namespace,
-    early_stopping_callback=False,
+    early_stopping_callback=None,
    logger=True,  # can pass WandbLogger() here
    extra_callbacks=[],
    checkpoint_callback=None,
@@ -355,6 +365,8 @@ def generic_train(
        checkpoint_callback = pl.callbacks.ModelCheckpoint(
            filepath=args.output_dir, prefix="checkpoint", monitor="val_loss", mode="min", save_top_k=1
        )
+    if early_stopping_callback:
+        extra_callbacks.append(early_stopping_callback)
    if logging_callback is None:
        logging_callback = LoggingCallback()

@@ -376,7 +388,6 @@ def generic_train(
        callbacks=[logging_callback] + extra_callbacks,
        logger=logger,
        checkpoint_callback=checkpoint_callback,
-        early_stop_callback=early_stopping_callback,
        **train_params,
    )

@@ -266,14 +266,14 @@ def find_top_rpn_proposals(
 ):
    """Args:
        proposals (list[Tensor]): (L, N, Hi*Wi*A, 4).
-        pred_objectness_logits: tensors of lenngth L.
+        pred_objectness_logits: tensors of length L.
        nms_thresh (float): IoU threshold to use for NMS
        pre_nms_topk (int): before nms
        post_nms_topk (int): after nms
        min_box_side_len (float): minimum proposal box side
        training (bool): True if proposals are to be used in training,
    Returns:
-        resuls (List[Dict]): stores post_nms_topk object proposals for image i.
+        results (List[Dict]): stores post_nms_topk object proposals for image i.
    """
    num_images = len(images)
    device = proposals[0].device
@@ -648,7 +648,7 @@ class RPNOutputs(object):
            images (ImageList): :class:`ImageList` instance representing N input images
            pred_objectness_logits (list[Tensor]): A list of L elements. Element i is a tensor of shape (N, A, Hi, W)
            pred_anchor_deltas (list[Tensor]): A list of L elements. Element i is a tensor of shape (N, A*4, Hi, Wi)
-            anchors (list[torch.Tensor]): nested list ofboxes. anchors[i][j] at (n, l) stores anchor array for feature map l
+            anchors (list[torch.Tensor]): nested list of boxes. anchors[i][j] at (n, l) stores anchor array for feature map l
            boundary_threshold (int): if >= 0, then anchors that extend beyond the image boundary by more than boundary_thresh are not used in training.
            gt_boxes (list[Boxes], optional): A list of N elements.
            smooth_l1_beta (float): The transition point between L1 and L2 lossn. When set to 0, the loss becomes L1. When +inf, it is ignored
@@ -1186,7 +1186,7 @@ class ROIOutputs(object):
        attr_probs_all, attrs_all = self._predict_attrs(attr_logits, preds_per_image)
        features = features.split(preds_per_image, dim=0)

-        # fun for each image too, also I can expirement and do multiple images
+        # fun for each image too, also I can experiment and do multiple images
        final_results = []
        zipped = zip(boxes_all, obj_scores_all, attr_probs_all, attrs_all, sizes)
        for i, (boxes, obj_scores, attr_probs, attrs, size) in enumerate(zipped):
@@ -1412,7 +1412,7 @@ class AnchorGenerator(nn.Module):

    def generate_cell_anchors(self, sizes=(32, 64, 128, 256, 512), aspect_ratios=(0.5, 1, 2)):
        """
-        anchors are continious geometric rectangles
+        anchors are continuous geometric rectangles
        centered on one feature map point sample.
        We can later build the set of anchors
        for the entire feature map by tiling these tensors
@@ -1801,7 +1801,7 @@ class GeneralizedRCNN(nn.Module):
                f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when "
                f"initializing {model.__class__.__name__}: {unexpected_keys}\n"
                f"- This IS expected if you are initializing {model.__class__.__name__} from the checkpoint of a model trained on another task "
-                f"or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPretraining model).\n"
+                f"or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n"
                f"- This IS NOT expected if you are initializing {model.__class__.__name__} from the checkpoint of a model that you expect "
                f"to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model)."
            )
@@ -1865,7 +1865,7 @@ class GeneralizedRCNN(nn.Module):
        scales_yx=None,
        **kwargs,
    ):
-        # run images through bacbone
+        # run images through backbone
        original_sizes = image_shapes * scales_yx
        features = self.backbone(images)

@@ -116,7 +116,7 @@ class Preprocess:
            images = self.aug(images)
            # transpose images and convert to torch tensors
            # images = [torch.as_tensor(i.astype("float32")).permute(2, 0, 1).to(self.device) for i in images]
-            # now normalize before pad to aoid useless arithmatic
+            # now normalize before pad to avoid useless arithmetic
            images = [self.normalizer(x) for x in images]
            # now pad them to do the following operations
            images, sizes = self.pad(images)
@@ -236,7 +236,7 @@ def compare(in_tensor):
    ), f"{sum([1 for x in np.isclose(n1, n2, rtol=0.01, atol=0.1).flatten() if x == False])/len(n1.flatten())*100:.4f} % element-wise mismatch"
    raise Exception("tensors are all good")

-    # Hugging face functiions below
+    # Hugging face functions below


 def is_remote_url(url_or_filename):
@@ -520,7 +520,7 @@ def get_image_from_url(url):
    return img


-# to load legace frcnn checkpoint from detectron
+# to load legacy frcnn checkpoint from detectron
 def load_frcnn_pkl_from_url(url):
    fn = url.split("/")[-1]
    if fn not in os.listdir(os.getcwd()):
@@ -33,7 +33,7 @@ def main(args):
    remaining_count = 0  # Number of remaining (not pruned) params in the encoder
    encoder_count = 0  # Number of params in the encoder

-    print("name".ljust(60, " "), "Remaining Weights %", "Remaning Weight")
+    print("name".ljust(60, " "), "Remaining Weights %", "Remaining Weight")
    for name, param in st.items():
        if "encoder" not in name:
            continue
@@ -28,7 +28,7 @@ from torch.nn import CrossEntropyLoss, MSELoss

 from emmental import MaskedBertConfig
 from emmental.modules import MaskedLinear
-from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
+from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
 from transformers.modeling_bert import ACT2FN, BertLayerNorm, load_tf_weights_in_bert
 from transformers.modeling_utils import PreTrainedModel, prune_linear_layer

@@ -498,7 +498,7 @@ class MaskedBertModel(MaskedBertPreTrainedModel):
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

-    @add_start_docstrings_to_callable(MASKED_BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(MASKED_BERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -591,7 +591,7 @@ class MaskedBertModel(MaskedBertPreTrainedModel):
        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0

        # If a 2D ou 3D attention mask is provided for the cross-attention
-        # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
        if self.config.is_decoder and encoder_hidden_states is not None:
            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
@@ -631,7 +631,7 @@ class MaskedBertModel(MaskedBertPreTrainedModel):
                )  # We can specify head_mask for each layer
            head_mask = head_mask.to(
                dtype=next(self.parameters()).dtype
-            )  # switch to fload if need + fp16 compatibility
+            )  # switch to float if need + fp16 compatibility
        else:
            head_mask = [None] * self.config.num_hidden_layers

@@ -671,7 +671,7 @@ class MaskedBertForSequenceClassification(MaskedBertPreTrainedModel):

        self.init_weights()

-    @add_start_docstrings_to_callable(MASKED_BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(MASKED_BERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -756,7 +756,7 @@ class MaskedBertForMultipleChoice(MaskedBertPreTrainedModel):

        self.init_weights()

-    @add_start_docstrings_to_callable(MASKED_BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(MASKED_BERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -846,7 +846,7 @@ class MaskedBertForTokenClassification(MaskedBertPreTrainedModel):

        self.init_weights()

-    @add_start_docstrings_to_callable(MASKED_BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(MASKED_BERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -932,7 +932,7 @@ class MaskedBertForQuestionAnswering(MaskedBertPreTrainedModel):

        self.init_weights()

-    @add_start_docstrings_to_callable(MASKED_BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_model_forward(MASKED_BERT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -225,7 +225,7 @@ def train(args, train_dataset, model, tokenizer, teacher=None):
        desc="Epoch",
        disable=args.local_rank not in [-1, 0],
    )
-    set_seed(args)  # Added here for reproductibility
+    set_seed(args)  # Added here for reproducibility
    for _ in train_iterator:
        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
        for step, batch in enumerate(epoch_iterator):
@@ -705,7 +705,7 @@ def main():
        "--final_lambda",
        default=0.0,
        type=float,
-        help="Regularization intensity (used in conjunction with `regulariation`.",
+        help="Regularization intensity (used in conjunction with `regularization`.",
    )

    parser.add_argument("--global_topk", action="store_true", help="Global TopK on the Scores.")
@@ -816,7 +816,7 @@ def main():
    if args.local_rank == -1 or args.no_cuda:
        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
-    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+    else:  # Initializes the distributed backend which will take care of synchronizing nodes/GPUs
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
        torch.distributed.init_process_group(backend="nccl")
@@ -231,7 +231,7 @@ def train(args, train_dataset, model, tokenizer, teacher=None):
    train_iterator = trange(
        epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]
    )
-    # Added here for reproductibility
+    # Added here for reproducibility
    set_seed(args)

    for _ in train_iterator:
@@ -824,7 +824,7 @@ def main():
        "--final_lambda",
        default=0.0,
        type=float,
-        help="Regularization intensity (used in conjunction with `regulariation`.",
+        help="Regularization intensity (used in conjunction with `regularization`.",
    )

    parser.add_argument("--global_topk", action="store_true", help="Global TopK on the Scores.")
@@ -977,7 +977,7 @@ def main():
    if args.local_rank == -1 or args.no_cuda:
        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
-    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+    else:  # Initializes the distributed backend which will take care of synchronizing nodes/GPUs
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
        torch.distributed.init_process_group(backend="nccl")
@@ -16,7 +16,7 @@ logger = logging.getLogger(__name__)
 class RagPyTorchDistributedRetriever(RagRetriever):
    """
    A distributed retriever built on top of the ``torch.distributed`` communication package. During training all workers
-    initalize their own instance of the retriever, however, only the main worker loads the index into memory. The index is stored
+    initialize their own instance of the retriever, however, only the main worker loads the index into memory. The index is stored
    in cpu memory. The index will also work well in a non-distributed setup.

    Args:
@@ -45,7 +45,7 @@ class RagPyTorchDistributedRetriever(RagRetriever):

    def init_retrieval(self, distributed_port: int):
        """
-        Retriever initalization function, needs to be called from the training process. The function sets some common parameters
+        Retriever initialization function, needs to be called from the training process. The function sets some common parameters
        and environment variables. On top of that, (only) the main process in the process group loads the index into memory.

        Args:
@@ -56,7 +56,7 @@ class RagPyTorchDistributedRetriever(RagRetriever):

        logger.info("initializing retrieval")

-        # initializing a separate process group for retrievel as the default
+        # initializing a separate process group for retrieval as the default
        # nccl backend doesn't support gather/scatter operations while gloo
        # is too slow to replace nccl for the core gpu communication
        if dist.is_initialized():
@@ -101,7 +101,7 @@ class RagPyTorchDistributedRetriever(RagRetriever):
            n_docs (:obj:`int`):
                The number of docs retrieved per query.

-        Ouput:
+        Output:
            retrieved_doc_embeds (:obj:`np.ndarray` of shape :obj:`(batch_size, n_docs, dim)`
                The retrieval embeddings of the retrieved docs per query.
            doc_ids (:obj:`np.ndarray` of shape :obj:`batch_size, n_docs`)
@@ -176,7 +176,7 @@ def get_args():
        choices=["e2e", "retrieval"],
        default="e2e",
        type=str,
-        help="Evaluation mode, e2e calculates exact match and F1 of the downstream task, retrieval calulates precision@k.",
+        help="Evaluation mode, e2e calculates exact match and F1 of the downstream task, retrieval calculates precision@k.",
    )
    parser.add_argument("--k", default=1, type=int, help="k for the precision@k calculation")
    parser.add_argument(
@@ -206,7 +206,7 @@ def get_args():
        "--predictions_path",
        type=str,
        default="predictions.txt",
-        help="Name of the predictions file, to be stored in the checkpoints directry",
+        help="Name of the predictions file, to be stored in the checkpoints directory",
    )
    parser.add_argument(
        "--eval_all_checkpoints",
@@ -16,6 +16,7 @@ from transformers.configuration_dpr import DPRConfig
 from transformers.configuration_rag import RagConfig
 from transformers.file_utils import is_datasets_available, is_faiss_available, is_psutil_available, is_torch_available
 from transformers.retrieval_rag import CustomHFIndex
+from transformers.testing_utils import require_torch_non_multigpu_but_fix_me
 from transformers.tokenization_bart import BartTokenizer
 from transformers.tokenization_bert import VOCAB_FILES_NAMES as DPR_VOCAB_FILES_NAMES
 from transformers.tokenization_dpr import DPRQuestionEncoderTokenizer
@@ -178,6 +179,7 @@ class RagRetrieverTest(TestCase):
            retriever.init_retrieval(port)
        return retriever

+    @require_torch_non_multigpu_but_fix_me
    def test_pytorch_distributed_retriever_retrieve(self):
        n_docs = 1
        retriever = self.get_dummy_pytorch_distributed_retriever(init_retrieval=True)
@@ -193,6 +195,7 @@ class RagRetrieverTest(TestCase):
        self.assertEqual(doc_dicts[1]["id"][0], "0")  # max inner product is reached with first doc
        self.assertListEqual(doc_ids.tolist(), [[1], [0]])

+    @require_torch_non_multigpu_but_fix_me
    def test_custom_hf_index_retriever_retrieve(self):
        n_docs = 1
        retriever = self.get_dummy_custom_hf_index_retriever(init_retrieval=True, from_disk=False)
@@ -208,6 +211,7 @@ class RagRetrieverTest(TestCase):
        self.assertEqual(doc_dicts[1]["id"][0], "0")  # max inner product is reached with first doc
        self.assertListEqual(doc_ids.tolist(), [[1], [0]])

+    @require_torch_non_multigpu_but_fix_me
    def test_custom_pytorch_distributed_retriever_retrieve_from_disk(self):
        n_docs = 1
        retriever = self.get_dummy_custom_hf_index_retriever(init_retrieval=True, from_disk=True)
@@ -26,7 +26,7 @@ device = "cuda" if torch.cuda.is_available() else "cpu"


 def split_text(text: str, n=100, character=" ") -> List[str]:
-    """Split the text every ``n``-th occurence of ``character``"""
+    """Split the text every ``n``-th occurrence of ``character``"""
    text = text.split(character)
    return [character.join(text[i : i + n]).strip() for i in range(0, len(text), n)]

@@ -5,7 +5,7 @@ psutil
 sacrebleu
 rouge-score
 tensorflow_datasets
-pytorch-lightning==0.9.0
+pytorch-lightning==1.0.4
 matplotlib
 git-python==1.0.3
 faiss-cpu
@@ -1,8 +1,8 @@
-## Sequence to Sequence
+## Sequence to Sequence Training and Evaluation

 This directory contains examples for finetuning and evaluating transformers on summarization and translation tasks.
-Please tag @sshleifer with any issues/unexpected behaviors, or send a PR!
-For `bertabs` instructions, see [`bertabs/README.md`](bertabs/README.md).
+Please tag @patil-suraj with any issues/unexpected behaviors, or send a PR!
+For deprecated `bertabs` instructions, see [`bertabs/README.md`](bertabs/README.md).

 ### Supported Architectures

@@ -66,7 +66,7 @@ Multiple eval datasets are available for download from:
 https://github.com/stas00/porting/tree/master/datasets/pegasus


-#### Private Data
+#### Your Data

 If you are using your own data, it must be formatted as one directory with 6 files:
 ```
@@ -232,7 +232,191 @@ Following command fine-tunes `sshleifer/student_marian_en_ro_6_3` on TPU V3-8 an
 ./builtin_trainer/train_distil_marian_enro_tpu.sh
 ```

-### Evaluation Commands
+# DistilBART
+<!---It should be called distilling bart and pegasus, but I don't want to break the link in the paper.-->
+This section describes all code and artifacts from our [Paper](http://arxiv.org/abs/2010.13002)
+
+![DBART](https://huggingface.co/front/thumbnails/distilbart_large.png)
+
+ For the CNN/DailyMail dataset, (relatively longer, more extractive summaries), we found a simple technique that works, which we call "Shrink and Fine-tune", or SFT.
+you just copy alternating layers from `facebook/bart-large-cnn` and fine-tune more on the cnn/dm data. `sshleifer/distill-pegasus-cnn-16-4`, `sshleifer/distilbart-cnn-12-6` and all other checkpoints under `sshleifer` that start with `distilbart-cnn` were trained this way. 
+ For the XSUM dataset, training on pseudo-labels worked best for Pegasus (`sshleifer/distill-pegasus-16-4`), while training with KD worked best for `distilbart-xsum-12-6`
+ For `sshleifer/dbart-xsum-12-3`
+ We ran 100s experiments, and didn't want to document 100s of commands. If you want a command to replicate a figure from the paper that is not documented below, feel free to ask on the [forums](https://discuss.huggingface.co/t/seq2seq-distillation-methodology-questions/1270) and tag `@sshleifer`. 
+ You can see the performance tradeoffs of model sizes [here](https://docs.google.com/spreadsheets/d/1EkhDMwVO02m8jCD1cG3RoFPLicpcL1GQHTQjfvDYgIM/edit#gid=0).
+and more granular timing results [here](https://docs.google.com/spreadsheets/d/1EkhDMwVO02m8jCD1cG3RoFPLicpcL1GQHTQjfvDYgIM/edit#gid=1753259047&range=B2:I23).
+
+### Evaluation
+
+use [run_distributed_eval](./run_distributed_eval.py), with the following convenient alias
+```bash
+deval () {
+	proc=$1
+	m=$2
+	dd=$3
+	sd=$4
+	shift
+	shift
+	shift
+	shift
+	python -m torch.distributed.launch --nproc_per_node=$proc  run_distributed_eval.py \
+		--model_name $m  --save_dir $sd --data_dir $dd $@
+}
+```
+On a 1 GPU system, here are four commands (that assume `xsum`, `cnn_dm` are downloaded, cmd-F for those links in this file).
+
+`distilBART`:
+```bash
+deval 1 sshleifer/distilbart-xsum-12-3 xsum dbart_12_3_xsum_eval --fp16  # --help for more choices.
+deval 1 sshleifer/distilbart-cnn_dm-12-6 cnn_dm dbart_12_6_cnn_eval --fp16
+```
+
+`distill-pegasus`:
+```bash
+deval 1 sshleifer/distill-pegasus-cnn-16-4 cnn_dm dpx_cnn_eval
+deval 1 sshleifer/distill-pegasus-xsum-16-4 xsum dpx_xsum_eval
+```
+
+### Distillation
+ For all of the following commands, you can get roughly equivalent result and faster run times by passing `--num_beams=4`. That's not what we did for the paper.
+ Besides the KD section, you can also run commands with the built-in transformers trainer. See, for example, [builtin_trainer/train_distilbart_cnn.sh](./builtin_trainer/train_distilbart_cnn.sh).
+ Large performance deviations (> 5X slower or more than 0.5 Rouge-2 worse), should be reported.
+ Multi-gpu (controlled with `--gpus` should work, but might require more epochs).
+
+#### Recommended Workflow
+ Get your dataset in the right format. (see 6 files above).
+ Find a teacher model [Pegasus](https://huggingface.co/models?search=pegasus) (slower, better ROUGE) or `facebook/bart-large-xsum`/`facebook/bart-large-cnn` (faster, slightly lower.).
+Choose the checkpoint where the corresponding dataset is most similar (or identical to) your dataset.
+ Follow the sections in order below. You can stop after SFT if you are satisfied, or move on to pseudo-labeling if you want more performance.
+ student size: If you want a close to free 50% speedup, cut the decoder in half. If you want a larger speedup, cut it in 4. 
+ If your SFT run starts at a validation ROUGE-2 that is more than 10 pts below the teacher's validation ROUGE-2,  you have a bug. Switching to a more expensive technique will not help. Try setting a breakpoint and looking at generation and truncation defaults/hyper-parameters, and share your experience on the forums!
+
+  
+#### Initialization
+We use [make_student.py](./make_student.py) to copy alternating layers from the teacher, and save the resulting model to disk
+```bash
+python make_student.py facebook/bart-large-xsum --save_path dbart_xsum_12_3  -e 12 -d 3
+```
+or for `pegasus-xsum`
+```bash
+python make_student.py google/pegasus-xsum --save_path dpx_xsum_16_4  --e 16 --d 4
+```
+we now have an initialized student saved to  `dbart_xsum_12_3`, which we will use for the following commands.
+ Extension: To replicate more complicated initialize experiments in section 6.1, or try your own. Use the `create_student_by_copying_alternating_layers` function.
+
+#### Pegasus 
+ The following commands are written for BART and will require, at minimum, the following modifications
+ reduce batch size, and increase gradient accumulation steps so that the product `gpus * batch size * gradient_accumulation_steps = 256`. We used `--learning-rate` = 1e-4 * gradient accumulation steps.
+ don't use fp16
+ `--tokenizer_name google/pegasus-large`
+
+### SFT (No Teacher Distillation)
+You don't need `distillation.py`, you can just run:
+
+```bash
+python finetune.py \
+  --data_dir xsum \
+  --freeze_encoder --freeze_embeds \
+  --learning_rate=3e-4 \
+  --do_train \
+  --do_predict \
+  --fp16 --fp16_opt_level=O1 \
+  --val_check_interval 0.1 --n_val 1000 --eval_beams 2 --length_penalty=0.5 \
+  --max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 \
+  --model_name_or_path dbart_xsum_12_3 \
+  --train_batch_size=64 --eval_batch_size=64 \
+  --sortish_sampler \
+  --num_train_epochs=6 \
+  --warmup_steps 500 \
+  --output_dir distilbart_xsum_sft_12_3 --gpus 1
+```
+
+ Note: The command that produced `sshleifer/distilbart-cnn-12-6` is at [train_distilbart_cnn.sh](./[train_distilbart_cnn.sh)
+
+```bash
+./train_distilbart_cnn.sh
+```
+<!--- runtime: 6H on NVIDIA RTX 24GB GPU -->
+ Tip: You can get the same simple distillation logic by using `distillation.py --no_teacher ` followed by identical arguments as the ones in `train_distilbart_cnn.sh`.
+If you are using `wandb` and comparing the two distillation methods, using this entry point will make your logs consistent,
+because you will have the same hyper-parameters logged in every run.
+
+### Pseudo-Labeling
+ You don't need `distillation.py`.
+ Instructions to generate pseudo-labels and use pre-computed pseudo-labels can be found [here](./precomputed_pseudo_labels.md).
+Simply run `finetune.py` with one of those pseudo-label datasets as `--data_dir` (`DATA`, below).
+
+```bash
+python finetune.py \
+  --teacher facebook/bart-large-xsum --data_dir DATA \
+  --freeze_encoder --freeze_embeds \
+  --learning_rate=3e-4 \
+  --do_train \
+  --do_predict \
+  --fp16 --fp16_opt_level=O1 \
+  --val_check_interval 0.1 --n_val 1000 --eval_beams 2 --length_penalty=0.5 \
+  --max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 \
+  --model_name_or_path dbart_xsum_12_3 \
+  --train_batch_size=32 --eval_batch_size=32 \
+  --sortish_sampler \
+  --num_train_epochs=5 \
+  --warmup_steps 500 \
+  --output_dir dbart_xsum_12_3_PL --gpus 1 --logger_name wandb
+```
+
+ 
+
+To combine datasets, as in Section 6.2, try something like:
+```bash
+curl -S https://cdn-datasets.huggingface.co/pseudo/xsum/bart_xsum_pl.tgz | tar -xvz -C .
+curl -S https://cdn-datasets.huggingface.co/pseudo/xsum/pegasus_xsum.tgz | tar -xvz -C .
+curl -S https://cdn-datasets.huggingface.co/summarization/xsum.tar.gz | tar -xvz -C .
+mkdir all_pl
+cat bart_xsum_pl/train.source pegasus_xsum/train.source xsum/train.source > all_pl/train.source
+cat bart_xsum_pl/train.target pegasus_xsum/train.target xsum/train.target > all_pl/train.target
+cp xsum/val* all_pl
+cp xsum/test* all_pl
+```
+then use `all_pl` as DATA in the command above.
+
+#### Direct Knowledge Distillation (KD)
+ In this method, we use try to enforce that the student and teacher produce similar encoder_outputs, logits, and hidden_states using `BartSummarizationDistiller`.
+ This method was used for `sshleifer/distilbart-xsum-12-6`, `6-6`, and `9-6` checkpoints were produced.
+ You must use [`distillation.py`](./distillation.py). Note that this command initializes the student for you.
+
+The command that produced `sshleifer/distilbart-xsum-12-6` is at [./train_distilbart_xsum.sh](train_distilbart_xsum.sh)
+```bash
+./train_distilbart_xsum.sh --logger_name wandb --gpus 1
+```
+
+ Expected ROUGE-2 between 21.3 and 21.6, run time ~13H.
+ direct KD + Pegasus is VERY slow and works best with `--supervise_forward --normalize_hidden`.
+
+<!--- runtime: 13H on V-100 16GB GPU. -->
+
+### Citation
+
+```bibtex
+@misc{shleifer2020pretrained,
+      title={Pre-trained Summarization Distillation}, 
+      author={Sam Shleifer and Alexander M. Rush},
+      year={2020},
+      eprint={2010.13002},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+@article{Wolf2019HuggingFacesTS,
+  title={HuggingFace's 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},
+  journal={ArXiv},
+  year={2019},
+  volume={abs/1910.03771}
+}
+```
+
+This is the end of the distillation section, the rest of this doc pertains to general seq2seq commands.
+
+## Evaluation Commands

 To create summaries for each article in dataset, we use `run_eval.py`, here are a few commands that run eval for different tasks and models.
 If 'translation' is in your task name, the computed metric will be BLEU. Otherwise, ROUGE will be used.
@@ -272,9 +456,8 @@ export DATA_DIR=cnn_dm
    --score_path cnn_rouge.json \
    --task summarization \
    --n_obs 100 \
-    --device cuda \
-    --max_source_length 1024 \
-    --max_target_length 56 \
+
+th 56 \
    --fp16 \
    --bs 32
 ```
@@ -356,47 +539,6 @@ stas/wmt19-en-ru data/en-ru/val.source data/en-ru/test_translations.txt --refere
 If you pass `--info "some experiment-specific info"` it will get printed before the results table - this is useful for scripting and multiple runs, so one can tell the different sets of results from each other.


-### DistilBART
-![DBART](https://huggingface.co/front/thumbnails/distilbart_large.png)
-
-For the CNN/DailyMail dataset, (relatively longer, more extractive summaries), we found a simple technique that works:
-you just copy alternating layers from `bart-large-cnn` and finetune more on the same data.
-
-For the XSUM dataset, that didn’t work as well so we used that same initialization strategy followed by a combination of Distillbert’s ce_loss and the hidden states MSE loss used in the tinybert paper.
-
-You can see the performance tradeoffs of model sizes [here](https://docs.google.com/spreadsheets/d/1EkhDMwVO02m8jCD1cG3RoFPLicpcL1GQHTQjfvDYgIM/edit#gid=0).
-and more granular timing results [here](https://docs.google.com/spreadsheets/d/1EkhDMwVO02m8jCD1cG3RoFPLicpcL1GQHTQjfvDYgIM/edit#gid=1753259047&range=B2:I23).
-
-#### No Teacher Distillation
-To run the simpler distilbart-cnn style distillation all you need is data, a GPU, and a properly initialized student.
-You don't even need `distillation.py`.
-
-Some [un-finetuned students](https://huggingface.co/models?search=sshleifer%2Fstudent) are available for replication purposes.
-They are initialized by copying layers from the associated `bart-large-{cnn|xsum}` teacher using `--init_strategy alternate`. (You can read about that in `initialization_utils.py`)
-The command that produced `sshleifer/distilbart-cnn-12-6` is
-```bash
-./train_distilbart_cnn.sh
-```
-runtime: 6H on NVIDIA RTX 24GB GPU
-
-*Note*: You can get the same simple distillation logic by using `./run_distiller.sh --no_teacher` followed by identical arguments as the ones in `train_distilbart_cnn.sh`.
-If you are using `wandb` and comparing the two distillation methods, using this entry point will make your logs consistent,
-because you will have the same hyperparameters logged in every run.
-
-#### With a teacher (Intermediate Supervision)
-*Note* only BART variants are supported
-
-In this method, we use try to enforce that the student and teacher produce similar encoder_outputs, logits, and hidden_states using `BartSummarizationDistiller`.
-This is how `sshleifer/distilbart-xsum*` checkpoints were produced.
-
-The command that produced `sshleifer/distilbart-xsum-12-6` is:
-
-```bash
-./train_distilbart_xsum.sh --logger_name wandb --gpus 1
-```
-
-runtime: 13H on V-100 16GB GPU.
-
 ### Contributing
 - follow the standard contributing guidelines and code of conduct.
 - add tests to `test_seq2seq_examples.py`
@@ -417,7 +559,7 @@ python convert_pl_checkpoint_to_hf.py PATH_TO_CKPT  randomly_initialized_hf_mode
 Then either `run_eval` or `run_distributed_eval` with `save_dir/best_tfmr` (see previous sections)


-## Experimental Features 
+# Experimental Features 
 These features are harder to use and not always useful.

 ###  Dynamic Batch Size for MT
@@ -444,3 +586,5 @@ uses 12,723 batches of length 48 and takes slightly more time 9.5 minutes.
 The feature is still experimental, because:
 + we can make it much more robust if we have memory mapped/preprocessed datasets.
 + The speedup over sortish sampler is not that large at the moment.
+
+
@@ -23,7 +23,7 @@ logger = logging.getLogger(__name__)


 BERTABS_FINETUNED_CONFIG_MAP = {
-    "bertabs-finetuned-cnndm": "https://s3.amazonaws.com/models.huggingface.co/bert/remi/bertabs-finetuned-cnndm-extractive-abstractive-summarization/config.json",
+    "bertabs-finetuned-cnndm": "https://huggingface.co/remi/bertabs-finetuned-cnndm-extractive-abstractive-summarization/resolve/main/config.json",
 }


@@ -44,7 +44,7 @@ class BertAbsConfig(PretrainedConfig):
        enc_ff_size: int
            The size of the encoder's feed-forward layers.
        enc_dropout: int
-            The dropout probabilitiy for all fully connected layers in the
+            The dropout probability for all fully connected layers in the
            embeddings, layers, pooler and also the attention probabilities in
            the encoder.
        dec_layer: int
@@ -56,7 +56,7 @@ class BertAbsConfig(PretrainedConfig):
        dec_ff_size: int
            The size of the decoder's feed-forward layers.
        dec_dropout: int
-            The dropout probabilitiy for all fully connected layers in the
+            The dropout probability for all fully connected layers in the
            embeddings, layers, pooler and also the attention probabilities in
            the decoder.
    """
@@ -152,7 +152,7 @@ class TransformerDecoder(nn.Module):
       dropout (float): dropout parameters
       embeddings (:obj:`onmt.modules.Embeddings`):
          embeddings to use, should have positional encodings
-       attn_type (str): if using a seperate copy attention
+       attn_type (str): if using a separate copy attention
    """

    def __init__(self, num_layers, d_model, heads, d_ff, dropout, embeddings, vocab_size):
@@ -817,11 +817,7 @@ class Translator(object):

        Args:
           batch (:obj:`Batch`): a batch from a dataset object
-           data (:obj:`Dataset`): the dataset object
           fast (bool): enables fast beam search (may not support all features)
-
-        Todo:
-           Shouldn't need the original dataset.
        """
        with torch.no_grad():
            return self._fast_translate_batch(batch, self.max_length, min_length=self.min_length)
@@ -1,19 +1,18 @@
-export WANDB_PROJECT=distilbart-cnn
+export WANDB_PROJECT=distilbart-trainer
 export BS=32
-export GAS=1
 export m=sshleifer/student_cnn_12_6
 export tok=facebook/bart-large
 export MAX_TGT_LEN=142

 python finetune_trainer.py \
    --model_name_or_path $m --tokenizer_name $tok \ 
-    --data_dir $CNN_DIR \
+    --data_dir cnn_dm \
    --output_dir distilbart-cnn-12-6 --overwrite_output_dir \
    --learning_rate=3e-5 \
    --warmup_steps 500 --sortish_sampler \
    --fp16 \
    --n_val 500 \
-    --gradient_accumulation_steps=$GAS \
+    --gradient_accumulation_steps=1 \
    --per_device_train_batch_size=$BS --per_device_eval_batch_size=$BS \
    --freeze_encoder --freeze_embeds \
    --num_train_epochs=2 \
@@ -21,5 +20,5 @@ python finetune_trainer.py \
    --logging_first_step \
    --max_target_length 56 --val_max_target_length $MAX_TGT_LEN --test_max_target_length $MAX_TGT_LEN \
    --do_train --do_eval --do_predict --evaluate_during_training \
-    --predict_with_generate \
+    --predict_with_generate --sortish_sampler \
    "$@"
@@ -102,7 +102,6 @@ def get_checkpoint_callback(output_dir, metric, save_top_k=1, lower_is_better=Fa
        monitor=f"val_{metric}",
        mode="min" if "loss" in metric else "max",
        save_top_k=save_top_k,
-        period=0,  # maybe save a checkpoint every time val is run, not just end of epoch.
    )
    return checkpoint_callback

@@ -12,7 +12,7 @@ def convert(src_path: str, map_location: str = "cpu", save_path: Union[str, None
    state_dict = torch.load(src_path, map_location=map_location)
    for k, v in tqdm(state_dict.items()):
        if not isinstance(v, torch.Tensor):
-            raise TypeError("FP16 conversion only works on paths that are saved state dics, like pytorch_model.bin")
+            raise TypeError("FP16 conversion only works on paths that are saved state dicts, like pytorch_model.bin")
        state_dict[k] = v.half()
    if save_path is None:  # overwrite src_path
        save_path = src_path
@@ -182,7 +182,6 @@ class SummarizationModule(BaseTransformer):
        return self._generative_step(batch)

    def validation_epoch_end(self, outputs, prefix="val") -> Dict:
-
        self.step_count += 1
        losses = {k: torch.stack([x[k] for x in outputs]).mean() for k in self.loss_names}
        loss = losses["loss"]
@@ -252,7 +251,7 @@ class SummarizationModule(BaseTransformer):
    def get_dataloader(self, type_path: str, batch_size: int, shuffle: bool = False) -> DataLoader:
        dataset = self.get_dataset(type_path)

-        if self.hparams.sortish_sampler and type_path != "test":
+        if self.hparams.sortish_sampler and type_path != "test" and type_path != "val":
            sampler = dataset.make_sortish_sampler(batch_size, distributed=self.hparams.gpus > 1)
            return DataLoader(
                dataset,
@@ -263,7 +262,7 @@ class SummarizationModule(BaseTransformer):
                sampler=sampler,
            )

-        elif self.hparams.max_tokens_per_batch is not None and type_path != "test":
+        elif self.hparams.max_tokens_per_batch is not None and type_path != "test" and type_path != "val":
            batch_sampler = dataset.make_dynamic_sampler(
                self.hparams.max_tokens_per_batch, distributed=self.hparams.gpus > 1
            )
@@ -4,16 +4,9 @@ import sys
 from dataclasses import dataclass, field
 from typing import Optional

-from seq2seq_trainer import Seq2SeqTrainer, arg_to_scheduler_choices
-from transformers import (
-    AutoConfig,
-    AutoModelForSeq2SeqLM,
-    AutoTokenizer,
-    HfArgumentParser,
-    MBartTokenizer,
-    TrainingArguments,
-    set_seed,
-)
+from seq2seq_trainer import Seq2SeqTrainer
+from seq2seq_training_args import Seq2SeqTrainingArguments
+from transformers import AutoConfig, AutoModelForSeq2SeqLM, AutoTokenizer, HfArgumentParser, MBartTokenizer, set_seed
 from transformers.trainer_utils import EvaluationStrategy
 from utils import (
    Seq2SeqDataCollator,
@@ -33,41 +26,6 @@ from utils import (
 logger = logging.getLogger(__name__)


-@dataclass
-class Seq2SeqTrainingArguments(TrainingArguments):
-    """
-    Parameters:
-        label_smoothing (:obj:`float`, `optional`, defaults to 0):
-            The label smoothing epsilon to apply (if not zero).
-        sortish_sampler (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether to SortishSamler or not. It sorts the inputs according to lenghts in-order to minimizing the padding size.
-        predict_with_generate (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether to use generate to calculate generative metrics (ROUGE, BLEU).
-    """
-
-    label_smoothing: Optional[float] = field(
-        default=0.0, metadata={"help": "The label smoothing epsilon to apply (if not zero)."}
-    )
-    sortish_sampler: bool = field(default=False, metadata={"help": "Whether to SortishSamler or not."})
-    predict_with_generate: bool = field(
-        default=False, metadata={"help": "Whether to use generate to calculate generative metrics (ROUGE, BLEU)."}
-    )
-    adafactor: bool = field(default=False, metadata={"help": "whether to use adafactor"})
-    encoder_layerdrop: Optional[float] = field(
-        default=None, metadata={"help": "Encoder layer dropout probability. Goes into model.config."}
-    )
-    decoder_layerdrop: Optional[float] = field(
-        default=None, metadata={"help": "Decoder layer dropout probability. Goes into model.config."}
-    )
-    dropout: Optional[float] = field(default=None, metadata={"help": "Dropout probability. Goes into model.config."})
-    attention_dropout: Optional[float] = field(
-        default=None, metadata={"help": "Attention dropout probability. Goes into model.config."}
-    )
-    lr_scheduler: Optional[str] = field(
-        default="linear", metadata={"help": f"Which lr scheduler to use. Selected in {arg_to_scheduler_choices}"}
-    )
-
-
@dataclass
 class ModelArguments:
    """
@@ -33,7 +33,7 @@ python -m torch.distributed.launch --nproc_per_node=8 run_distributed_eval.py \
    --type_path train
 ```

-+ These command takes a while to run. For example,  pegasus_cnn_cnn_pls.tgz took 8 hours on 8 GPUs.
+ These commands takes a while to run. For example, `pegasus_cnn_cnn_pls.tgz` took 8 hours on 8 GPUs.
 + Pegasus does not work in fp16 :(, Bart, mBART and Marian do.
 + Even if you have 1 GPU, `run_distributed_eval.py` is 10-20% faster than `run_eval.py` because it uses `SortishSampler` to minimize padding computation.

@@ -1,10 +0,0 @@
-# Add parent directory to python path to access lightning_base.py
-export PYTHONPATH="../":"${PYTHONPATH}"
-
-python distillation.py \
--learning_rate=3e-4 \
--do_train \
--do_predict \
--fp16 \
--val_check_interval 0.1 \
-"$@"
@@ -20,12 +20,6 @@ from transformers.optimization import (
 from transformers.trainer_pt_utils import get_tpu_sampler


-try:
-    from .utils import label_smoothed_nll_loss
-except ImportError:
-    from utils import label_smoothed_nll_loss
-
-
 logger = logging.get_logger(__name__)

 arg_to_scheduler = {
@@ -36,7 +30,6 @@ arg_to_scheduler = {
    "constant": get_constant_schedule,
    "constant_w_warmup": get_constant_schedule_with_warmup,
 }
-arg_to_scheduler_choices = sorted(arg_to_scheduler.keys())


 class Seq2SeqTrainer(Trainer):
@@ -64,6 +57,14 @@ class Seq2SeqTrainer(Trainer):
                f"The `config.pad_token_id` is `None`. Using `config.eos_token_id` = {self.config.eos_token_id} for padding.."
            )

+        if self.args.label_smoothing == 0:
+            self.loss_fn = torch.nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)
+        else:
+            # dynamically import label_smoothed_nll_loss
+            from utils import label_smoothed_nll_loss
+
+            self.loss_fn = label_smoothed_nll_loss
+
    def create_optimizer_and_scheduler(self, num_training_steps: int):
        """
        Setup the optimizer and the learning rate scheduler.
@@ -135,9 +136,7 @@ class Seq2SeqTrainer(Trainer):
            if self.data_args is not None and self.data_args.ignore_pad_token_for_loss:
                # force training to ignore pad token
                logits = model(**inputs, use_cache=False)[0]
-
-                loss_fct = torch.nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)
-                loss = loss_fct(logits.view(-1, logits.shape[-1]), labels.view(-1))
+                loss = self.loss_fn(logits.view(-1, logits.shape[-1]), labels.view(-1))
            else:
                # compute usual loss via models
                loss, logits = model(**inputs, labels=labels, use_cache=False)[:2]
@@ -145,9 +144,7 @@ class Seq2SeqTrainer(Trainer):
            # compute label smoothed loss
            logits = model(**inputs, use_cache=False)[0]
            lprobs = torch.nn.functional.log_softmax(logits, dim=-1)
-            loss, _ = label_smoothed_nll_loss(
-                lprobs, labels, self.args.label_smoothing, ignore_index=self.config.pad_token_id
-            )
+            loss, _ = self.loss_fn(lprobs, labels, self.args.label_smoothing, ignore_index=self.config.pad_token_id)
        return loss, logits

    def compute_loss(self, model, inputs):
@@ -0,0 +1,45 @@
+import logging
+from dataclasses import dataclass, field
+from typing import Optional
+
+from seq2seq_trainer import arg_to_scheduler
+from transformers import TrainingArguments
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class Seq2SeqTrainingArguments(TrainingArguments):
+    """
+    Parameters:
+        label_smoothing (:obj:`float`, `optional`, defaults to 0):
+            The label smoothing epsilon to apply (if not zero).
+        sortish_sampler (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to SortishSamler or not. It sorts the inputs according to lenghts in-order to minimizing the padding size.
+        predict_with_generate (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to use generate to calculate generative metrics (ROUGE, BLEU).
+    """
+
+    label_smoothing: Optional[float] = field(
+        default=0.0, metadata={"help": "The label smoothing epsilon to apply (if not zero)."}
+    )
+    sortish_sampler: bool = field(default=False, metadata={"help": "Whether to SortishSamler or not."})
+    predict_with_generate: bool = field(
+        default=False, metadata={"help": "Whether to use generate to calculate generative metrics (ROUGE, BLEU)."}
+    )
+    adafactor: bool = field(default=False, metadata={"help": "whether to use adafactor"})
+    encoder_layerdrop: Optional[float] = field(
+        default=None, metadata={"help": "Encoder layer dropout probability. Goes into model.config."}
+    )
+    decoder_layerdrop: Optional[float] = field(
+        default=None, metadata={"help": "Decoder layer dropout probability. Goes into model.config."}
+    )
+    dropout: Optional[float] = field(default=None, metadata={"help": "Dropout probability. Goes into model.config."})
+    attention_dropout: Optional[float] = field(
+        default=None, metadata={"help": "Attention dropout probability. Goes into model.config."}
+    )
+    lr_scheduler: Optional[str] = field(
+        default="linear",
+        metadata={"help": f"Which lr scheduler to use. Selected in {sorted(arg_to_scheduler.keys())}"},
+    )
@@ -3,93 +3,110 @@
 import argparse
 import os
 import sys
-from pathlib import Path
 from unittest.mock import patch

-import pytest
 import pytorch_lightning as pl
 import timeout_decorator
 import torch

 from distillation import BartSummarizationDistiller, distill_main
 from finetune import SummarizationModule, main
-from test_seq2seq_examples import CUDA_AVAILABLE, MBART_TINY
-from transformers import BartForConditionalGeneration, MarianMTModel
-from transformers.testing_utils import TestCasePlus, slow
+from transformers import MarianMTModel
+from transformers.file_utils import cached_path
+from transformers.testing_utils import TestCasePlus, require_torch_gpu, require_torch_non_multigpu_but_fix_me, slow
 from utils import load_json


-MODEL_NAME = MBART_TINY
-MARIAN_MODEL = "sshleifer/student_marian_en_ro_6_1"
+MARIAN_MODEL = "sshleifer/mar_enro_6_3_student"


-class TestAll(TestCasePlus):
+class TestMbartCc25Enro(TestCasePlus):
+    def setUp(self):
+        super().setUp()
+
+        data_cached = cached_path(
+            "https://cdn-datasets.huggingface.co/translation/wmt_en_ro-tr40k-va0.5k-te0.5k.tar.gz",
+            extract_compressed_file=True,
+        )
+        self.data_dir = f"{data_cached}/wmt_en_ro-tr40k-va0.5k-te0.5k"
+
    @slow
-    @pytest.mark.skipif(not CUDA_AVAILABLE, reason="too slow to run on CPU")
+    @require_torch_gpu
+    @require_torch_non_multigpu_but_fix_me
    def test_model_download(self):
        """This warms up the cache so that we can time the next test without including download time, which varies between machines."""
-        BartForConditionalGeneration.from_pretrained(MODEL_NAME)
        MarianMTModel.from_pretrained(MARIAN_MODEL)

-    @timeout_decorator.timeout(120)
+    # @timeout_decorator.timeout(1200)
    @slow
-    @pytest.mark.skipif(not CUDA_AVAILABLE, reason="too slow to run on CPU")
+    @require_torch_gpu
+    @require_torch_non_multigpu_but_fix_me
    def test_train_mbart_cc25_enro_script(self):
-        data_dir = "examples/seq2seq/test_data/wmt_en_ro"
        env_vars_to_replace = {
-            "--fp16_opt_level=O1": "",
-            "$MAX_LEN": 128,
-            "$BS": 4,
+            "$MAX_LEN": 64,
+            "$BS": 64,
            "$GAS": 1,
-            "$ENRO_DIR": data_dir,
-            "facebook/mbart-large-cc25": MODEL_NAME,
-            # Download is 120MB in previous test.
-            "val_check_interval=0.25": "val_check_interval=1.0",
+            "$ENRO_DIR": self.data_dir,
+            "facebook/mbart-large-cc25": MARIAN_MODEL,
+            # "val_check_interval=0.25": "val_check_interval=1.0",
+            "--learning_rate=3e-5": "--learning_rate 3e-4",
+            "--num_train_epochs 6": "--num_train_epochs 1",
        }

        # Clean up bash script
-        bash_script = Path("examples/seq2seq/train_mbart_cc25_enro.sh").open().read().split("finetune.py")[1].strip()
+        bash_script = (self.test_file_dir / "train_mbart_cc25_enro.sh").open().read().split("finetune.py")[1].strip()
        bash_script = bash_script.replace("\\\n", "").strip().replace('"$@"', "")
        for k, v in env_vars_to_replace.items():
            bash_script = bash_script.replace(k, str(v))
        output_dir = self.get_auto_remove_tmp_dir()

-        bash_script = bash_script.replace("--fp16 ", "")
-        testargs = (
-            ["finetune.py"]
-            + bash_script.split()
-            + [
-                f"--output_dir={output_dir}",
-                "--gpus=1",
-                "--learning_rate=3e-1",
-                "--warmup_steps=0",
-                "--val_check_interval=1.0",
-                "--tokenizer_name=facebook/mbart-large-en-ro",
-            ]
-        )
+        # bash_script = bash_script.replace("--fp16 ", "")
+        args = f"""
+            --output_dir {output_dir}
+            --tokenizer_name Helsinki-NLP/opus-mt-en-ro
+            --sortish_sampler
+            --do_predict
+            --gpus 1
+            --freeze_encoder
+            --n_train 40000
+            --n_val 500
+            --n_test 500
+            --fp16_opt_level O1
+            --num_sanity_val_steps 0
+            --eval_beams 2
+        """.split()
+        # XXX: args.gpus > 1 : handle multigpu in the future
+
+        testargs = ["finetune.py"] + bash_script.split() + args
        with patch.object(sys, "argv", testargs):
            parser = argparse.ArgumentParser()
            parser = pl.Trainer.add_argparse_args(parser)
            parser = SummarizationModule.add_model_specific_args(parser, os.getcwd())
            args = parser.parse_args()
-            args.do_predict = False
-            # assert args.gpus == gpus THIS BREAKS for multigpu
            model = main(args)

        # Check metrics
        metrics = load_json(model.metrics_save_path)
        first_step_stats = metrics["val"][0]
        last_step_stats = metrics["val"][-1]
-        assert (
-            len(metrics["val"]) == (args.max_epochs / args.val_check_interval) + 1
-        )  # +1 accounts for val_sanity_check
-
-        assert last_step_stats["val_avg_gen_time"] >= 0.01
-
-        assert first_step_stats["val_avg_bleu"] < last_step_stats["val_avg_bleu"]  # model learned nothing
-        assert 1.0 >= last_step_stats["val_avg_gen_time"]  # model hanging on generate. Maybe bad config was saved.
+        self.assertEqual(len(metrics["val"]), (args.max_epochs / args.val_check_interval))
        assert isinstance(last_step_stats[f"val_avg_{model.val_metric}"], float)

+        self.assertGreater(last_step_stats["val_avg_gen_time"], 0.01)
+        # model hanging on generate. Maybe bad config was saved. (XXX: old comment/assert?)
+        self.assertLessEqual(last_step_stats["val_avg_gen_time"], 1.0)
+
+        # test learning requirements:
+
+        # 1. BLEU improves over the course of training by more than 2 pts
+        self.assertGreater(last_step_stats["val_avg_bleu"] - first_step_stats["val_avg_bleu"], 2)
+
+        # 2. BLEU finishes above 17
+        self.assertGreater(last_step_stats["val_avg_bleu"], 17)
+
+        # 3. test BLEU and val BLEU within ~1.1 pt.
+        self.assertLess(abs(metrics["val"][-1]["val_avg_bleu"] - metrics["test"][-1]["test_avg_bleu"]), 1.1)
+
        # check lightning ckpt can be loaded and has a reasonable statedict
        contents = os.listdir(output_dir)
        ckpt_path = [x for x in contents if x.endswith(".ckpt")][0]
@@ -107,11 +124,14 @@ class TestAll(TestCasePlus):
            # assert len(metrics["val"]) ==  desired_n_evals
            assert len(metrics["test"]) == 1

+
+class TestDistilMarianNoTeacher(TestCasePlus):
    @timeout_decorator.timeout(600)
    @slow
-    @pytest.mark.skipif(not CUDA_AVAILABLE, reason="too slow to run on CPU")
+    @require_torch_gpu
+    @require_torch_non_multigpu_but_fix_me
    def test_opus_mt_distill_script(self):
-        data_dir = "examples/seq2seq/test_data/wmt_en_ro"
+        data_dir = f"{self.test_file_dir_str}/test_data/wmt_en_ro"
        env_vars_to_replace = {
            "--fp16_opt_level=O1": "",
            "$MAX_LEN": 128,
@@ -124,7 +144,7 @@ class TestAll(TestCasePlus):

        # Clean up bash script
        bash_script = (
-            Path("examples/seq2seq/distil_marian_no_teacher.sh").open().read().split("distillation.py")[1].strip()
+            (self.test_file_dir / "distil_marian_no_teacher.sh").open().read().split("distillation.py")[1].strip()
        )
        bash_script = bash_script.replace("\\\n", "").strip().replace('"$@"', "")
        bash_script = bash_script.replace("--fp16 ", " ")
@@ -144,6 +164,7 @@ class TestAll(TestCasePlus):
                f"--num_train_epochs={epochs}",
                "--warmup_steps=10",
                "--val_check_interval=1.0",
+                "--do_predict",
            ]
        )
        with patch.object(sys, "argv", testargs):
@@ -151,7 +172,6 @@ class TestAll(TestCasePlus):
            parser = pl.Trainer.add_argparse_args(parser)
            parser = BartSummarizationDistiller.add_model_specific_args(parser, os.getcwd())
            args = parser.parse_args()
-            args.do_predict = False
            # assert args.gpus == gpus THIS BREAKS for multigpu

            model = distill_main(args)
@@ -11,7 +11,7 @@ from save_len_file import save_len_file
 from test_seq2seq_examples import ARTICLES, BART_TINY, MARIAN_TINY, MBART_TINY, SUMMARIES, T5_TINY, make_test_data_dir
 from transformers import AutoTokenizer
 from transformers.modeling_bart import shift_tokens_right
-from transformers.testing_utils import TestCasePlus, slow
+from transformers.testing_utils import TestCasePlus, require_torch_non_multigpu_but_fix_me, slow
 from utils import FAIRSEQ_AVAILABLE, DistributedSortishSampler, LegacySeq2SeqDataset, Seq2SeqDataset


@@ -30,6 +30,7 @@ class TestAll(TestCasePlus):
        ],
    )
    @slow
+    @require_torch_non_multigpu_but_fix_me
    def test_seq2seq_dataset_truncation(self, tok_name):
        tokenizer = AutoTokenizer.from_pretrained(tok_name)
        tmp_dir = make_test_data_dir(tmp_dir=self.get_auto_remove_tmp_dir())
@@ -69,6 +70,7 @@ class TestAll(TestCasePlus):
            break  # No need to test every batch

    @parameterized.expand([BART_TINY, BERT_BASE_CASED])
+    @require_torch_non_multigpu_but_fix_me
    def test_legacy_dataset_truncation(self, tok):
        tokenizer = AutoTokenizer.from_pretrained(tok)
        tmp_dir = make_test_data_dir(tmp_dir=self.get_auto_remove_tmp_dir())
@@ -93,6 +95,7 @@ class TestAll(TestCasePlus):
            assert max_len_target > trunc_target  # Truncated
            break  # No need to test every batch

+    @require_torch_non_multigpu_but_fix_me
    def test_pack_dataset(self):
        tokenizer = AutoTokenizer.from_pretrained("facebook/mbart-large-cc25")

@@ -111,6 +114,7 @@ class TestAll(TestCasePlus):
        assert orig_paths == new_paths

    @pytest.mark.skipif(not FAIRSEQ_AVAILABLE, reason="This test requires fairseq")
+    @require_torch_non_multigpu_but_fix_me
    def test_dynamic_batch_size(self):
        if not FAIRSEQ_AVAILABLE:
            return
@@ -135,6 +139,7 @@ class TestAll(TestCasePlus):
        if failures:
            raise AssertionError(f"too many tokens in {len(failures)} batches")

+    @require_torch_non_multigpu_but_fix_me
    def test_sortish_sampler_reduces_padding(self):
        ds, _, tokenizer = self._get_dataset(max_len=512)
        bs = 2
@@ -174,6 +179,7 @@ class TestAll(TestCasePlus):
        )
        return ds, max_tokens, tokenizer

+    @require_torch_non_multigpu_but_fix_me
    def test_distributed_sortish_sampler_splits_indices_between_procs(self):
        ds, max_tokens, tokenizer = self._get_dataset()
        ids1 = set(DistributedSortishSampler(ds, 256, num_replicas=2, rank=0, add_extra_examples=False))
@@ -189,6 +195,7 @@ class TestAll(TestCasePlus):
            PEGASUS_XSUM,
        ],
    )
+    @require_torch_non_multigpu_but_fix_me
    def test_dataset_kwargs(self, tok_name):
        tokenizer = AutoTokenizer.from_pretrained(tok_name)
        if tok_name == MBART_TINY:
@@ -1,25 +1,18 @@
 import os
 import sys
-from pathlib import Path
 from unittest.mock import patch

-import pytest
-
-from transformers import BertTokenizer, EncoderDecoderModel, is_torch_available
+from transformers import BertTokenizer, EncoderDecoderModel
 from transformers.file_utils import is_datasets_available
-from transformers.testing_utils import TestCasePlus, slow
+from transformers.testing_utils import TestCasePlus, execute_subprocess_async, get_gpu_count, slow
 from transformers.trainer_callback import TrainerState
 from transformers.trainer_utils import set_seed

 from .finetune_trainer import Seq2SeqTrainingArguments, main
 from .seq2seq_trainer import Seq2SeqTrainer
 from .test_seq2seq_examples import MBART_TINY
-from .utils import execute_async_std


-if is_torch_available():
-    import torch
-
 set_seed(42)
 MARIAN_MODEL = "sshleifer/student_marian_en_ro_6_1"

@@ -166,11 +159,9 @@ class TestFinetuneTrainer(TestCasePlus):
        trainer.train()

    def run_trainer(self, eval_steps: int, max_len: str, model_name: str, num_train_epochs: int):
-
-        # XXX: remove hardcoded path
-        data_dir = "examples/seq2seq/test_data/wmt_en_ro"
+        data_dir = self.examples_dir / "seq2seq/test_data/wmt_en_ro"
        output_dir = self.get_auto_remove_tmp_dir()
-        argv = f"""
+        args = f"""
            --model_name_or_path {model_name}
            --data_dir {data_dir}
            --output_dir {output_dir}
@@ -186,7 +177,7 @@ class TestFinetuneTrainer(TestCasePlus):
            --num_train_epochs {str(num_train_epochs)}
            --per_device_train_batch_size 4
            --per_device_eval_batch_size 4
-            --learning_rate 3e-4
+            --learning_rate 3e-3
            --warmup_steps 8
            --evaluate_during_training
            --predict_with_generate
@@ -202,33 +193,18 @@ class TestFinetuneTrainer(TestCasePlus):
        """.split()
        # --eval_beams  2

-        n_gpu = torch.cuda.device_count()
+        n_gpu = get_gpu_count()
        if n_gpu > 1:
-
-            path = Path(__file__).resolve()
-            cur_path = path.parents[0]
-
-            path = Path(__file__).resolve()
-            examples_path = path.parents[1]
-            src_path = f"{path.parents[2]}/src"
-            env = os.environ.copy()
-            env["PYTHONPATH"] = f"{examples_path}:{src_path}:{env.get('PYTHONPATH', '')}"
-
-            distributed_args = (
-                f"-m torch.distributed.launch --nproc_per_node={n_gpu} {cur_path}/finetune_trainer.py".split()
-            )
-            cmd = [sys.executable] + distributed_args + argv
-
-            print("\nRunning: ", " ".join(cmd))
-
-            result = execute_async_std(cmd, env=env, stdin=None, timeout=180, quiet=False, echo=False)
-
-            assert result.stdout, "produced no output"
-            if result.returncode > 0:
-                pytest.fail(f"failed with returncode {result.returncode}")
+            distributed_args = f"""
+                -m torch.distributed.launch
+                --nproc_per_node={n_gpu}
+                {self.test_file_dir}/finetune_trainer.py
+            """.split()
+            cmd = [sys.executable] + distributed_args + args
+            execute_subprocess_async(cmd, env=self.get_env())
        else:
            # 0 or 1 gpu
-            testargs = ["finetune_trainer.py"] + argv
+            testargs = ["finetune_trainer.py"] + args
            with patch.object(sys, "argv", testargs):
                main()

@@ -19,7 +19,13 @@ import unittest

 from parameterized import parameterized
 from transformers import FSMTForConditionalGeneration, FSMTTokenizer
-from transformers.testing_utils import get_tests_dir, require_torch, slow, torch_device
+from transformers.testing_utils import (
+    get_tests_dir,
+    require_torch,
+    require_torch_non_multigpu_but_fix_me,
+    slow,
+    torch_device,
+)
 from utils import calculate_bleu


@@ -48,6 +54,7 @@ class ModelEvalTester(unittest.TestCase):
        ]
    )
    @slow
+    @require_torch_non_multigpu_but_fix_me
    def test_bleu_scores(self, pair, min_bleu_score):
        # note: this test is not testing the best performance since it only evals a small batch
        # but it should be enough to detect a regression in the output quality
@@ -4,7 +4,7 @@ import unittest
 from make_student import create_student_by_copying_alternating_layers
 from transformers import AutoConfig
 from transformers.file_utils import cached_property
-from transformers.testing_utils import require_torch
+from transformers.testing_utils import require_torch, require_torch_non_multigpu_but_fix_me


 TINY_BART = "sshleifer/bart-tiny-random"
@@ -17,23 +17,28 @@ class MakeStudentTester(unittest.TestCase):
    def teacher_config(self):
        return AutoConfig.from_pretrained(TINY_BART)

+    @require_torch_non_multigpu_but_fix_me
    def test_valid_t5(self):
        student, *_ = create_student_by_copying_alternating_layers(TINY_T5, tempfile.mkdtemp(), e=1, d=1)
        self.assertEqual(student.config.num_hidden_layers, 1)

+    @require_torch_non_multigpu_but_fix_me
    def test_asymmetric_t5(self):
        student, *_ = create_student_by_copying_alternating_layers(TINY_T5, tempfile.mkdtemp(), e=1, d=None)

+    @require_torch_non_multigpu_but_fix_me
    def test_same_decoder_small_encoder(self):
        student, *_ = create_student_by_copying_alternating_layers(TINY_BART, tempfile.mkdtemp(), e=1, d=None)
        self.assertEqual(student.config.encoder_layers, 1)
        self.assertEqual(student.config.decoder_layers, self.teacher_config.encoder_layers)

+    @require_torch_non_multigpu_but_fix_me
    def test_small_enc_small_dec(self):
        student, *_ = create_student_by_copying_alternating_layers(TINY_BART, tempfile.mkdtemp(), e=1, d=1)
        self.assertEqual(student.config.encoder_layers, 1)
        self.assertEqual(student.config.decoder_layers, 1)

+    @require_torch_non_multigpu_but_fix_me
    def test_raises_assert(self):
        with self.assertRaises(AssertionError):
            create_student_by_copying_alternating_layers(TINY_BART, tempfile.mkdtemp(), e=None, d=None)
@@ -19,7 +19,14 @@ from run_eval import generate_summaries_or_translations, run_generate
 from run_eval_search import run_search
 from transformers import AutoConfig, AutoModelForSeq2SeqLM
 from transformers.hf_api import HfApi
-from transformers.testing_utils import CaptureStderr, CaptureStdout, TestCasePlus, require_torch_gpu, slow
+from transformers.testing_utils import (
+    CaptureStderr,
+    CaptureStdout,
+    TestCasePlus,
+    require_torch_gpu,
+    require_torch_non_multigpu_but_fix_me,
+    slow,
+)
 from utils import ROUGE_KEYS, label_smoothed_nll_loss, lmap, load_json


@@ -126,6 +133,7 @@ class TestSummarizationDistiller(TestCasePlus):

    @slow
    @require_torch_gpu
+    @require_torch_non_multigpu_but_fix_me
    def test_hub_configs(self):
        """I put require_torch_gpu cause I only want this to run with self-scheduled."""

@@ -143,10 +151,12 @@ class TestSummarizationDistiller(TestCasePlus):
                failures.append(m)
        assert not failures, f"The following models could not be loaded through AutoConfig: {failures}"

+    @require_torch_non_multigpu_but_fix_me
    def test_distill_no_teacher(self):
        updates = dict(student_encoder_layers=2, student_decoder_layers=1, no_teacher=True)
        self._test_distiller_cli(updates)

+    @require_torch_non_multigpu_but_fix_me
    def test_distill_checkpointing_with_teacher(self):
        updates = dict(
            student_encoder_layers=2,
@@ -171,6 +181,7 @@ class TestSummarizationDistiller(TestCasePlus):
        convert_pl_to_hf(ckpts[0], transformer_ckpts[0].parent, out_path_new)
        assert os.path.exists(os.path.join(out_path_new, "pytorch_model.bin"))

+    @require_torch_non_multigpu_but_fix_me
    def test_loss_fn(self):
        model = AutoModelForSeq2SeqLM.from_pretrained(BART_TINY, return_dict=True)
        input_ids, mask = model.dummy_inputs["input_ids"], model.dummy_inputs["attention_mask"]
@@ -191,6 +202,7 @@ class TestSummarizationDistiller(TestCasePlus):
            # TODO: understand why this breaks
            self.assertEqual(nll_loss, model_computed_loss)

+    @require_torch_non_multigpu_but_fix_me
    def test_distill_mbart(self):
        updates = dict(
            student_encoder_layers=2,
@@ -215,6 +227,7 @@ class TestSummarizationDistiller(TestCasePlus):
        assert len(all_files) > 2
        self.assertEqual(len(transformer_ckpts), 2)

+    @require_torch_non_multigpu_but_fix_me
    def test_distill_t5(self):
        updates = dict(
            student_encoder_layers=1,
@@ -296,18 +309,21 @@ class TestTheRest(TestCasePlus):

    # test one model to quickly (no-@slow) catch simple problems and do an
    # extensive testing of functionality with multiple models as @slow separately
+    @require_torch_non_multigpu_but_fix_me
    def test_run_eval(self):
        self.run_eval_tester(T5_TINY)

    # any extra models should go into the list here - can be slow
    @parameterized.expand([BART_TINY, MBART_TINY])
    @slow
+    @require_torch_non_multigpu_but_fix_me
    def test_run_eval_slow(self, model):
        self.run_eval_tester(model)

    # testing with 2 models to validate: 1. translation (t5) 2. summarization (mbart)
    @parameterized.expand([T5_TINY, MBART_TINY])
    @slow
+    @require_torch_non_multigpu_but_fix_me
    def test_run_eval_search(self, model):
        input_file_name = Path(self.get_auto_remove_tmp_dir()) / "utest_input.source"
        output_file_name = input_file_name.parent / "utest_output.txt"
@@ -358,6 +374,7 @@ class TestTheRest(TestCasePlus):
    @parameterized.expand(
        [T5_TINY, BART_TINY, MBART_TINY, MARIAN_TINY, FSMT_TINY],
    )
+    @require_torch_non_multigpu_but_fix_me
    def test_finetune(self, model):
        args_d: dict = CHEAP_ARGS.copy()
        task = "translation" if model in [MBART_TINY, MARIAN_TINY, FSMT_TINY] else "summarization"
@@ -409,6 +426,7 @@ class TestTheRest(TestCasePlus):
        assert isinstance(example_batch, dict)
        assert len(example_batch) >= 4

+    @require_torch_non_multigpu_but_fix_me
    def test_finetune_extra_model_args(self):
        args_d: dict = CHEAP_ARGS.copy()

@@ -459,6 +477,7 @@ class TestTheRest(TestCasePlus):
            model = main(args)
        assert str(excinfo.value) == f"model config doesn't have a `{unsupported_param}` attribute"

+    @require_torch_non_multigpu_but_fix_me
    def test_finetune_lr_schedulers(self):
        args_d: dict = CHEAP_ARGS.copy()

@@ -1,119 +1,24 @@
 # as due to their complexity multi-gpu tests could impact other tests, and to aid debug we have those in a separate module.

-import logging
 import os
 import sys
-from pathlib import Path

-import pytest
+from transformers.testing_utils import (
+    TestCasePlus,
+    execute_subprocess_async,
+    get_gpu_count,
+    require_torch_gpu,
+    require_torch_multigpu,
+    slow,
+)

-from transformers import is_torch_available
-from transformers.testing_utils import TestCasePlus, require_torch_multigpu
-
-from .utils import execute_async_std, load_json
-
-
-if is_torch_available():
-    import torch
-
-
-logging.basicConfig(level=logging.DEBUG)
-
-logger = logging.getLogger()
-CUDA_AVAILABLE = torch.cuda.is_available()
-CHEAP_ARGS = {
-    "max_tokens_per_batch": None,
-    "supervise_forward": True,
-    "normalize_hidden": True,
-    "label_smoothing": 0.2,
-    "eval_max_gen_length": None,
-    "eval_beams": 1,
-    "val_metric": "loss",
-    "save_top_k": 1,
-    "adafactor": True,
-    "early_stopping_patience": 2,
-    "logger_name": "default",
-    "length_penalty": 0.5,
-    "cache_dir": "",
-    "task": "summarization",
-    "num_workers": 2,
-    "alpha_hid": 0,
-    "freeze_embeds": True,
-    "enc_only": False,
-    "tgt_suffix": "",
-    "resume_from_checkpoint": None,
-    "sortish_sampler": True,
-    "student_decoder_layers": 1,
-    "val_check_interval": 1.0,
-    "output_dir": "",
-    "fp16": False,  # TODO(SS): set this to CUDA_AVAILABLE if ci installs apex or start using native amp
-    "no_teacher": False,
-    "fp16_opt_level": "O1",
-    "gpus": 1 if CUDA_AVAILABLE else 0,
-    "n_tpu_cores": 0,
-    "max_grad_norm": 1.0,
-    "do_train": True,
-    "do_predict": True,
-    "accumulate_grad_batches": 1,
-    "server_ip": "",
-    "server_port": "",
-    "seed": 42,
-    "model_name_or_path": "sshleifer/bart-tiny-random",
-    "config_name": "",
-    "tokenizer_name": "facebook/bart-large",
-    "do_lower_case": False,
-    "learning_rate": 0.3,
-    "lr_scheduler": "linear",
-    "weight_decay": 0.0,
-    "adam_epsilon": 1e-08,
-    "warmup_steps": 0,
-    "max_epochs": 1,
-    "train_batch_size": 2,
-    "eval_batch_size": 2,
-    "max_source_length": 12,
-    "max_target_length": 12,
-    "val_max_target_length": 12,
-    "test_max_target_length": 12,
-    "fast_dev_run": False,
-    "no_cache": False,
-    "n_train": -1,
-    "n_val": -1,
-    "n_test": -1,
-    "student_encoder_layers": 1,
-    "freeze_encoder": False,
-    "auto_scale_batch_size": False,
-}
-
-
-def _dump_articles(path: Path, articles: list):
-    content = "\n".join(articles)
-    Path(path).open("w").writelines(content)
-
-
-ARTICLES = [" Sam ate lunch today.", "Sams lunch ingredients."]
-SUMMARIES = ["A very interesting story about what I ate for lunch.", "Avocado, celery, turkey, coffee"]
-T5_TINY = "patrickvonplaten/t5-tiny-random"
-BART_TINY = "sshleifer/bart-tiny-random"
-MBART_TINY = "sshleifer/tiny-mbart"
-MARIAN_TINY = "sshleifer/tiny-marian-en-de"
-
-
-stream_handler = logging.StreamHandler(sys.stdout)
-logger.addHandler(stream_handler)
-logging.disable(logging.CRITICAL)  # remove noisy download output from tracebacks
-
-
-def make_test_data_dir(tmp_dir):
-    for split in ["train", "val", "test"]:
-        _dump_articles(os.path.join(tmp_dir, f"{split}.source"), ARTICLES)
-        _dump_articles(os.path.join(tmp_dir, f"{split}.target"), SUMMARIES)
-    return tmp_dir
+from .test_seq2seq_examples import CHEAP_ARGS, make_test_data_dir
+from .utils import load_json


 class TestSummarizationDistillerMultiGPU(TestCasePlus):
    @classmethod
    def setUpClass(cls):
-        logging.disable(logging.CRITICAL)  # remove noisy download output from tracebacks
        return cls

    @require_torch_multigpu
@@ -157,23 +62,9 @@ class TestSummarizationDistillerMultiGPU(TestCasePlus):
                return f"--{k}"
            return f"--{k}={v}"

-        path = Path(__file__).resolve()
-        cur_path = path.parents[0]
-        examples_path = path.parents[1]
-        src_path = f"{path.parents[2]}/src"
-        env = os.environ.copy()
-        env["PYTHONPATH"] = f"{examples_path}:{src_path}:{env.get('PYTHONPATH', '')}"
-
        cli_args = [x for x in (convert(k, v) for k, v in args_d.items()) if len(x)]
-        cmd = [sys.executable, f"{cur_path}/distillation.py"] + cli_args
-
-        print("\nRunning: ", " ".join(cmd))
-
-        result = execute_async_std(cmd, env=env, stdin=None, timeout=180, quiet=False, echo=False)
-
-        assert result.stdout, "produced no output"
-        if result.returncode > 0:
-            pytest.fail(f"failed with returncode {result.returncode}")
+        cmd = [sys.executable, f"{self.test_file_dir}/distillation.py"] + cli_args
+        execute_subprocess_async(cmd, env=self.get_env())

        contents = os.listdir(output_dir)
        contents = {os.path.basename(p) for p in contents}
@@ -192,8 +83,34 @@ class TestSummarizationDistillerMultiGPU(TestCasePlus):
        print(metrics)
        last_step_stats = metrics["val"][-1]
        self.assertGreaterEqual(last_step_stats["val_avg_gen_time"], 0.01)
-        self.assertGreaterEqual(1.0, last_step_stats["val_avg_gen_time"])
        self.assertIsInstance(last_step_stats[f"val_avg_{val_metric}"], float)
        self.assertEqual(len(metrics["test"]), 1)
        desired_n_evals = int(args_d["max_epochs"] * (1 / args_d["val_check_interval"]) / 2 + 1)
        self.assertEqual(len(metrics["val"]), desired_n_evals)
+
+    @slow
+    @require_torch_gpu
+    def test_distributed_eval(self):
+        output_dir = self.get_auto_remove_tmp_dir()
+        args = f"""
+            --model_name Helsinki-NLP/opus-mt-en-ro
+            --save_dir {output_dir}
+            --data_dir {self.test_file_dir_str}/test_data/wmt_en_ro
+            --num_beams 2
+            --task translation
+        """.split()
+
+        # we want this test to run even if there is only one GPU, but if there are more we use them all
+        n_gpu = get_gpu_count()
+        distributed_args = f"""
+            -m torch.distributed.launch
+            --nproc_per_node={n_gpu}
+            {self.test_file_dir}/run_distributed_eval.py
+        """.split()
+        cmd = [sys.executable] + distributed_args + args
+        execute_subprocess_async(cmd, env=self.get_env())
+
+        metrics_save_path = os.path.join(output_dir, "test_bleu.json")
+        metrics = load_json(metrics_save_path)
+        # print(metrics)
+        self.assertGreaterEqual(metrics["bleu"], 25)
@@ -1,11 +1,13 @@
+import os
 import tempfile
 import unittest

-from transformers.convert_marian_tatoeba_to_pytorch import TatoebaConverter
+from transformers.convert_marian_tatoeba_to_pytorch import DEFAULT_REPO, TatoebaConverter
 from transformers.file_utils import cached_property
-from transformers.testing_utils import slow
+from transformers.testing_utils import require_torch_non_multigpu_but_fix_me, slow


+@unittest.skipUnless(os.path.exists(DEFAULT_REPO), "Tatoeba directory does not exist.")
 class TatoebaConversionTester(unittest.TestCase):
    @cached_property
    def resolver(self):
@@ -13,10 +15,12 @@ class TatoebaConversionTester(unittest.TestCase):
        return TatoebaConverter(save_dir=tmp_dir)

    @slow
+    @require_torch_non_multigpu_but_fix_me
    def test_resolver(self):
        self.resolver.convert_models(["heb-eng"])

    @slow
+    @require_torch_non_multigpu_but_fix_me
    def test_model_card(self):
        content, mmeta = self.resolver.write_model_card("opus-mt-he-en", dry_run=True)
        assert mmeta["long_pair"] == "heb-eng"
@@ -13,7 +13,7 @@ python finetune.py \
    --val_check_interval 0.25 \
    --n_val 500 \
    --num_train_epochs 2 \
-    --freeze_encoder --freeze_embeds --data_dir $CNN_DIR \
+    --freeze_encoder --freeze_embeds --data_dir cnn_dm \
    --max_target_length 142 --val_max_target_length=142 \
    --train_batch_size=$BS --eval_batch_size=$BS --gradient_accumulation_steps=$GAS \
    --model_name_or_path sshleifer/student_cnn_12_6 \
@@ -2,6 +2,7 @@
 export PYTHONPATH="../":"${PYTHONPATH}"
 python distillation.py \
  --teacher facebook/bart-large-xsum --data_dir xsum \
+  --tokenizer_name facebook/bart-large-xsum \
  --student_decoder_layers 6 --student_encoder_layers 12 \
  --freeze_encoder --freeze_embeds \
  --learning_rate=3e-4 \
@@ -5,7 +5,6 @@ import math
 import os
 import pickle
 import socket
-import sys
 from logging import getLogger
 from pathlib import Path
 from typing import Callable, Dict, Iterable, List, Tuple, Union
@@ -270,7 +269,7 @@ class Seq2SeqDataCollator:
        ), f"pad_token_id is not defined for ({self.tokenizer.__class__.__name__}), it must be defined."
        self.data_args = data_args
        self.tpu_num_cores = tpu_num_cores
-        self.dataset_kwargs = {"add_prefix_space": isinstance(tokenizer, BartTokenizer)}
+        self.dataset_kwargs = {"add_prefix_space": True} if isinstance(tokenizer, BartTokenizer) else {}
        if data_args.src_lang is not None:
            self.dataset_kwargs["src_lang"] = data_args.src_lang
        if data_args.tgt_lang is not None:
@@ -641,74 +640,6 @@ def check_output_dir(args, expected_items=0):
    ):
        raise ValueError(
            f"Output directory ({args.output_dir}) already exists and "
-            "has {len(os.listdir(args.output_dir))} items in it (expected {expected_items} items). "
+            f"has {len(os.listdir(args.output_dir))} items in it (expected {expected_items} items). "
            "Use --overwrite_output_dir to overcome."
        )
-
-
-# the following code deals with async io between processes
-
-# adapted from https://stackoverflow.com/a/59041913/9201239
-import asyncio  # noqa
-
-
-class _RunOutput:
-    def __init__(self, returncode, stdout, stderr):
-        self.returncode = returncode
-        self.stdout = stdout
-        self.stderr = stderr
-
-
-async def _read_stream(stream, callback):
-    while True:
-        line = await stream.readline()
-        if line:
-            callback(line)
-        else:
-            break
-
-
-async def _stream_subprocess(cmd, env=None, stdin=None, timeout=None, quiet=False, echo=False) -> _RunOutput:
-    if echo:
-        print(cmd)
-
-    p = await asyncio.create_subprocess_exec(
-        cmd[0],
-        *cmd[1:],
-        stdin=stdin,
-        stdout=asyncio.subprocess.PIPE,
-        stderr=asyncio.subprocess.PIPE,
-        env=env,
-    )
-    out = []
-    err = []
-
-    def tee(line, sink, pipe, label=""):
-        line = line.decode("utf-8").rstrip()
-        sink.append(line)
-        if not quiet:
-            print(label, line, file=pipe)
-
-    await asyncio.wait(
-        [
-            _read_stream(p.stdout, lambda l: tee(l, out, sys.stdout)),
-            _read_stream(p.stderr, lambda l: tee(l, err, sys.stderr, label="stderr:")),
-        ],
-        timeout=timeout,
-    )
-
-    # XXX: warning for a possible deadlock when using `wait` with huge amounts of data in the pipe
-    # https://docs.python.org/3/library/asyncio-subprocess.html#asyncio.asyncio.subprocess.Process.wait
-    #
-    # If it starts hanging, will need to switch s/wait/communicate/ - so perhaps for debug we will enable
-    # `wait` as it's easier to see in real time, but for normal runs use `communicate`
-    return _RunOutput(await p.wait(), out, err)
-
-
-def execute_async_std(cmd, env=None, stdin=None, timeout=None, quiet=False, echo=False) -> _RunOutput:
-    loop = asyncio.get_event_loop()
-    result = loop.run_until_complete(
-        _stream_subprocess(cmd, env=env, stdin=stdin, timeout=timeout, quiet=quiet, echo=echo)
-    )
-
-    return result
@@ -23,20 +23,28 @@ from unittest.mock import patch
 import torch

 from transformers.file_utils import is_apex_available
-from transformers.testing_utils import TestCasePlus, torch_device
+from transformers.testing_utils import TestCasePlus, require_torch_non_multigpu_but_fix_me, torch_device


 SRC_DIRS = [
    os.path.join(os.path.dirname(__file__), dirname)
-    for dirname in ["text-generation", "text-classification", "language-modeling", "question-answering"]
+    for dirname in [
+        "text-generation",
+        "text-classification",
+        "token-classification",
+        "language-modeling",
+        "question-answering",
+    ]
 ]
 sys.path.extend(SRC_DIRS)


 if SRC_DIRS is not None:
+    import run_clm
    import run_generation
    import run_glue
-    import run_language_modeling
+    import run_mlm
+    import run_ner
    import run_pl_glue
    import run_squad

@@ -59,6 +67,7 @@ def is_cuda_and_apex_available():


 class ExamplesTests(TestCasePlus):
+    @require_torch_non_multigpu_but_fix_me
    def test_run_glue(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)
@@ -91,6 +100,7 @@ class ExamplesTests(TestCasePlus):
            for value in result.values():
                self.assertGreaterEqual(value, 0.75)

+    @require_torch_non_multigpu_but_fix_me
    def test_run_pl_glue(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)
@@ -128,33 +138,96 @@ class ExamplesTests(TestCasePlus):
            #         self.assertGreaterEqual(v, 0.75, f"({k})")
            #

-    def test_run_language_modeling(self):
+    @require_torch_non_multigpu_but_fix_me
+    def test_run_clm(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
-            run_language_modeling.py
-            --model_name_or_path distilroberta-base
-            --model_type roberta
-            --mlm
-            --line_by_line
-            --train_data_file ./tests/fixtures/sample_text.txt
-            --eval_data_file ./tests/fixtures/sample_text.txt
-            --output_dir {tmp_dir}
-            --overwrite_output_dir
+            run_clm.py
+            --model_name_or_path distilgpt2
+            --train_file ./tests/fixtures/sample_text.txt
+            --validation_file ./tests/fixtures/sample_text.txt
            --do_train
            --do_eval
-            --num_train_epochs=1
+            --block_size 128
+            --per_device_train_batch_size 5
+            --per_device_eval_batch_size 5
+            --num_train_epochs 2
+            --output_dir {tmp_dir}
+            --overwrite_output_dir
            """.split()

+        if torch.cuda.device_count() > 1:
+            # Skipping because there are not enough batches to train the model + would need a drop_last to work.
+            return
+
        if torch_device != "cuda":
            testargs.append("--no_cuda")

        with patch.object(sys, "argv", testargs):
-            result = run_language_modeling.main()
+            result = run_clm.main()
+            self.assertLess(result["perplexity"], 100)
+
+    @require_torch_non_multigpu_but_fix_me
+    def test_run_mlm(self):
+        stream_handler = logging.StreamHandler(sys.stdout)
+        logger.addHandler(stream_handler)
+
+        tmp_dir = self.get_auto_remove_tmp_dir()
+        testargs = f"""
+            run_mlm.py
+            --model_name_or_path distilroberta-base
+            --train_file ./tests/fixtures/sample_text.txt
+            --validation_file ./tests/fixtures/sample_text.txt
+            --output_dir {tmp_dir}
+            --overwrite_output_dir
+            --do_train
+            --do_eval
+            --prediction_loss_only
+            --num_train_epochs=1
+        """.split()
+
+        if torch_device != "cuda":
+            testargs.append("--no_cuda")
+
+        with patch.object(sys, "argv", testargs):
+            result = run_mlm.main()
            self.assertLess(result["perplexity"], 42)

+    @require_torch_non_multigpu_but_fix_me
+    def test_run_ner(self):
+        stream_handler = logging.StreamHandler(sys.stdout)
+        logger.addHandler(stream_handler)
+
+        tmp_dir = self.get_auto_remove_tmp_dir()
+        testargs = f"""
+            run_ner.py
+            --model_name_or_path bert-base-uncased
+            --train_file tests/fixtures/tests_samples/conll/sample.json
+            --validation_file tests/fixtures/tests_samples/conll/sample.json
+            --output_dir {tmp_dir}
+            --overwrite_output_dir
+            --do_train
+            --do_eval
+            --warmup_steps=2
+            --learning_rate=2e-4
+            --per_gpu_train_batch_size=2
+            --per_gpu_eval_batch_size=2
+            --num_train_epochs=2
+        """.split()
+
+        if torch_device != "cuda":
+            testargs.append("--no_cuda")
+
+        with patch.object(sys, "argv", testargs):
+            result = run_ner.main()
+            self.assertGreaterEqual(result["eval_accuracy_score"], 0.75)
+            self.assertGreaterEqual(result["eval_precision"], 0.75)
+            self.assertLess(result["eval_loss"], 0.5)
+
+    @require_torch_non_multigpu_but_fix_me
    def test_run_squad(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)
@@ -183,6 +256,7 @@ class ExamplesTests(TestCasePlus):
            self.assertGreaterEqual(result["f1"], 25)
            self.assertGreaterEqual(result["exact"], 21)

+    @require_torch_non_multigpu_but_fix_me
    def test_generation(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)
@@ -20,7 +20,7 @@ import unittest
 from time import time
 from unittest.mock import patch

-from transformers.testing_utils import require_torch_tpu
+from transformers.testing_utils import require_torch_non_multigpu_but_fix_me, require_torch_tpu


 logging.basicConfig(level=logging.DEBUG)
@@ -30,6 +30,7 @@ logger = logging.getLogger()

@require_torch_tpu
 class TorchXLAExamplesTests(unittest.TestCase):
+    @require_torch_non_multigpu_but_fix_me
    def test_run_glue(self):
        import xla_spawn

@@ -81,6 +82,7 @@ class TorchXLAExamplesTests(unittest.TestCase):
            # Assert that the script takes less than 300 seconds to make sure it doesn't hang.
            self.assertLess(end - start, 500)

+    @require_torch_non_multigpu_but_fix_me
    def test_trainer_tpu(self):
        import xla_spawn

@@ -1,6 +1,5 @@
 # coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+# Copyright 2020 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -178,7 +177,7 @@ def main():
    set_seed(training_args.seed)

    # Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
-    # or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub
+    # or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub).
    #
    # For CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the
    # sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named
@@ -192,7 +192,7 @@ def main():

    # Optionally, predict on dev set and write to output_dir
    if args.do_predict:
-        checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "checkpointepoch=*.ckpt"), recursive=True)))
+        checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "checkpoint-epoch=*.ckpt"), recursive=True)))
        model = model.load_from_checkpoint(checkpoints[-1])
        return trainer.test(model)

@@ -1,6 +1,40 @@
-## Named Entity Recognition
+## Token classification

-Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/token-classification/run_ner.py) for Pytorch and
+Fine-tuning the library models for token classification task such as Named Entity Recognition (NER) or Parts-of-speech
+tagging (POS). The main scrip `run_ner.py` leverages the 🤗 Datasets library and the Trainer API. You can easily
+customize it to your needs if you need extra processing on your datasets.
+
+It will either run on a datasets hosted on our [hub](https://huggingface.co/datasets) or with your own text files for
+training and validation.
+
+The following example fine-tunes BERT on CoNLL-2003:
+
+```bash
+python run_ner.py \
+  --model_name_or_path bert-base-uncased \
+  --dataset_name conll2003 \
+  --output_dir /tmp/test-ner \
+  --do_train \
+  --do_eval
+```
+
+or just can just run the bash script `run.sh`.
+
+To run on your own training and validation files, use the following command:
+
+```bash
+python run_ner.py \
+  --model_name_or_path bert-base-uncased \
+  --train_file path_to_train_file \
+  --validation_file path_to_validation_file \
+  --output_dir /tmp/test-ner \
+  --do_train \
+  --do_eval
+```
+
+## Old version of the script
+
+Based on the scripts [`run_ner_old.py`](https://github.com/huggingface/transformers/blob/master/examples/token-classification/run_ner_old.py) for Pytorch and
 [`run_tf_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/token-classification/run_tf_ner.py) for Tensorflow 2.

 The following examples are covered in this section:
@@ -69,7 +103,7 @@ export SEED=1
 To start training, just run:

 ```bash
-python3 run_ner.py --data_dir ./ \
+python3 run_ner_old.py --data_dir ./ \
 --labels ./labels.txt \
 --model_name_or_path $BERT_MODEL \
 --output_dir $OUTPUT_DIR \
@@ -87,7 +121,7 @@ If your GPU supports half-precision training, just add the `--fp16` flag. After

 #### JSON-based configuration file

-Instead of passing all parameters via commandline arguments, the `run_ner.py` script also supports reading parameters from a json-based configuration file:
+Instead of passing all parameters via commandline arguments, the `run_ner_old.py` script also supports reading parameters from a json-based configuration file:

 ```json
 {
@@ -106,7 +140,7 @@ Instead of passing all parameters via commandline arguments, the `run_ner.py` sc
 }
 ```

-It must be saved with a `.json` extension and can be used by running `python3 run_ner.py config.json`.
+It must be saved with a `.json` extension and can be used by running `python3 run_ner_old.py config.json`.

 #### Evaluation

@@ -250,7 +284,7 @@ cat data_wnut_17/train.txt data_wnut_17/dev.txt data_wnut_17/test.txt | cut -d "

 #### Run the Pytorch version

-Fine-tuning with the PyTorch version can be started using the `run_ner.py` script. In this example we use a JSON-based configuration file.
+Fine-tuning with the PyTorch version can be started using the `run_ner_old.py` script. In this example we use a JSON-based configuration file.

 This configuration file looks like:

@@ -274,7 +308,7 @@ This configuration file looks like:

 If your GPU supports half-precision training, please set `fp16` to `true`.

-Save this JSON-based configuration under `wnut_17.json`. The fine-tuning can be started with `python3 run_ner.py wnut_17.json`.
+Save this JSON-based configuration under `wnut_17.json`. The fine-tuning can be started with `python3 run_ner_old.py wnut_17.json`.

 #### Evaluation

@@ -1,36 +1,6 @@
-## The relevant files are currently on a shared Google
-## drive at https://drive.google.com/drive/folders/1kC0I2UGl2ltrluI9NqDjaQJGw5iliw_J
-## Monitor for changes and eventually migrate to nlp dataset
-curl -L 'https://drive.google.com/uc?export=download&id=1Jjhbal535VVz2ap4v4r_rN1UEHTdLK5P' \
-| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > train.txt.tmp
-curl -L 'https://drive.google.com/uc?export=download&id=1ZfRcQThdtAR5PPRjIDtrVP7BtXSCUBbm' \
-| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > dev.txt.tmp
-curl -L 'https://drive.google.com/uc?export=download&id=1u9mb7kNJHWQCWyweMDRMuTFoOHOfeBTH' \
-| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > test.txt.tmp
-
-export MAX_LENGTH=128
-export BERT_MODEL=bert-base-multilingual-cased
-python3 scripts/preprocess.py train.txt.tmp $BERT_MODEL $MAX_LENGTH > train.txt
-python3 scripts/preprocess.py dev.txt.tmp $BERT_MODEL $MAX_LENGTH > dev.txt
-python3 scripts/preprocess.py test.txt.tmp $BERT_MODEL $MAX_LENGTH > test.txt
-cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
-export OUTPUT_DIR=germeval-model
-export BATCH_SIZE=32
-export NUM_EPOCHS=3
-export SAVE_STEPS=750
-export SEED=1
-
 python3 run_ner.py \
--task_type NER \
--data_dir . \
--labels ./labels.txt \
--model_name_or_path $BERT_MODEL \
--output_dir $OUTPUT_DIR \
--max_seq_length  $MAX_LENGTH \
--num_train_epochs $NUM_EPOCHS \
--per_gpu_train_batch_size $BATCH_SIZE \
--save_steps $SAVE_STEPS \
--seed $SEED \
--do_train \
--do_eval \
--do_predict
+  --model_name_or_path bert-base-uncased \
+  --dataset_name conll2003 \
+  --output_dir /tmp/test-ner \
+  --do_train \
+  --do_eval
@@ -21,7 +21,7 @@ export NUM_EPOCHS=3
 export SAVE_STEPS=750
 export SEED=1

-python3 run_ner.py \
+python3 run_ner_old.py \
 --task_type Chunk \
 --data_dir . \
 --model_name_or_path $BERT_MODEL \
@@ -1,6 +1,5 @@
 # coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+# 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.
@@ -13,29 +12,33 @@
 # 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 named entity recognition on CoNLL-2003. """
+"""
+Fine-tuning the library models for token classification.
+"""
+# You can also adapt this script on your own token classification task and datasets. Pointers for this are left as comments.
+
 import logging
 import os
 import sys
 from dataclasses import dataclass, field
-from importlib import import_module
-from typing import Dict, List, Optional, Tuple
+from typing import Optional

 import numpy as np
+from datasets import load_dataset
 from seqeval.metrics import accuracy_score, f1_score, precision_score, recall_score
-from torch import nn

+import transformers
 from transformers import (
    AutoConfig,
    AutoModelForTokenClassification,
    AutoTokenizer,
-    EvalPrediction,
+    DataCollatorForTokenClassification,
    HfArgumentParser,
    Trainer,
    TrainingArguments,
    set_seed,
 )
-from utils_ner import Split, TokenClassificationDataset, TokenClassificationTask
+from transformers.trainer_utils import is_main_process


 logger = logging.getLogger(__name__)
@@ -53,15 +56,9 @@ class ModelArguments:
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
    )
-    task_type: Optional[str] = field(
-        default="NER", metadata={"help": "Task type to fine tune in training (e.g. NER, POS, etc)"}
-    )
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
    )
-    use_fast: bool = field(default=False, metadata={"help": "Set this flag to use fast tokenization."})
-    # If you want to tweak more attributes on your tokenizer, you should do it in a distinct script,
-    # or just modify its tokenizer_config.json.
    cache_dir: Optional[str] = field(
        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
    )
@@ -73,23 +70,58 @@ class DataTrainingArguments:
    Arguments pertaining to what data we are going to input our model for training and eval.
    """

-    data_dir: str = field(
-        metadata={"help": "The input data dir. Should contain the .txt files for a CoNLL-2003-formatted task."}
+    task_name: Optional[str] = field(default="ner", metadata={"help": "The name of the task (ner, pos...)."})
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
    )
-    labels: Optional[str] = field(
+    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 csv or JSON file)."}
+    )
+    validation_file: Optional[str] = field(
        default=None,
-        metadata={"help": "Path to a file containing all labels. If not specified, CoNLL-2003 labels are used."},
+        metadata={"help": "An optional input evaluation data file to evaluate on (a csv or JSON file)."},
    )
-    max_seq_length: int = field(
-        default=128,
-        metadata={
-            "help": "The maximum total input sequence length after tokenization. Sequences longer "
-            "than this will be truncated, sequences shorter will be padded."
-        },
+    test_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input test data file to predict on (a csv or JSON file)."},
    )
    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."},
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to model maximum sentence length. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch. More "
+            "efficient on GPU but very bad for TPU."
+        },
+    )
+    label_all_tokens: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to put the label for one word on all tokens of generated by that word or just on the "
+            "one (in which case the other tokens will have a padding index)."
+        },
+    )
+
+    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."
+        self.task_name = self.task_name.lower()


 def main():
@@ -112,60 +144,90 @@ def main():
        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."
-        )
-
-    module = import_module("tasks")
-    try:
-        token_classification_task_clazz = getattr(module, model_args.task_type)
-        token_classification_task: TokenClassificationTask = token_classification_task_clazz()
-    except AttributeError:
-        raise ValueError(
-            f"Task {model_args.task_type} needs to be defined as a TokenClassificationTask subclass in {module}. "
-            f"Available tasks classes are: {TokenClassificationTask.__subclasses__()}"
+            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",
        datefmt="%m/%d/%Y %H:%M:%S",
-        level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
+        level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
    )
+
+    # Log on each process the small summary:
    logger.warning(
-        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
-        training_args.local_rank,
-        training_args.device,
-        training_args.n_gpu,
-        bool(training_args.local_rank != -1),
-        training_args.fp16,
+        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):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
    logger.info("Training/evaluation parameters %s", training_args)

-    # Set seed
+    # Set seed before initializing model.
    set_seed(training_args.seed)

-    # Prepare CONLL-2003 task
-    labels = token_classification_task.get_labels(data_args.labels)
-    label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
-    num_labels = len(labels)
+    # 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 guarantee 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.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+    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
+        if data_args.test_file is not None:
+            data_files["test"] = data_args.test_file
+        extension = data_args.train_file.split(".")[-1]
+        datasets = load_dataset(extension, data_files=data_files)
+    # 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.
+
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "words" if "words" in column_names else column_names[0]
+    label_column_name = data_args.task_name if data_args.task_name in column_names else column_names[1]
+
+    # Labeling (this part will be easier when https://github.com/huggingface/datasets/issues/797 is solved)
+    def get_label_list(labels):
+        unique_labels = set()
+        for label in labels:
+            unique_labels = unique_labels | set(label)
+        label_list = list(unique_labels)
+        label_list.sort()
+        return label_list
+
+    label_list = get_label_list(datasets["train"][label_column_name])
+    label_to_id = {l: i for i, l in enumerate(label_list)}
+    num_labels = len(label_list)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
-
    config = AutoConfig.from_pretrained(
        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
        num_labels=num_labels,
-        id2label=label_map,
-        label2id={label: i for i, label in enumerate(labels)},
+        finetuning_task=data_args.task_name,
        cache_dir=model_args.cache_dir,
    )
    tokenizer = AutoTokenizer.from_pretrained(
        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
-        use_fast=model_args.use_fast,
+        use_fast=True,
    )
    model = AutoModelForTokenClassification.from_pretrained(
        model_args.model_name_or_path,
@@ -174,67 +236,85 @@ def main():
        cache_dir=model_args.cache_dir,
    )

-    # Get datasets
-    train_dataset = (
-        TokenClassificationDataset(
-            token_classification_task=token_classification_task,
-            data_dir=data_args.data_dir,
-            tokenizer=tokenizer,
-            labels=labels,
-            model_type=config.model_type,
-            max_seq_length=data_args.max_seq_length,
-            overwrite_cache=data_args.overwrite_cache,
-            mode=Split.train,
+    # Preprocessing the dataset
+    # Padding strategy
+    padding = "max_length" if data_args.pad_to_max_length else False
+
+    # Tokenize all texts and align the labels with them.
+    def tokenize_and_align_labels(examples):
+        tokenized_inputs = tokenizer(
+            examples[text_column_name],
+            padding=padding,
+            truncation=True,
+            # We use this argument because the texts in our dataset are lists of words (with a label for each word).
+            is_split_into_words=True,
+            return_offsets_mapping=True,
        )
-        if training_args.do_train
-        else None
-    )
-    eval_dataset = (
-        TokenClassificationDataset(
-            token_classification_task=token_classification_task,
-            data_dir=data_args.data_dir,
-            tokenizer=tokenizer,
-            labels=labels,
-            model_type=config.model_type,
-            max_seq_length=data_args.max_seq_length,
-            overwrite_cache=data_args.overwrite_cache,
-            mode=Split.dev,
-        )
-        if training_args.do_eval
-        else None
+        offset_mappings = tokenized_inputs.pop("offset_mapping")
+        labels = []
+        for label, offset_mapping in zip(examples[label_column_name], offset_mappings):
+            label_index = 0
+            current_label = -100
+            label_ids = []
+            for offset in offset_mapping:
+                # We set the label for the first token of each word. Special characters will have an offset of (0, 0)
+                # so the test ignores them.
+                if offset[0] == 0 and offset[1] != 0:
+                    current_label = label_to_id[label[label_index]]
+                    label_index += 1
+                    label_ids.append(current_label)
+                # For special tokens, we set the label to -100 so it's automatically ignored in the loss function.
+                elif offset[0] == 0 and offset[1] == 0:
+                    label_ids.append(-100)
+                # For the other tokens in a word, we set the label to either the current label or -100, depending on
+                # the label_all_tokens flag.
+                else:
+                    label_ids.append(current_label if data_args.label_all_tokens else -100)
+
+            labels.append(label_ids)
+        tokenized_inputs["labels"] = labels
+        return tokenized_inputs
+
+    tokenized_datasets = datasets.map(
+        tokenize_and_align_labels,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        load_from_cache_file=not data_args.overwrite_cache,
    )

-    def align_predictions(predictions: np.ndarray, label_ids: np.ndarray) -> Tuple[List[int], List[int]]:
-        preds = np.argmax(predictions, axis=2)
+    # Data collator
+    data_collator = DataCollatorForTokenClassification(tokenizer)

-        batch_size, seq_len = preds.shape
+    # Metrics
+    def compute_metrics(p):
+        predictions, labels = p
+        predictions = np.argmax(predictions, axis=2)

-        out_label_list = [[] for _ in range(batch_size)]
-        preds_list = [[] for _ in range(batch_size)]
+        # Remove ignored index (special tokens)
+        true_predictions = [
+            [label_list[p] for (p, l) in zip(prediction, label) if l != -100]
+            for prediction, label in zip(predictions, labels)
+        ]
+        true_labels = [
+            [label_list[l] for (p, l) in zip(prediction, label) if l != -100]
+            for prediction, label in zip(predictions, labels)
+        ]

-        for i in range(batch_size):
-            for j in range(seq_len):
-                if label_ids[i, j] != nn.CrossEntropyLoss().ignore_index:
-                    out_label_list[i].append(label_map[label_ids[i][j]])
-                    preds_list[i].append(label_map[preds[i][j]])
-
-        return preds_list, out_label_list
-
-    def compute_metrics(p: EvalPrediction) -> Dict:
-        preds_list, out_label_list = align_predictions(p.predictions, p.label_ids)
        return {
-            "accuracy_score": accuracy_score(out_label_list, preds_list),
-            "precision": precision_score(out_label_list, preds_list),
-            "recall": recall_score(out_label_list, preds_list),
-            "f1": f1_score(out_label_list, preds_list),
+            "accuracy_score": accuracy_score(true_labels, true_predictions),
+            "precision": precision_score(true_labels, true_predictions),
+            "recall": recall_score(true_labels, true_predictions),
+            "f1": f1_score(true_labels, true_predictions),
        }

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
-        train_dataset=train_dataset,
-        eval_dataset=eval_dataset,
+        train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
        compute_metrics=compute_metrics,
    )

@@ -243,58 +323,50 @@ def main():
        trainer.train(
            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
        )
-        trainer.save_model()
-        # For convenience, we also re-save the tokenizer to the same directory,
-        # so that you can share your model easily on huggingface.co/models =)
-        if trainer.is_world_master():
-            tokenizer.save_pretrained(training_args.output_dir)
+        trainer.save_model()  # Saves the tokenizer too for easy upload

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

-        result = trainer.evaluate()
+        results = trainer.evaluate()

-        output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
-        if trainer.is_world_master():
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results_ner.txt")
+        if trainer.is_world_process_zero():
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
-                for key, value in result.items():
-                    logger.info("  %s = %s", key, value)
-                    writer.write("%s = %s\n" % (key, value))
-
-            results.update(result)
+                for key, value in results.items():
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")

    # Predict
    if training_args.do_predict:
-        test_dataset = TokenClassificationDataset(
-            token_classification_task=token_classification_task,
-            data_dir=data_args.data_dir,
-            tokenizer=tokenizer,
-            labels=labels,
-            model_type=config.model_type,
-            max_seq_length=data_args.max_seq_length,
-            overwrite_cache=data_args.overwrite_cache,
-            mode=Split.test,
-        )
+        logger.info("*** Predict ***")

-        predictions, label_ids, metrics = trainer.predict(test_dataset)
-        preds_list, _ = align_predictions(predictions, label_ids)
+        test_dataset = datasets["test"]
+        predictions, labels, metrics = trainer.predict(test_dataset)
+        predictions = np.argmax(predictions, axis=2)
+
+        # Remove ignored index (special tokens)
+        true_predictions = [
+            [label_list[p] for (p, l) in zip(prediction, label) if l != -100]
+            for prediction, label in zip(predictions, labels)
+        ]

        output_test_results_file = os.path.join(training_args.output_dir, "test_results.txt")
        if trainer.is_world_master():
            with open(output_test_results_file, "w") as writer:
                for key, value in metrics.items():
-                    logger.info("  %s = %s", key, value)
-                    writer.write("%s = %s\n" % (key, value))
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")

        # Save predictions
        output_test_predictions_file = os.path.join(training_args.output_dir, "test_predictions.txt")
        if trainer.is_world_master():
            with open(output_test_predictions_file, "w") as writer:
-                with open(os.path.join(data_args.data_dir, "test.txt"), "r") as f:
-                    token_classification_task.write_predictions_to_file(writer, f, preds_list)
+                for prediction in true_predictions:
+                    writer.write(" ".join(prediction) + "\n")

    return results

@@ -0,0 +1,308 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Fine-tuning the library models for named entity recognition on CoNLL-2003. """
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from importlib import import_module
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+from seqeval.metrics import accuracy_score, f1_score, precision_score, recall_score
+from torch import nn
+
+from transformers import (
+    AutoConfig,
+    AutoModelForTokenClassification,
+    AutoTokenizer,
+    EvalPrediction,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    set_seed,
+)
+from utils_ner import Split, TokenClassificationDataset, TokenClassificationTask
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    task_type: Optional[str] = field(
+        default="NER", metadata={"help": "Task type to fine tune in training (e.g. NER, POS, etc)"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    use_fast: bool = field(default=False, metadata={"help": "Set this flag to use fast tokenization."})
+    # If you want to tweak more attributes on your tokenizer, you should do it in a distinct script,
+    # or just modify its tokenizer_config.json.
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    data_dir: str = field(
+        metadata={"help": "The input data dir. Should contain the .txt files for a CoNLL-2003-formatted task."}
+    )
+    labels: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to a file containing all labels. If not specified, CoNLL-2003 labels are used."},
+    )
+    max_seq_length: int = field(
+        default=128,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated, sequences shorter will be padded."
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+
+
+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."
+        )
+
+    module = import_module("tasks")
+    try:
+        token_classification_task_clazz = getattr(module, model_args.task_type)
+        token_classification_task: TokenClassificationTask = token_classification_task_clazz()
+    except AttributeError:
+        raise ValueError(
+            f"Task {model_args.task_type} needs to be defined as a TokenClassificationTask subclass in {module}. "
+            f"Available tasks classes are: {TokenClassificationTask.__subclasses__()}"
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        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,
+    )
+    logger.warning(
+        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+        training_args.local_rank,
+        training_args.device,
+        training_args.n_gpu,
+        bool(training_args.local_rank != -1),
+        training_args.fp16,
+    )
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed
+    set_seed(training_args.seed)
+
+    # Prepare CONLL-2003 task
+    labels = token_classification_task.get_labels(data_args.labels)
+    label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
+    num_labels = len(labels)
+
+    # Load pretrained model and tokenizer
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+
+    config = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        num_labels=num_labels,
+        id2label=label_map,
+        label2id={label: i for i, label in enumerate(labels)},
+        cache_dir=model_args.cache_dir,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_fast=model_args.use_fast,
+    )
+    model = AutoModelForTokenClassification.from_pretrained(
+        model_args.model_name_or_path,
+        from_tf=bool(".ckpt" in model_args.model_name_or_path),
+        config=config,
+        cache_dir=model_args.cache_dir,
+    )
+
+    # Get datasets
+    train_dataset = (
+        TokenClassificationDataset(
+            token_classification_task=token_classification_task,
+            data_dir=data_args.data_dir,
+            tokenizer=tokenizer,
+            labels=labels,
+            model_type=config.model_type,
+            max_seq_length=data_args.max_seq_length,
+            overwrite_cache=data_args.overwrite_cache,
+            mode=Split.train,
+        )
+        if training_args.do_train
+        else None
+    )
+    eval_dataset = (
+        TokenClassificationDataset(
+            token_classification_task=token_classification_task,
+            data_dir=data_args.data_dir,
+            tokenizer=tokenizer,
+            labels=labels,
+            model_type=config.model_type,
+            max_seq_length=data_args.max_seq_length,
+            overwrite_cache=data_args.overwrite_cache,
+            mode=Split.dev,
+        )
+        if training_args.do_eval
+        else None
+    )
+
+    def align_predictions(predictions: np.ndarray, label_ids: np.ndarray) -> Tuple[List[int], List[int]]:
+        preds = np.argmax(predictions, axis=2)
+
+        batch_size, seq_len = preds.shape
+
+        out_label_list = [[] for _ in range(batch_size)]
+        preds_list = [[] for _ in range(batch_size)]
+
+        for i in range(batch_size):
+            for j in range(seq_len):
+                if label_ids[i, j] != nn.CrossEntropyLoss().ignore_index:
+                    out_label_list[i].append(label_map[label_ids[i][j]])
+                    preds_list[i].append(label_map[preds[i][j]])
+
+        return preds_list, out_label_list
+
+    def compute_metrics(p: EvalPrediction) -> Dict:
+        preds_list, out_label_list = align_predictions(p.predictions, p.label_ids)
+        return {
+            "accuracy_score": accuracy_score(out_label_list, preds_list),
+            "precision": precision_score(out_label_list, preds_list),
+            "recall": recall_score(out_label_list, preds_list),
+            "f1": f1_score(out_label_list, preds_list),
+        }
+
+    # Initialize our Trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        compute_metrics=compute_metrics,
+    )
+
+    # Training
+    if training_args.do_train:
+        trainer.train(
+            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
+        )
+        trainer.save_model()
+        # For convenience, we also re-save the tokenizer to the same directory,
+        # so that you can share your model easily on huggingface.co/models =)
+        if trainer.is_world_master():
+            tokenizer.save_pretrained(training_args.output_dir)
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+
+        result = trainer.evaluate()
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
+        if trainer.is_world_master():
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Eval results *****")
+                for key, value in result.items():
+                    logger.info("  %s = %s", key, value)
+                    writer.write("%s = %s\n" % (key, value))
+
+            results.update(result)
+
+    # Predict
+    if training_args.do_predict:
+        test_dataset = TokenClassificationDataset(
+            token_classification_task=token_classification_task,
+            data_dir=data_args.data_dir,
+            tokenizer=tokenizer,
+            labels=labels,
+            model_type=config.model_type,
+            max_seq_length=data_args.max_seq_length,
+            overwrite_cache=data_args.overwrite_cache,
+            mode=Split.test,
+        )
+
+        predictions, label_ids, metrics = trainer.predict(test_dataset)
+        preds_list, _ = align_predictions(predictions, label_ids)
+
+        output_test_results_file = os.path.join(training_args.output_dir, "test_results.txt")
+        if trainer.is_world_master():
+            with open(output_test_results_file, "w") as writer:
+                for key, value in metrics.items():
+                    logger.info("  %s = %s", key, value)
+                    writer.write("%s = %s\n" % (key, value))
+
+        # Save predictions
+        output_test_predictions_file = os.path.join(training_args.output_dir, "test_predictions.txt")
+        if trainer.is_world_master():
+            with open(output_test_predictions_file, "w") as writer:
+                with open(os.path.join(data_args.data_dir, "test.txt"), "r") as f:
+                    token_classification_task.write_predictions_to_file(writer, f, preds_list)
+
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,36 @@
+## The relevant files are currently on a shared Google
+## drive at https://drive.google.com/drive/folders/1kC0I2UGl2ltrluI9NqDjaQJGw5iliw_J
+## Monitor for changes and eventually migrate to nlp dataset
+curl -L 'https://drive.google.com/uc?export=download&id=1Jjhbal535VVz2ap4v4r_rN1UEHTdLK5P' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > train.txt.tmp
+curl -L 'https://drive.google.com/uc?export=download&id=1ZfRcQThdtAR5PPRjIDtrVP7BtXSCUBbm' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > dev.txt.tmp
+curl -L 'https://drive.google.com/uc?export=download&id=1u9mb7kNJHWQCWyweMDRMuTFoOHOfeBTH' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > test.txt.tmp
+
+export MAX_LENGTH=128
+export BERT_MODEL=bert-base-multilingual-cased
+python3 scripts/preprocess.py train.txt.tmp $BERT_MODEL $MAX_LENGTH > train.txt
+python3 scripts/preprocess.py dev.txt.tmp $BERT_MODEL $MAX_LENGTH > dev.txt
+python3 scripts/preprocess.py test.txt.tmp $BERT_MODEL $MAX_LENGTH > test.txt
+cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
+export OUTPUT_DIR=germeval-model
+export BATCH_SIZE=32
+export NUM_EPOCHS=3
+export SAVE_STEPS=750
+export SEED=1
+
+python3 run_ner_old.py \
+--task_type NER \
+--data_dir . \
+--labels ./labels.txt \
+--model_name_or_path $BERT_MODEL \
+--output_dir $OUTPUT_DIR \
+--max_seq_length  $MAX_LENGTH \
+--num_train_epochs $NUM_EPOCHS \
+--per_gpu_train_batch_size $BATCH_SIZE \
+--save_steps $SAVE_STEPS \
+--seed $SEED \
+--do_train \
+--do_eval \
+--do_predict
@@ -207,9 +207,9 @@ if __name__ == "__main__":

    if args.do_predict:
        # See https://github.com/huggingface/transformers/issues/3159
-        # pl use this format to create a checkpoint:
+        # pl use this default format to create a checkpoint:
        # https://github.com/PyTorchLightning/pytorch-lightning/blob/master\
-        # /pytorch_lightning/callbacks/model_checkpoint.py#L169
-        checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "checkpointepoch=*.ckpt"), recursive=True)))
+        # /pytorch_lightning/callbacks/model_checkpoint.py#L322
+        checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "checkpoint-epoch=*.ckpt"), recursive=True)))
        model = model.load_from_checkpoint(checkpoints[-1])
        trainer.test(model)
@@ -21,7 +21,7 @@ export NUM_EPOCHS=3
 export SAVE_STEPS=750
 export SEED=1

-python3 run_ner.py \
+python3 run_ner_old.py \
 --task_type POS \
 --data_dir . \
 --model_name_or_path $BERT_MODEL \
@@ -3,8 +3,8 @@ import sys
 import unittest
 from unittest.mock import patch

-import run_ner
-from transformers.testing_utils import slow
+import run_ner_old as run_ner
+from transformers.testing_utils import require_torch_non_multigpu_but_fix_me, slow


 logging.basicConfig(level=logging.INFO)
@@ -14,6 +14,7 @@ logger = logging.getLogger()

 class ExamplesTests(unittest.TestCase):
    @slow
+    @require_torch_non_multigpu_but_fix_me
    def test_run_ner(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)
@@ -34,6 +35,7 @@ class ExamplesTests(unittest.TestCase):
            result = run_ner.main()
            self.assertLess(result["eval_loss"], 1.5)

+    @require_torch_non_multigpu_but_fix_me
    def test_run_ner_pl(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)
@@ -66,14 +66,16 @@ class Split(Enum):


 class TokenClassificationTask:
-    def read_examples_from_file(self, data_dir, mode: Union[Split, str]) -> List[InputExample]:
+    @staticmethod
+    def read_examples_from_file(data_dir, mode: Union[Split, str]) -> List[InputExample]:
        raise NotImplementedError

-    def get_labels(self, path: str) -> List[str]:
+    @staticmethod
+    def get_labels(path: str) -> List[str]:
        raise NotImplementedError

+    @staticmethod
    def convert_examples_to_features(
-        self,
        examples: List[InputExample],
        label_list: List[str],
        max_seq_length: int,
@@ -10,11 +10,12 @@ widget:

 [MagBERT-NER] is a state-of-the-art NER model for Moroccan French language (Maghreb). The MagBERT-NER model was fine-tuned for NER Task based the language model for French Camembert (based on the RoBERTa architecture).

-For further information or requests, please go to [Typica.AI Website](https://typicasoft.io/)
+For further information or requests, please visite our website at [typica.ai Website](https://typica.ai/) or send us an email at contactus@typica.ai

 ## How to use MagBERT-NER with HuggingFace

 ##### Load MagBERT-NER and its sub-word tokenizer :
+
 ```python
 from transformers import AutoTokenizer, AutoModelForTokenClassification

@@ -45,11 +46,14 @@ nlp("Saad Dine El Otmani, né le 16 janvier 1956 à Inezgane, est un homme d'Ét

 ```

-```
-

 ## Authors 

-MagBert-NER was trained and evaluated by Hicham Assoudi, Ph.D.
+MagBert-NER Model was trained by Hicham Assoudi, Ph.D. 
+For any questions, comments you can contact me at assoudi@typica.ai


+## Citation
+
+If you use our work, please cite:
+Hicham Assoudi, Ph.D., MagBERT-NER: a state-of-the-art NER model for Moroccan French language (Maghreb), (2020)
@@ -6,5 +6,5 @@ license: apache-2.0
 ---

 <a href="https://huggingface.co/exbert/?model=albert-base-v1">
-	<img width="300px" src="https://hf-dinosaur.huggingface.co/exbert/button.png">
+	<img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
 </a>
@@ -6,5 +6,5 @@ license: apache-2.0
 ---

 <a href="https://huggingface.co/exbert/?model=albert-xxlarge-v2">
-	<img width="300px" src="https://hf-dinosaur.huggingface.co/exbert/button.png">
+	<img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
 </a>
@@ -63,5 +63,5 @@ The model achieved 13.7582/13.8026 BLEU scores on SemEval2020 Task4: Commonsense
 ```

 <a href="https://huggingface.co/exbert/?model=aliosm/ComVE-distilgpt2">
-	<img width="300px" src="https://hf-dinosaur.huggingface.co/exbert/button.png">
+	<img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
 </a>
@@ -64,5 +64,5 @@ The model achieved 16.5110/15.9299 BLEU scores on SemEval2020 Task4: Commonsense
 ```

 <a href="https://huggingface.co/exbert/?model=aliosm/ComVE-gpt2-large">
-	<img width="300px" src="https://hf-dinosaur.huggingface.co/exbert/button.png">
+	<img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
 </a>
--- a/Show More
+++ b/Show More