try

.circleci/config.yml

@@ -160,7 +160,7 @@ jobs:
         environment:
             TRANSFORMERS_IS_CI: yes
             PYTEST_TIMEOUT: 120
-        parallelism: 1
+        parallelism: 4
         steps:
             - checkout
             - run: uv pip install -e ".[quality]"
@@ -169,19 +169,19 @@ jobs:
                 command: pip freeze | tee installed.txt
             - store_artifacts:
                   path: ~/transformers/installed.txt
-            - run: python utils/check_copies.py
-            - run: python utils/check_modular_conversion.py --num_workers 4
-            - run: python utils/check_table.py
-            - run: python utils/check_dummies.py
-            - run: python utils/check_repo.py
-            - run: python utils/check_inits.py
-            - run: python utils/check_config_docstrings.py
-            - run: python utils/check_config_attributes.py
-            - run: python utils/check_doctest_list.py
-            - run: make deps_table_check_updated
-            - run: python utils/update_metadata.py --check-only
-            - run: python utils/check_docstrings.py
-            - run: python utils/check_support_list.py
+            - run: if [ "$CIRCLE_NODE_INDEX" == "0" ]; then python utils/check_copies.py; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "1" ]; then python utils/check_modular_conversion.py --num_workers 4; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "3" ]; then python utils/check_table.py; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "0" ]; then python utils/check_dummies.py; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "2" ]; then python utils/check_repo.py; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "2" ]; then python utils/check_inits.py; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "2" ]; then python utils/check_config_docstrings.py; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "2" ]; then python utils/check_config_attributes.py; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "0" ]; then python utils/check_doctest_list.py; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "0" ]; then make deps_table_check_updated; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "3" ]; then python utils/update_metadata.py --check-only; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "3" ]; then python utils/check_docstrings.py; fi
+            - run: if [ "$CIRCLE_NODE_INDEX" == "0" ]; then python utils/check_support_list.py; fi
 
 workflows:
     version: 2

.circleci/create_circleci_config.py

@@ -28,6 +28,8 @@ COMMON_ENV_VARIABLES = {
     "TRANSFORMERS_IS_CI": True,
     "PYTEST_TIMEOUT": 120,
     "RUN_PIPELINE_TESTS": False,
+    "RUN_PT_TF_CROSS_TESTS": False,
+    "RUN_PT_FLAX_CROSS_TESTS": False,
 }
 # Disable the use of {"s": None} as the output is way too long, causing the navigation on CircleCI impractical
 COMMON_PYTEST_OPTIONS = {"max-worker-restart": 0, "dist": "loadfile", "vvv": None, "rsfE":None}
@@ -175,6 +177,23 @@ class CircleCIJob:
 
 
 # JOBS
+torch_and_tf_job = CircleCIJob(
+    "torch_and_tf",
+    docker_image=[{"image":"huggingface/transformers-torch-tf-light"}],
+    additional_env={"RUN_PT_TF_CROSS_TESTS": True},
+    marker="is_pt_tf_cross_test",
+    pytest_options={"rA": None, "durations": 0},
+)
+
+
+torch_and_flax_job = CircleCIJob(
+    "torch_and_flax",
+    additional_env={"RUN_PT_FLAX_CROSS_TESTS": True},
+    docker_image=[{"image":"huggingface/transformers-torch-jax-light"}],
+    marker="is_pt_flax_cross_test",
+    pytest_options={"rA": None, "durations": 0},
+)
+
 torch_job = CircleCIJob(
     "torch",
     docker_image=[{"image": "huggingface/transformers-torch-light"}],
@@ -334,7 +353,7 @@ doc_test_job = CircleCIJob(
     pytest_num_workers=1,
 )
 
-REGULAR_TESTS = [torch_job, tf_job, flax_job, hub_job, onnx_job, tokenization_job, processor_job, generate_job, non_model_job] # fmt: skip
+REGULAR_TESTS = [torch_and_tf_job, torch_and_flax_job, torch_job, tf_job, flax_job, hub_job, onnx_job, tokenization_job, processor_job, generate_job, non_model_job] # fmt: skip
 EXAMPLES_TESTS = [examples_torch_job, examples_tensorflow_job]
 PIPELINE_TESTS = [pipelines_torch_job, pipelines_tf_job]
 REPO_UTIL_TESTS = [repo_utils_job]

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

@@ -26,7 +26,7 @@ jobs:
 
     strategy:
       matrix:
-        file: ["quality", "consistency", "custom-tokenizers", "torch-light", "tf-light", "exotic-models", "torch-tf-light", "jax-light", "examples-torch",  "examples-tf"]
+        file: ["quality", "consistency", "custom-tokenizers", "torch-light", "tf-light", "exotic-models", "torch-tf-light", "torch-jax-light", "jax-light", "examples-torch",  "examples-tf"]
     continue-on-error: true
 
     steps:
@@ -34,11 +34,11 @@ jobs:
         name: Set tag
         run: |
               if ${{contains(github.event.head_commit.message, '[build-ci-image]')}}; then
-                  echo "TAG=huggingface/transformers-${{ matrix.file }}:dev" >> "$GITHUB_ENV"
+                  echo "TAG=huggingface/transformers-${{ matrix.file }}:dev" >> "$GITHUB_ENV" 
                   echo "setting it to DEV!"
               else
                   echo "TAG=huggingface/transformers-${{ matrix.file }}" >> "$GITHUB_ENV"
-
+                  
               fi
       -
         name: Set up Docker Buildx

.github/workflows/check_failed_model_tests.yml

@@ -22,6 +22,7 @@ env:
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
   SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   TF_FORCE_GPU_ALLOW_GROWTH: true
+  RUN_PT_TF_CROSS_TESTS: 1
   CUDA_VISIBLE_DEVICES: 0,1
 
 

.github/workflows/model_jobs.yml

@@ -30,6 +30,7 @@ env:
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
   SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   TF_FORCE_GPU_ALLOW_GROWTH: true
+  RUN_PT_TF_CROSS_TESTS: 1
   CUDA_VISIBLE_DEVICES: 0,1
 
 jobs:

.github/workflows/model_jobs_amd.yml

@@ -30,6 +30,7 @@ env:
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
   SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   TF_FORCE_GPU_ALLOW_GROWTH: true
+  RUN_PT_TF_CROSS_TESTS: 1
   CUDA_VISIBLE_DEVICES: 0,1
 
 jobs:

.github/workflows/push-important-models.yml

@@ -7,13 +7,14 @@ on:
 env:
   OUTPUT_SLACK_CHANNEL_ID: "C06L2SGMEEA"
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
-  HF_HOME: /mnt/cache
-  TRANSFORMERS_IS_CI: yes
-  OMP_NUM_THREADS: 8
-  MKL_NUM_THREADS: 8
-  RUN_SLOW: yes # For gated repositories, we still need to agree to share information on the Hub repo. page in order to get access. # This token is created under the bot `hf-transformers-bot`.
-  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
-  TF_FORCE_GPU_ALLOW_GROWTH: true
+  HF_HOME: /mnt/cache 
+  TRANSFORMERS_IS_CI: yes 
+  OMP_NUM_THREADS: 8 
+  MKL_NUM_THREADS: 8 
+  RUN_SLOW: yes # For gated repositories, we still need to agree to share information on the Hub repo. page in order to get access. # This token is created under the bot `hf-transformers-bot`. 
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }} 
+  TF_FORCE_GPU_ALLOW_GROWTH: true 
+  RUN_PT_TF_CROSS_TESTS: 1
 
 jobs:
   get_modified_models:
@@ -24,13 +25,13 @@ jobs:
     steps:
       - name: Check out code
         uses: actions/checkout@v4
-
+      
       - name: Get changed files
         id: changed-files
         uses: tj-actions/changed-files@3f54ebb830831fc121d3263c1857cfbdc310cdb9 #v42
         with:
           files: src/transformers/models/**
-
+      
       - name: Run step if only the files listed above change
         if: steps.changed-files.outputs.any_changed == 'true'
         id: set-matrix
@@ -59,41 +60,41 @@ jobs:
     if: ${{ needs.get_modified_models.outputs.matrix != '[]' && needs.get_modified_models.outputs.matrix != '' && fromJson(needs.get_modified_models.outputs.matrix)[0] != null }}
     strategy:
       fail-fast: false
-      matrix:
+      matrix: 
         model-name: ${{ fromJson(needs.get_modified_models.outputs.matrix) }}
 
     steps:
       - name: Check out code
         uses: actions/checkout@v4
-
+      
       - name: Install locally transformers & other libs
         run: |
           apt install sudo
           sudo -H pip install --upgrade pip
-          sudo -H pip uninstall -y transformers
-          sudo -H pip install -U -e ".[testing]"
+          sudo -H pip uninstall -y transformers 
+          sudo -H pip install -U -e ".[testing]" 
           MAX_JOBS=4 pip install flash-attn --no-build-isolation
           pip install bitsandbytes
-
+      
       - name: NVIDIA-SMI
         run: |
           nvidia-smi
-
+      
       - name: Show installed libraries and their versions
         run: pip freeze
-
+      
       - name: Run FA2 tests
         id: run_fa2_tests
         run:
           pytest -rsfE -m "flash_attn_test" --make-reports=${{ matrix.model-name }}_fa2_tests/ tests/${{ matrix.model-name }}/test_modeling_*
-
+      
       - name: "Test suite reports artifacts: ${{ matrix.model-name }}_fa2_tests"
         if: ${{ always() }}
         uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.model-name }}_fa2_tests
           path: /transformers/reports/${{ matrix.model-name }}_fa2_tests
-
+      
       - name: Post to Slack
         if: always()
         uses: huggingface/hf-workflows/.github/actions/post-slack@main
@@ -102,13 +103,13 @@ jobs:
           title: 🤗 Results of the FA2 tests - ${{ matrix.model-name }}
           status: ${{ steps.run_fa2_tests.conclusion}}
           slack_token: ${{ secrets.CI_SLACK_BOT_TOKEN }}
-
+      
       - name: Run integration tests
         id: run_integration_tests
         if: always()
         run:
           pytest -rsfE -k "IntegrationTest"  --make-reports=tests_integration_${{ matrix.model-name }} tests/${{ matrix.model-name }}/test_modeling_*
-
+      
       - name: "Test suite reports artifacts: tests_integration_${{ matrix.model-name }}"
         if: ${{ always() }}
         uses: actions/upload-artifact@v4
@@ -118,7 +119,7 @@ jobs:
 
       - name: Post to Slack
         if: always()
-        uses: huggingface/hf-workflows/.github/actions/post-slack@main
+        uses: huggingface/hf-workflows/.github/actions/post-slack@main 
         with:
           slack_channel: ${{ env.OUTPUT_SLACK_CHANNEL_ID }}
           title: 🤗 Results of the Integration tests - ${{ matrix.model-name }}

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

@@ -22,6 +22,7 @@ env:
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
   SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   TF_FORCE_GPU_ALLOW_GROWTH: true
+  RUN_PT_TF_CROSS_TESTS: 1
   CUDA_VISIBLE_DEVICES: 0,1
 
 jobs:

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

@@ -14,6 +14,7 @@ env:
   MKL_NUM_THREADS: 8
   PYTEST_TIMEOUT: 60
   TF_FORCE_GPU_ALLOW_GROWTH: true
+  RUN_PT_TF_CROSS_TESTS: 1
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
 
 jobs:

.github/workflows/self-push.yml

@@ -24,6 +24,7 @@ env:
   MKL_NUM_THREADS: 8
   PYTEST_TIMEOUT: 60
   TF_FORCE_GPU_ALLOW_GROWTH: true
+  RUN_PT_TF_CROSS_TESTS: 1
   CUDA_VISIBLE_DEVICES: 0,1
 
 jobs:
@@ -292,7 +293,7 @@ jobs:
 
           echo "$machine_type"
           echo "machine_type=$machine_type" >> $GITHUB_ENV
-
+          
       - name: Update clone using environment variables
         working-directory: /transformers
         run: |
@@ -405,7 +406,7 @@ jobs:
 
           echo "$machine_type"
           echo "machine_type=$machine_type" >> $GITHUB_ENV
-
+          
       - name: Update clone using environment variables
         working-directory: /workspace/transformers
         run: |
@@ -515,7 +516,7 @@ jobs:
 
           echo "$machine_type"
           echo "machine_type=$machine_type" >> $GITHUB_ENV
-
+          
       - name: Update clone using environment variables
         working-directory: /workspace/transformers
         run: |
@@ -647,6 +648,6 @@ jobs:
         # `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
         run: |
           pip install huggingface_hub
-          pip install slack_sdk
+          pip install slack_sdk 
           pip show slack_sdk
           python utils/notification_service.py "${{ needs.setup.outputs.matrix }}"

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

@@ -7,19 +7,19 @@ on:
     - cron: "17 2 * * *"
   push:
     branches:
-      - refactor-from-pretrained-base-commit
+      - run_scheduled_ci*
 
 jobs:
-#  model-ci:
-#    name: Model CI
-#    uses: ./.github/workflows/self-scheduled.yml
-#    with:
-#      job: run_models_gpu
-#      slack_report_channel: "#transformers-ci-daily-models"
-#      runner: daily-ci
-#      docker: huggingface/transformers-all-latest-gpu
-#      ci_event: Daily CI
-#    secrets: inherit
+  model-ci:
+    name: Model CI
+    uses: ./.github/workflows/self-scheduled.yml
+    with:
+      job: run_models_gpu
+      slack_report_channel: "#transformers-ci-daily-models"
+      runner: daily-ci
+      docker: huggingface/transformers-all-latest-gpu
+      ci_event: Daily CI
+    secrets: inherit
 
   torch-pipeline:
     name: Torch pipeline CI

.github/workflows/self-scheduled.yml

@@ -40,6 +40,7 @@ env:
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
   SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   TF_FORCE_GPU_ALLOW_GROWTH: true
+  RUN_PT_TF_CROSS_TESTS: 1
   CUDA_VISIBLE_DEVICES: 0,1
   NUM_SLICES: 2
 
@@ -570,4 +571,4 @@ jobs:
     with:
       docker: ${{ inputs.docker }}
       start_sha: ${{ github.sha }}
-    secrets: inherit
+    secrets: inherit

.github/workflows/ssh-runner.yml

@@ -5,7 +5,7 @@ on:
     inputs:
       runner_type:
         description: 'Type of runner to test (a10 or t4)'
-        required: true
+        required: true 
       docker_image:
         description: 'Name of the Docker image'
         required: true
@@ -15,14 +15,15 @@ on:
 
 env:
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
-  HF_HOME: /mnt/cache
-  TRANSFORMERS_IS_CI: yes
-  OMP_NUM_THREADS: 8
-  MKL_NUM_THREADS: 8
-  RUN_SLOW: yes # For gated repositories, we still need to agree to share information on the Hub repo. page in order to get access. # This token is created under the bot `hf-transformers-bot`.
-  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
-  TF_FORCE_GPU_ALLOW_GROWTH: true
+  HF_HOME: /mnt/cache 
+  TRANSFORMERS_IS_CI: yes 
+  OMP_NUM_THREADS: 8 
+  MKL_NUM_THREADS: 8 
+  RUN_SLOW: yes # For gated repositories, we still need to agree to share information on the Hub repo. page in order to get access. # This token is created under the bot `hf-transformers-bot`. 
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }} 
+  TF_FORCE_GPU_ALLOW_GROWTH: true 
   CUDA_VISIBLE_DEVICES: 0,1
+  RUN_PT_TF_CROSS_TESTS: 1
 
 jobs:
   get_runner:
@@ -77,7 +78,7 @@ jobs:
       - name: Show installed libraries and their versions
         working-directory: /transformers
         run: pip freeze
-
+      
       - name: NVIDIA-SMI
         run: |
           nvidia-smi

CONTRIBUTING.md

@@ -343,6 +343,8 @@ RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./examples/pytorch/t
 
 Like the slow tests, there are other environment variables available which are not enabled by default during testing:
 - `RUN_CUSTOM_TOKENIZERS`: Enables tests for custom tokenizers.
+- `RUN_PT_FLAX_CROSS_TESTS`: Enables tests for PyTorch + Flax integration.
+- `RUN_PT_TF_CROSS_TESTS`: Enables tests for TensorFlow + PyTorch integration.
 
 More environment variables and additional information can be found in the [testing_utils.py](https://github.com/huggingface/transformers/blob/main/src/transformers/testing_utils.py).
 

conftest.py

@@ -61,6 +61,7 @@ NOT_DEVICE_TESTS = {
     "test_load_save_without_tied_weights",
     "test_tied_weights_keys",
     "test_model_weights_reload_no_missing_tied_weights",
+    "test_pt_tf_model_equivalence",
     "test_mismatched_shapes_have_properly_initialized_weights",
     "test_matched_shapes_have_loaded_weights_when_some_mismatched_shapes_exist",
     "test_model_is_small",
@@ -84,6 +85,12 @@ warnings.simplefilter(action="ignore", category=FutureWarning)
 
 
 def pytest_configure(config):
+    config.addinivalue_line(
+        "markers", "is_pt_tf_cross_test: mark test to run only when PT and TF interactions are tested"
+    )
+    config.addinivalue_line(
+        "markers", "is_pt_flax_cross_test: mark test to run only when PT and FLAX interactions are tested"
+    )
     config.addinivalue_line("markers", "is_pipeline_test: mark test to run only when pipelines are tested")
     config.addinivalue_line("markers", "is_staging_test: mark test to run only in the staging environment")
     config.addinivalue_line("markers", "accelerate_tests: mark test that require accelerate")

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

@@ -2,10 +2,10 @@ FROM rocm/dev-ubuntu-22.04:6.2.4
 LABEL maintainer="Hugging Face"
 
 ARG DEBIAN_FRONTEND=noninteractive
-ARG PYTORCH='2.6.0'
-ARG TORCH_VISION='0.21.0'
-ARG TORCH_AUDIO='2.6.0'
-ARG ROCM='6.2.4'
+ARG PYTORCH='2.5.1'
+ARG TORCH_VISION='0.20.0'
+ARG TORCH_AUDIO='2.5.0'
+ARG ROCM='6.2'
 
 RUN apt update && \
     apt install -y --no-install-recommends \
@@ -16,11 +16,9 @@ RUN apt update && \
     python-is-python3 \
     rocrand-dev \
     rocthrust-dev \
-    rocblas-dev \
-    hipsolver-dev \
     hipsparse-dev \
     hipblas-dev \
-    hipblaslt-dev && \
+    rocblas-dev && \
     apt clean && \
     rm -rf /var/lib/apt/lists/*
 

docker/transformers-quantization-latest-gpu/Dockerfile

@@ -76,9 +76,6 @@ RUN python3 -m pip install git+https://github.com/NetEase-FuXi/EETQ.git
 RUN python3 -m pip install --no-cache-dir flute-kernel==0.3.0 -i https://flute-ai.github.io/whl/cu118
 RUN python3 -m pip install --no-cache-dir fast_hadamard_transform==1.0.4.post1
 
-# Add compressed-tensors for quantization testing
-RUN python3 -m pip install --no-cache-dir compressed-tensors
-
 # When installing in editable mode, `transformers` is not recognized as a package.
 # this line must be added in order for python to be aware of transformers.
 RUN cd transformers && python3 setup.py develop

docs/source/ar/trainer.md

@@ -673,29 +673,6 @@ tpu_use_sudo: false
 use_cpu: false
 ```
 
-</hfoption>
-<hfoption id="Tensor Parallelism with PyTorch 2">
-
-```yml
-compute_environment: LOCAL_MACHINE
-tp_config:
-  tp_size: 4
-distributed_type: TP
-downcast_bf16: 'no'
-machine_rank: 0
-main_training_function: main
-mixed_precision: 'no'
-num_machines: 1
-num_processes: 4
-rdzv_backend: static
-same_network: true
-tpu_env: []
-tpu_use_cluster: false
-tpu_use_sudo: false
-use_cpu: false
-
-```
-
 </hfoption>
 </hfoptions>
 يُعد أمر  [`accelerate_launch`](https://huggingface.co/docs/accelerate/package_reference/cli#accelerate-launch) هو الطريقة المُوصى بها لتشغيل نص البرمجى للتدريب على نظام موزع باستخدام Accelerate و [`Trainer`] مع المعلمات المحددة في `config_file.yaml`. يتم حفظ هذا الملف في مجلد ذاكرة التخزين المؤقت لـ Accelerate ويتم تحميله تلقائيًا عند تشغيل `accelerate_launch`.

docs/source/de/contributing.md

@@ -283,6 +283,8 @@ RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./examples/pytorch/t
 Wie bei den langsamen Tests gibt es auch andere Umgebungsvariablen, die standardmäßig beim Testen nicht gesetzt sind:
 
 * `RUN_CUSTOM_TOKENIZERS`: Aktiviert Tests für benutzerdefinierte Tokenizer.
+* `RUN_PT_FLAX_CROSS_TESTS`: Aktiviert Tests für die Integration von PyTorch + Flax.
+* `RUN_PT_TF_CROSS_TESTS`: Aktiviert Tests für die Integration von TensorFlow + PyTorch.
 
 Weitere Umgebungsvariablen und zusätzliche Informationen finden Sie in der [testing_utils.py](src/transformers/testing_utils.py).
 

docs/source/en/_toctree.yml

@@ -461,8 +461,6 @@
         title: Granite
       - local: model_doc/granitemoe
         title: GraniteMoe
-      - local: model_doc/granitemoeshared
-        title: GraniteMoeShared
       - local: model_doc/granitevision
         title: GraniteVision
       - local: model_doc/helium

docs/source/en/index.md

@@ -173,7 +173,6 @@ Flax), PyTorch, and/or TensorFlow.
 |               [GPTSAN-japanese](model_doc/gptsan-japanese)               |       ✅        |         ❌         |      ❌      |
 |                       [Granite](model_doc/granite)                       |       ✅        |         ❌         |      ❌      |
 |                  [GraniteMoeMoe](model_doc/granitemoe)                   |       ✅        |         ❌         |      ❌      |
-|            [GraniteMoeSharedMoe](model_doc/granitemoeshared)             |       ✅        |         ❌         |      ❌      |
 |                    [Graphormer](model_doc/graphormer)                    |       ✅        |         ❌         |      ❌      |
 |                [Grounding DINO](model_doc/grounding-dino)                |       ✅        |         ❌         |      ❌      |
 |                      [GroupViT](model_doc/groupvit)                      |       ✅        |         ✅         |      ❌      |

docs/source/en/llm_tutorial_optimization.md

@@ -55,7 +55,7 @@ To give some examples of how much VRAM it roughly takes to load a model in bfloa
 
 As of writing this document, the largest GPU chip on the market is the A100 & H100 offering 80GB of VRAM. Most of the models listed before require more than 80GB just to be loaded and therefore necessarily require [tensor parallelism](https://huggingface.co/docs/transformers/perf_train_gpu_many#tensor-parallelism) and/or [pipeline parallelism](https://huggingface.co/docs/transformers/perf_train_gpu_many#naive-model-parallelism-vertical-and-pipeline-parallelism).
 
-🤗 Transformers now supports tensor parallelism for supported models having `base_tp_plan` in their respecitve config classes. Learn more about Tensor Parallelism [here](perf_train_gpu_many#tensor-parallelism). Furthermore, if you're interested in writing models in a tensor-parallelism-friendly way, feel free to have a look at [the text-generation-inference library](https://github.com/huggingface/text-generation-inference/tree/main/server/text_generation_server/models/custom_modeling).
+🤗 Transformers does not support tensor parallelism out of the box as it requires the model architecture to be written in a specific way. If you're interested in writing models in a tensor-parallelism-friendly way, feel free to have a look at [the text-generation-inference library](https://github.com/huggingface/text-generation-inference/tree/main/server/text_generation_server/models/custom_modeling).
 
 Naive pipeline parallelism is supported out of the box. For this, simply load the model with `device="auto"` which will automatically place the different layers on the available GPUs as explained [here](https://huggingface.co/docs/accelerate/v0.22.0/en/concept_guides/big_model_inference).
 Note, however that while very effective, this naive pipeline parallelism does not tackle the issues of GPU idling. For this more advanced pipeline parallelism is required as explained [here](https://huggingface.co/docs/transformers/en/perf_train_gpu_many#naive-model-parallelism-vertical-and-pipeline-parallelism).

docs/source/en/model_doc/granitemoeshared.md

@@ -1,66 +0,0 @@
-<!--Copyright 2025 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
-
-⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
-rendered properly in your Markdown viewer.
-
--->
-
-# GraniteMoeShared
-
-## Overview
-
-
-The GraniteMoe model was proposed in [Power Scheduler: A Batch Size and Token Number Agnostic Learning Rate Scheduler](https://arxiv.org/abs/2408.13359) by Yikang Shen, Matthew Stallone, Mayank Mishra, Gaoyuan Zhang, Shawn Tan, Aditya Prasad, Adriana Meza Soria, David D. Cox and Rameswar Panda.
-
-Additionally this class GraniteMoeSharedModel adds shared experts for Moe.
-
-```python
-import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-model_path = "ibm-research/moe-7b-1b-active-shared-experts"
-tokenizer = AutoTokenizer.from_pretrained(model_path)
-
-# drop device_map if running on CPU
-model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto")
-model.eval()
-
-# change input text as desired
-prompt = "Write a code to find the maximum value in a list of numbers."
-
-# tokenize the text
-input_tokens = tokenizer(prompt, return_tensors="pt")
-# generate output tokens
-output = model.generate(**input_tokens, max_new_tokens=100)
-# decode output tokens into text
-output = tokenizer.batch_decode(output)
-# loop over the batch to print, in this example the batch size is 1
-for i in output:
-    print(i)
-```
-
-This HF implementation is contributed by [Mayank Mishra](https://huggingface.co/mayank-mishra), [Shawn Tan](https://huggingface.co/shawntan) and [Sukriti Sharma](https://huggingface.co/SukritiSharma).
-
-
-## GraniteMoeSharedConfig
-
-[[autodoc]] GraniteMoeSharedConfig
-
-## GraniteMoeSharedModel
-
-[[autodoc]] GraniteMoeSharedModel
-    - forward
-
-## GraniteMoeSharedForCausalLM
-
-[[autodoc]] GraniteMoeSharedForCausalLM
-    - forward

docs/source/en/perf_infer_gpu_one.md

@@ -60,7 +60,6 @@ FlashAttention-2 is currently supported for the following architectures:
 * [GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj#transformers.GPTJModel)
 * [Granite](https://huggingface.co/docs/transformers/model_doc/granite#transformers.GraniteModel)
 * [GraniteMoe](https://huggingface.co/docs/transformers/model_doc/granitemoe#transformers.GraniteMoeModel)
-* [GraniteMoeShared](https://huggingface.co/docs/transformers/model_doc/granitemoeshared#transformers.GraniteMoeSharedModel)
 * [Idefics2](https://huggingface.co/docs/transformers/model_doc/idefics2#transformers.Idefics2Model)
 * [Idefics3](https://huggingface.co/docs/transformers/model_doc/idefics3#transformers.Idefics3Model)
 * [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
@@ -267,7 +266,6 @@ For now, Transformers supports SDPA inference and training for the following arc
 * [Idefics3](https://huggingface.co/docs/transformers/model_doc/idefics3#transformers.Idefics3Model)
 * [I-JEPA](https://huggingface.co/docs/transformers/model_doc/ijepa#transformers.IJepaModel)
 * [GraniteMoe](https://huggingface.co/docs/transformers/model_doc/granitemoe#transformers.GraniteMoeModel)
-* [GraniteMoeShared](https://huggingface.co/docs/transformers/model_doc/granitemoeshared#transformers.GraniteMoeSharedModel)
 * [JetMoe](https://huggingface.co/docs/transformers/model_doc/jetmoe#transformers.JetMoeModel)
 * [Jamba](https://huggingface.co/docs/transformers/model_doc/jamba#transformers.JambaModel)
 * [Llama](https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel)

docs/source/en/perf_train_gpu_many.md

@@ -450,13 +450,12 @@ Implementations:
 - [parallelformers](https://github.com/tunib-ai/parallelformers) (only inference at the moment)
 - [SageMaker](https://arxiv.org/abs/2111.05972) - this is a proprietary solution that can only be used on AWS.
 - [OSLO](https://github.com/tunib-ai/oslo) has the tensor parallelism implementation based on the Transformers.
-- [`transformers` integration](main_classes/trainer) tensor parallelism is available through tp_size attribute for models having `base_tp_plan`. Further you can look at [example usage](perf_infer_gpu_multi)
 
 SageMaker combines TP with DP for a more efficient processing.
 
 🤗 Transformers status:
-- core: uses PyTorch 2 APIs to support tensor parallelism to models having base_tp_plan in their respective config classes.
-- Alternatively, you can as well try [parallelformers](https://github.com/tunib-ai/parallelformers) that provides this support for most of our models. Training mode with TP is as well supported natively in transformers.
+- core: not yet implemented in the core
+- but if you want inference [parallelformers](https://github.com/tunib-ai/parallelformers) provides this support for most of our models. So until this is implemented in the core you can use theirs. And hopefully training mode will be supported too.
 - Deepspeed-Inference also supports our BERT, GPT-2, and GPT-Neo models in their super-fast CUDA-kernel-based inference mode, see more [here](https://www.deepspeed.ai/tutorials/inference-tutorial/)
 
 🤗 Accelerate integrates with [TP from Megatron-LM](https://huggingface.co/docs/accelerate/v0.23.0/en/usage_guides/megatron_lm).
@@ -536,7 +535,7 @@ Important papers:
 - [Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model](
 https://arxiv.org/abs/2201.11990)
 
-🤗 Transformers status: not yet implemented, since we have no PP.
+🤗 Transformers status: not yet implemented, since we have no PP and TP.
 
 ## FlexFlow
 

docs/source/en/trainer.md

@@ -799,29 +799,6 @@ tpu_use_sudo: false
 use_cpu: false
 ```
 
-</hfoption>
-<hfoption id="Tensor Parallelism with PyTorch 2">
-
-```yml
-compute_environment: LOCAL_MACHINE
-tp_config:
-  tp_size: 4
-distributed_type: TP
-downcast_bf16: 'no'
-machine_rank: 0
-main_training_function: main
-mixed_precision: 'no'
-num_machines: 1
-num_processes: 4
-rdzv_backend: static
-same_network: true
-tpu_env: []
-tpu_use_cluster: false
-tpu_use_sudo: false
-use_cpu: false
-
-```
-
 </hfoption>
 </hfoptions>
 

docs/source/es/trainer.md

@@ -361,30 +361,6 @@ use_cpu: false
 
 ```
 
-</hfoption>
-
-<hfoption id="Tensor Parallelism with PyTorch 2">
-
-```yml
-compute_environment: LOCAL_MACHINE
-tp_config:
-  tp_size: 4
-distributed_type: TP
-downcast_bf16: 'no'
-machine_rank: 0
-main_training_function: main
-mixed_precision: 'no'
-num_machines: 1
-num_processes: 4
-rdzv_backend: static
-same_network: true
-tpu_env: []
-tpu_use_cluster: false
-tpu_use_sudo: false
-use_cpu: false
-
-```
-
 </hfoption>
 </hfoptions>
 

docs/source/ko/contributing.md

@@ -85,7 +85,7 @@ python src/transformers/commands/transformers_cli.py env
 3. 해당 기능의 사용법을 보여주는 *코드 스니펫*을 제공해 주세요.
 4. 기능과 관련된 논문이 있는 경우 링크를 포함해 주세요.
 
-이슈가 잘 작성되었다면 이슈가 생성된 순간, 이미 80% 정도의 작업이 완료된 것입니다.
+이슈가 잘 작성되었다면 이슈가 생성된 순간, 이미 80% 정도의 작업이 완료된 것입니다. 
 
 이슈를 제기하는 데 도움이 될 만한 [템플릿](https://github.com/huggingface/transformers/tree/main/templates)도 준비되어 있습니다.
 
@@ -140,7 +140,7 @@ python src/transformers/commands/transformers_cli.py env
    ```
 
    만약 이미 가상 환경에 🤗 Transformers가 설치되어 있다면, `-e` 플래그를 사용하여 설치하기 전에 `pip uninstall transformers`로 제거해주세요.
-
+   
    여러분의 운영체제에 따라서, 그리고 🤗 Transformers의 선택적 의존성의 수가 증가하면서, 이 명령이 실패할 수도 있습니다. 그럴 경우 사용하려는 딥러닝 프레임워크(PyTorch, TensorFlow, 그리고/또는 Flax)를 설치한 후 아래 명령을 실행해주세요:
 
    ```bash
@@ -188,7 +188,7 @@ python src/transformers/commands/transformers_cli.py env
    이러한 검사에 대해 자세히 알아보고 관련 문제를 해결하는 방법은 [Pull Request에 대한 검사](https://huggingface.co/docs/transformers/pr_checks) 가이드를 확인하세요.
 
    만약 `docs/source` 디렉터리 아래의 문서를 수정하는 경우, 문서가 빌드될 수 있는지 확인하세요. 이 검사는 Pull Request를 열 때도 CI에서 실행됩니다. 로컬 검사를 실행하려면 문서 빌더를 설치해야 합니다:
-
+   
    ```bash
    pip install ".[docs]"
    ```
@@ -216,7 +216,7 @@ python src/transformers/commands/transformers_cli.py env
    git fetch upstream
    git rebase upstream/main
    ```
-
+   
    변경 사항을 브랜치에 푸시하세요:
 
    ```bash
@@ -238,7 +238,7 @@ python src/transformers/commands/transformers_cli.py env
 ☐ 새로운 기능을 추가하는 경우, 해당 기능에 대한 테스트도 추가하세요.<br>
    - 새 모델을 추가하는 경우, `ModelTester.all_model_classes = (MyModel, MyModelWithLMHead,...)`을 사용하여 일반적인 테스트를 활성화하세요.
    - 새 `@slow` 테스트를 추가하는 경우, 다음 명령으로 테스트를 통과하는지 확인하세요: `RUN_SLOW=1 python -m pytest tests/models/my_new_model/test_my_new_model.py`.
-   - 새 토크나이저를 추가하는 경우, 테스트를 작성하고 다음 명령으로 테스트를 통과하는지 확인하세요: `RUN_SLOW=1 python -m pytest tests/models/{your_model_name}/test_tokenization_{your_model_name}.py`.
+   - 새 토크나이저를 추가하는 경우, 테스트를 작성하고 다음 명령으로 테스트를 통과하는지 확인하세요: `RUN_SLOW=1 python -m pytest tests/models/{your_model_name}/test_tokenization_{your_model_name}.py`. 
    - CircleCI에서는 느린 테스트를 실행하지 않지만, GitHub Actions에서는 매일 밤 실행됩니다!<br>
 
 ☐ 모든 공개 메소드는 유용한 기술문서를 가져야 합니다 (예를 들어 [`modeling_bert.py`](https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py) 참조).<br>
@@ -282,6 +282,8 @@ RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./examples/pytorch/t
 
 느린 테스트와 마찬가지로, 다음과 같이 테스트 중에 기본적으로 활성화되지 않는 다른 환경 변수도 있습니다:
 - `RUN_CUSTOM_TOKENIZERS`: 사용자 정의 토크나이저 테스트를 활성화합니다.
+- `RUN_PT_FLAX_CROSS_TESTS`: PyTorch + Flax 통합 테스트를 활성화합니다.
+- `RUN_PT_TF_CROSS_TESTS`: TensorFlow + PyTorch 통합 테스트를 활성화합니다.
 
 더 많은 환경 변수와 추가 정보는 [testing_utils.py](src/transformers/testing_utils.py)에서 찾을 수 있습니다.
 

docs/source/ko/trainer.md

@@ -548,29 +548,6 @@ tpu_use_sudo: false
 use_cpu: false
 ```
 
-</hfoption>
-<hfoption id="Tensor Parallelism with PyTorch 2">
-
-```yml
-compute_environment: LOCAL_MACHINE
-tp_config:
-  tp_size: 4
-distributed_type: TP
-downcast_bf16: 'no'
-machine_rank: 0
-main_training_function: main
-mixed_precision: 'no'
-num_machines: 1
-num_processes: 4
-rdzv_backend: static
-same_network: true
-tpu_env: []
-tpu_use_cluster: false
-tpu_use_sudo: false
-use_cpu: false
-
-```
-
 </hfoption>
 </hfoptions>
 

docs/source/zh/contributing.md

@@ -33,7 +33,7 @@ limitations under the License.
 * 实现新的模型。
 * 为示例或文档做贡献。
 
-如果你不知道从哪里开始，有一个特别的 [Good First Issue](https://github.com/huggingface/transformers/contribute) 列表。它会列出一些适合初学者的开放的 issues，并帮助你开始为开源项目做贡献。只需要在你想要处理的 issue 下发表评论就行。
+如果你不知道从哪里开始，有一个特别的 [Good First Issue](https://github.com/huggingface/transformers/contribute) 列表。它会列出一些适合初学者的开放的 issues，并帮助你开始为开源项目做贡献。只需要在你想要处理的 issue 下发表评论就行。 
 
 如果想要稍微更有挑战性的内容，你也可以查看 [Good Second Issue](https://github.com/huggingface/transformers/labels/Good%20Second%20Issue) 列表。总的来说，如果你觉得自己知道该怎么做，就去做吧，我们会帮助你达到目标的！🚀
 
@@ -139,7 +139,7 @@ python src/transformers/commands/transformers_cli.py env
    ```
 
    如果在虚拟环境中已经安装了 🤗 Transformers，请先使用 `pip uninstall transformers` 卸载它，然后再用 `-e` 参数以可编辑模式重新安装。
-
+   
    根据你的操作系统，以及 Transformers 的可选依赖项数量的增加，可能会在执行此命令时出现失败。如果出现这种情况，请确保已经安装了你想使用的深度学习框架（PyTorch, TensorFlow 和 Flax），然后执行以下操作：
 
    ```bash
@@ -187,7 +187,7 @@ python src/transformers/commands/transformers_cli.py env
    想要了解有关这些检查及如何解决相关问题的更多信息，请阅读 [检查 Pull Request](https://huggingface.co/docs/transformers/pr_checks) 指南。
 
    如果你修改了 `docs/source` 目录下的文档，请确保文档仍然能够被构建。这个检查也会在你创建 PR 时在 CI 中运行。如果要进行本地检查，请确保安装了文档构建工具：
-
+   
    ```bash
    pip install ".[docs]"
    ```
@@ -281,6 +281,8 @@ RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./examples/pytorch/t
 
 和时间较长的测试一样，还有其他环境变量在测试过程中，在默认情况下是未启用的：
 - `RUN_CUSTOM_TOKENIZERS`: 启用自定义分词器的测试。
+- `RUN_PT_FLAX_CROSS_TESTS`: 启用 PyTorch + Flax 整合的测试。
+- `RUN_PT_TF_CROSS_TESTS`: 启用 TensorFlow + PyTorch 整合的测试。
 
 更多环境变量和额外信息可以在 [testing_utils.py](src/transformers/testing_utils.py) 中找到。
 

examples/flax/question-answering/run_qa.py

@@ -61,7 +61,7 @@ from transformers.utils import check_min_version, send_example_telemetry
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 Array = Any
 Dataset = datasets.arrow_dataset.Dataset

examples/flax/speech-recognition/run_flax_speech_recognition_seq2seq.py

@@ -60,7 +60,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risk.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=2.14.0", "To fix: pip install -r examples/flax/speech-recognition/requirements.txt")
 

examples/flax/text-classification/run_flax_glue.py

@@ -56,7 +56,7 @@ from transformers.utils import check_min_version, send_example_telemetry
 
 logger = logging.getLogger(__name__)
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 Array = Any
 Dataset = datasets.arrow_dataset.Dataset

examples/flax/token-classification/run_flax_ner.py

@@ -57,7 +57,7 @@ from transformers.utils.versions import require_version
 
 logger = logging.getLogger(__name__)
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/token-classification/requirements.txt")
 

examples/pytorch/audio-classification/run_audio_classification.py

@@ -45,7 +45,7 @@ from transformers.utils.versions import require_version
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.14.0", "To fix: pip install -r examples/pytorch/audio-classification/requirements.txt")
 

examples/pytorch/contrastive-image-text/run_clip.py

@@ -54,7 +54,7 @@ from transformers.utils.versions import require_version
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/contrastive-image-text/requirements.txt")
 

examples/pytorch/image-classification/run_image_classification.py

@@ -57,7 +57,7 @@ from transformers.utils.versions import require_version
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/image-classification/requirements.txt")
 

examples/pytorch/image-classification/run_image_classification_no_trainer.py

@@ -49,7 +49,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = get_logger(__name__)
 

examples/pytorch/image-pretraining/run_mae.py

@@ -43,7 +43,7 @@ from transformers.utils.versions import require_version
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/image-pretraining/requirements.txt")
 

examples/pytorch/image-pretraining/run_mim.py

@@ -48,7 +48,7 @@ Any model supported by the AutoModelForMaskedImageModeling API can be used.
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/image-pretraining/requirements.txt")
 

examples/pytorch/image-pretraining/run_mim_no_trainer.py

@@ -53,7 +53,7 @@ Any model supported by the AutoModelForMaskedImageModeling API can be used.
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/image-pretraining/requirements.txt")
 

examples/pytorch/instance-segmentation/run_instance_segmentation.py

@@ -46,7 +46,7 @@ from transformers.utils.versions import require_version
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=2.0.0", "To fix: pip install -r examples/pytorch/instance-segmentation/requirements.txt")
 

examples/pytorch/instance-segmentation/run_instance_segmentation_no_trainer.py

@@ -52,7 +52,7 @@ from transformers.utils.versions import require_version
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=2.0.0", "To fix: pip install -r examples/pytorch/instance-segmentation/requirements.txt")
 

examples/pytorch/language-modeling/run_clm.py

@@ -55,7 +55,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")
 

examples/pytorch/language-modeling/run_clm_no_trainer.py

@@ -57,7 +57,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = get_logger(__name__)
 

examples/pytorch/language-modeling/run_fim.py

@@ -58,7 +58,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")
 

examples/pytorch/language-modeling/run_fim_no_trainer.py

@@ -60,7 +60,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = get_logger(__name__)
 

examples/pytorch/language-modeling/run_mlm.py

@@ -54,7 +54,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")
 

examples/pytorch/language-modeling/run_mlm_no_trainer.py

@@ -57,7 +57,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = get_logger(__name__)
 require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")

examples/pytorch/language-modeling/run_plm.py

@@ -47,7 +47,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")
 

examples/pytorch/multiple-choice/run_swag.py

@@ -46,7 +46,7 @@ from transformers.utils import check_min_version, send_example_telemetry
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = logging.getLogger(__name__)
 

examples/pytorch/multiple-choice/run_swag_no_trainer.py

@@ -54,7 +54,7 @@ from transformers.utils import check_min_version, send_example_telemetry
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = get_logger(__name__)
 # You should update this to your particular problem to have better documentation of `model_type`

examples/pytorch/object-detection/run_object_detection.py

@@ -48,7 +48,7 @@ from transformers.utils.versions import require_version
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=2.0.0", "To fix: pip install -r examples/pytorch/object-detection/requirements.txt")
 

examples/pytorch/object-detection/run_object_detection_no_trainer.py

@@ -51,7 +51,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logging.basicConfig(level=logging.INFO)
 logger = get_logger(__name__)

examples/pytorch/question-answering/run_qa.py

@@ -50,7 +50,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/question-answering/requirements.txt")
 

examples/pytorch/question-answering/run_qa_beam_search.py

@@ -48,7 +48,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/question-answering/requirements.txt")
 

examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py

@@ -56,7 +56,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/question-answering/requirements.txt")
 

examples/pytorch/question-answering/run_qa_no_trainer.py

@@ -57,7 +57,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/question-answering/requirements.txt")
 

examples/pytorch/question-answering/run_seq2seq_qa.py

@@ -46,7 +46,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/question-answering/requirements.txt")
 

examples/pytorch/semantic-segmentation/run_semantic_segmentation.py

@@ -51,7 +51,7 @@ from transformers.utils.versions import require_version
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=2.0.0", "To fix: pip install -r examples/pytorch/semantic-segmentation/requirements.txt")
 

examples/pytorch/semantic-segmentation/run_semantic_segmentation_no_trainer.py

@@ -50,7 +50,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = get_logger(__name__)
 

examples/pytorch/speech-recognition/run_speech_recognition_ctc.py

@@ -50,7 +50,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.18.0", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
 

examples/pytorch/speech-recognition/run_speech_recognition_ctc_adapter.py

@@ -53,7 +53,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.18.0", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
 

examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py

@@ -48,7 +48,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.18.0", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
 

examples/pytorch/summarization/run_summarization.py

@@ -52,7 +52,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/summarization/requirements.txt")
 

examples/pytorch/summarization/run_summarization_no_trainer.py

@@ -56,7 +56,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = get_logger(__name__)
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/summarization/requirements.txt")

examples/pytorch/text-classification/run_classification.py

@@ -47,7 +47,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
 

examples/pytorch/text-classification/run_glue.py

@@ -48,7 +48,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
 

examples/pytorch/text-classification/run_glue_no_trainer.py

@@ -49,7 +49,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = get_logger(__name__)
 

examples/pytorch/text-classification/run_xnli.py

@@ -48,7 +48,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
 

examples/pytorch/token-classification/run_ner.py

@@ -49,7 +49,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/token-classification/requirements.txt")
 

examples/pytorch/token-classification/run_ner_no_trainer.py

@@ -56,7 +56,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = get_logger(__name__)
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/token-classification/requirements.txt")

examples/pytorch/translation/run_translation.py

@@ -52,7 +52,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/translation/requirements.txt")
 

examples/pytorch/translation/run_translation_no_trainer.py

@@ -57,7 +57,7 @@ from transformers.utils.versions import require_version
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = get_logger(__name__)
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/translation/requirements.txt")

examples/quantization/custom_quantization.py

@@ -1,78 +0,0 @@
-import json
-from typing import Any, Dict
-
-import torch
-
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from transformers.quantizers import HfQuantizer, register_quantization_config, register_quantizer
-from transformers.utils.quantization_config import QuantizationConfigMixin
-
-
-@register_quantization_config("custom")
-class CustomConfig(QuantizationConfigMixin):
-    def __init__(self):
-        self.quant_method = "custom"
-        self.bits = 8
-
-    def to_dict(self) -> Dict[str, Any]:
-        output = {
-            "num_bits": self.bits,
-        }
-        return output
-
-    def __repr__(self):
-        config_dict = self.to_dict()
-        return f"{self.__class__.__name__} {json.dumps(config_dict, indent=2, sort_keys=True)}\n"
-
-    def to_diff_dict(self) -> Dict[str, Any]:
-        config_dict = self.to_dict()
-
-        default_config_dict = CustomConfig().to_dict()
-
-        serializable_config_dict = {}
-
-        for key, value in config_dict.items():
-            if value != default_config_dict[key]:
-                serializable_config_dict[key] = value
-
-        return serializable_config_dict
-
-
-@register_quantizer("custom")
-class CustomQuantizer(HfQuantizer):
-    def __init__(self, quantization_config: QuantizationConfigMixin, **kwargs):
-        super().__init__(quantization_config, **kwargs)
-        self.quantization_config = quantization_config
-        self.scale_map = {}
-        self.device = kwargs.get("device", "cuda" if torch.cuda.is_available() else "cpu")
-        self.torch_dtype = kwargs.get("torch_dtype", torch.float32)
-
-    def _process_model_before_weight_loading(self, model, **kwargs):
-        return True
-
-    def _process_model_after_weight_loading(self, model, **kwargs):
-        return True
-
-    def is_serializable(self) -> bool:
-        return True
-
-    def is_trainable(self) -> bool:
-        return False
-
-
-model_8bit = AutoModelForCausalLM.from_pretrained(
-    "facebook/opt-350m", quantization_config=CustomConfig(), torch_dtype="auto"
-)
-
-tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
-input_text = "once there is"
-inputs = tokenizer(input_text, return_tensors="pt")
-output = model_8bit.generate(
-    **inputs,
-    max_length=100,
-    num_return_sequences=1,
-    no_repeat_ngram_size=2,
-)
-generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
-
-print(generated_text)

examples/quantization/custom_quantization_int8_example.py

@@ -1,257 +0,0 @@
-import json
-from typing import Any, Dict, List, Optional
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from accelerate import init_empty_weights
-from huggingface_hub import HfApi
-
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from transformers.quantizers import HfQuantizer, get_module_from_name, register_quantization_config, register_quantizer
-from transformers.utils.quantization_config import QuantizationConfigMixin
-
-
-# Implement INT8 Symmetric Linear layer
-class Int8SymmetricLinear(torch.nn.Module):
-    def __init__(self, in_features, out_features, bias, dtype=torch.float32):
-        super().__init__()
-        self.in_features = in_features
-        self.out_features = out_features
-
-        self.register_buffer("weight", torch.zeros((out_features, in_features), dtype=torch.int8))
-        self.register_buffer("weight_scale", torch.zeros((out_features, 1), dtype=dtype))
-
-        if bias:
-            self.register_buffer("bias", torch.zeros((self.out_features), dtype=dtype))
-        else:
-            self.bias = None
-
-    def forward(self, x):
-        dequant_weight = self.weight * self.weight_scale
-        output = F.linear(x, dequant_weight)
-        if self.bias is not None:
-            output = output + self.bias
-        return output
-
-
-# Function to replace standard linear layers with INT8 symmetric quantized layers
-def _replace_with_int8_symmetric_linear(
-    model,
-    modules_to_not_convert=None,
-    current_key_name=None,
-    quantization_config=None,
-    has_been_replaced=False,
-    pre_quantized=False,
-):
-    """
-    Recursively replaces nn.Linear modules with Int8SymmetricLinear modules.
-    """
-    if current_key_name is None:
-        current_key_name = []
-
-    for name, module in model.named_children():
-        current_key_name.append(name)
-
-        if (isinstance(module, nn.Linear)) and name not in modules_to_not_convert:
-            # Check if the current key is not in the `modules_to_not_convert`
-            current_key_name_str = ".".join(current_key_name)
-            if not any(
-                (key + "." in current_key_name_str) or (key == current_key_name_str) for key in modules_to_not_convert
-            ):
-                with init_empty_weights(include_buffers=True):
-                    in_features = module.in_features
-                    out_features = module.out_features
-                    model._modules[name] = Int8SymmetricLinear(
-                        in_features, out_features, module.bias is not None, dtype=module.weight.dtype
-                    )
-                    has_been_replaced = True
-                    model._modules[name].requires_grad_(False)
-
-        if len(list(module.children())) > 0:
-            _, has_been_replaced = _replace_with_int8_symmetric_linear(
-                module,
-                modules_to_not_convert,
-                current_key_name,
-                quantization_config,
-                has_been_replaced=has_been_replaced,
-                pre_quantized=pre_quantized,
-            )
-        # Remove the last key for recursion
-        current_key_name.pop(-1)
-    return model, has_been_replaced
-
-
-def replace_with_int8_symmetric_linear(
-    model, modules_to_not_convert=None, current_key_name=None, quantization_config=None, pre_quantized=False
-):
-    """
-    Main function to replace model layers with INT8 symmetric quantized versions.
-    """
-    modules_to_not_convert = ["lm_head"] if modules_to_not_convert is None else modules_to_not_convert
-
-    if quantization_config.modules_to_not_convert is not None:
-        modules_to_not_convert.extend(quantization_config.modules_to_not_convert)
-    modules_to_not_convert = list(set(modules_to_not_convert))
-
-    model, has_been_replaced = _replace_with_int8_symmetric_linear(
-        model, modules_to_not_convert, current_key_name, quantization_config, pre_quantized=pre_quantized
-    )
-
-    if not has_been_replaced:
-        raise ValueError(
-            "You are loading your model using INT8 symmetric quantization but no linear modules were found in your model."
-        )
-
-    return model
-
-
-@register_quantization_config("int8_symmetric")
-class Int8SymmetricConfig(QuantizationConfigMixin):
-    """
-    Configuration for INT8 symmetric quantization.
-    """
-
-    def __init__(self, modules_to_not_convert: Optional[List[str]] = None, **kwargs):
-        self.quant_method = "int8_symmetric"
-        self.modules_to_not_convert = modules_to_not_convert
-
-    def __repr__(self):
-        config_dict = self.to_dict()
-        return f"{self.__class__.__name__} {json.dumps(config_dict, indent=2, sort_keys=True)}\n"
-
-    def to_diff_dict(self) -> Dict[str, Any]:
-        config_dict = self.to_dict()
-        default_config_dict = Int8SymmetricConfig().to_dict()
-
-        serializable_config_dict = {}
-        for key, value in config_dict.items():
-            if value != default_config_dict[key]:
-                serializable_config_dict[key] = value
-
-        return serializable_config_dict
-
-
-@register_quantizer("int8_symmetric")
-class Int8SymmetricQuantizer(HfQuantizer):
-    """
-    Implementation of INT8 symmetric quantization.
-
-    """
-
-    requires_calibration = False
-    requires_parameters_quantization = True
-
-    def __init__(self, quantization_config: QuantizationConfigMixin, **kwargs):
-        super().__init__(quantization_config, **kwargs)
-        self.quantization_config = quantization_config
-
-    def _process_model_before_weight_loading(self, model, **kwargs):
-        """
-        Replace model's linear layers with quantized versions before loading weights.
-        """
-        self.modules_to_not_convert = self.quantization_config.modules_to_not_convert
-
-        model = replace_with_int8_symmetric_linear(
-            model,
-            modules_to_not_convert=self.modules_to_not_convert,
-            quantization_config=self.quantization_config,
-            pre_quantized=self.pre_quantized,
-        )
-
-    def check_quantized_param(
-        self,
-        model,
-        param_value: "torch.Tensor",
-        param_name: str,
-        state_dict: Dict[str, Any],
-        **kwargs,
-    ):
-        module, tensor_name = get_module_from_name(model, param_name)
-
-        if isinstance(module, Int8SymmetricLinear):
-            if self.pre_quantized or tensor_name == "bias":
-                if tensor_name == "weight" and param_value.dtype != torch.int8:
-                    raise ValueError("Expect quantized weights but got an unquantized weight")
-                return False
-            else:
-                if tensor_name == "weight_scale":
-                    raise ValueError("Expect unquantized weights but got a quantized weight_scale")
-                return True
-        return False
-
-    def create_quantized_param(
-        self,
-        model,
-        param_value: "torch.Tensor",
-        param_name: str,
-        target_device: "torch.device",
-        state_dict: Dict[str, Any],
-        unexpected_keys: Optional[List[str]] = None,
-    ):
-        """
-        Quantizes weights to INT8 symmetric format.
-        """
-        abs_max_per_row = torch.max(torch.abs(param_value), dim=1, keepdim=True)[0].clamp(min=1e-5)
-
-        weight_scale = abs_max_per_row / 127.0
-
-        weight_quantized = torch.round(param_value / weight_scale).clamp(-128, 127).to(torch.int8)
-
-        module, tensor_name = get_module_from_name(model, param_name)
-        module._buffers[tensor_name] = weight_quantized.to(target_device)
-        module._buffers["weight_scale"] = weight_scale.to(target_device)
-
-    def update_missing_keys(self, model, missing_keys: List[str], prefix: str) -> List[str]:
-        not_missing_keys = []
-        for name, module in model.named_modules():
-            if isinstance(module, Int8SymmetricLinear):
-                for missing in missing_keys:
-                    if (
-                        (name in missing or name in f"{prefix}.{missing}")
-                        and not missing.endswith(".weight")
-                        and not missing.endswith(".bias")
-                    ):
-                        not_missing_keys.append(missing)
-        return [k for k in missing_keys if k not in not_missing_keys]
-
-    def _process_model_after_weight_loading(self, model, **kwargs):
-        """
-        Post-processing after weights are loaded.
-        """
-        return True
-
-    def is_serializable(self, safe_serialization=None):
-        return True
-
-    @property
-    def is_trainable(self) -> bool:
-        return False
-
-
-# Example usage
-if __name__ == "__main__":
-    model_int8 = AutoModelForCausalLM.from_pretrained(
-        "meta-llama/Llama-3.2-1B", quantization_config=Int8SymmetricConfig(), torch_dtype=torch.float, device_map="cpu"
-    )
-
-    tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")
-    input_text = "once there is"
-    inputs = tokenizer(input_text, return_tensors="pt").to("cpu")
-    output = model_int8.generate(
-        **inputs,
-        max_length=100,
-        num_return_sequences=1,
-        no_repeat_ngram_size=2,
-    )
-    generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
-    print(generated_text)
-
-    # Save and upload to HUB
-    output_model_dir = "Llama-3.2-1B-INT8-CUSTOM"
-    model_int8.save_pretrained(output_model_dir)
-    tokenizer.save_pretrained(output_model_dir)
-    api = HfApi()
-    repo_id = "medmekk/Llama-3.2-1B-INT8-CUSTOM"
-    api.create_repo(repo_id, private=False)
-    api.upload_folder(folder_path=output_model_dir, repo_id=repo_id, repo_type="model")

examples/research_projects/codeparrot/examples/requirements.txt

@@ -1,5 +1,5 @@
 datasets==2.3.2
-transformers==4.48.0
+transformers==4.38.0
 wandb==0.13.1
 evaluate==0.2.2
 scikit-learn==1.5.0

examples/tensorflow/contrastive-image-text/run_clip.py

@@ -51,7 +51,7 @@ from transformers.utils.versions import require_version
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version(
     "datasets>=1.8.0", "To fix: pip install -r examples/tensorflow/contrastive-image-text/requirements.txt"

examples/tensorflow/image-classification/run_image_classification.py

@@ -55,7 +55,7 @@ from transformers.utils.versions import require_version
 logger = logging.getLogger(__name__)
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/image-classification/requirements.txt")
 

examples/tensorflow/multiple-choice/run_swag.py

@@ -50,7 +50,7 @@ from transformers.utils import check_min_version, send_example_telemetry
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = logging.getLogger(__name__)
 

examples/tensorflow/question-answering/run_qa.py

@@ -62,7 +62,7 @@ except (ModuleNotFoundError, ImportError):
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 logger = logging.getLogger(__name__)
 

examples/tensorflow/summarization/run_summarization.py

@@ -53,7 +53,7 @@ from transformers.utils.versions import require_version
 
 # region Checking dependencies
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/summarization/requirements.txt")
 

examples/tensorflow/text-classification/run_glue.py

@@ -47,7 +47,7 @@ from transformers.utils import check_min_version, send_example_telemetry
 
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 task_to_keys = {
     "cola": ("sentence", None),

examples/tensorflow/translation/run_translation.py

@@ -56,7 +56,7 @@ from transformers.utils.versions import require_version
 
 # region Dependencies and constants
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.50.0.dev0")
+check_min_version("4.49.0.dev0")
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/summarization/requirements.txt")
 

i18n/README_zh-hans.md

@@ -83,7 +83,7 @@ checkpoint: 检查点
 
 🤗 Transformers 提供了数以千计的预训练模型，支持 100 多种语言的文本分类、信息抽取、问答、摘要、翻译、文本生成。它的宗旨是让最先进的 NLP 技术人人易用。
 
-🤗 Transformers 提供了便于快速下载和使用的API，让你可以把预训练模型用在给定文本、在你的数据集上微调然后通过 [model hub](https://huggingface.co/models) 与社区共享。同时，每个定义的 Python 模块都是完全独立的，便于修改和快速进行研究实验。
+🤗 Transformers 提供了便于快速下载和使用的API，让你可以把预训练模型用在给定文本、在你的数据集上微调然后通过 [model hub](https://huggingface.co/models) 与社区共享。同时，每个定义的 Python 模块均完全独立，方便修改和快速研究实验。
 
 🤗 Transformers 支持三个最热门的深度学习库： [Jax](https://jax.readthedocs.io/en/latest/), [PyTorch](https://pytorch.org/) 以及 [TensorFlow](https://www.tensorflow.org/) — 并与之无缝整合。你可以直接使用一个框架训练你的模型然后用另一个加载和推理。
 

setup.py

@@ -437,7 +437,7 @@ install_requires = [
 
 setup(
     name="transformers",
-    version="4.50.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="4.49.0.dev0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     author="The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)",
     author_email="transformers@huggingface.co",
     description="State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow",
@@ -473,6 +473,8 @@ setup(
 extras["tests_torch"] = deps_list()
 extras["tests_tf"] = deps_list()
 extras["tests_flax"] = deps_list()
+extras["tests_torch_and_tf"] = deps_list()
+extras["tests_torch_and_flax"] = deps_list()
 extras["tests_hub"] = deps_list()
 extras["tests_pipelines_torch"] = deps_list()
 extras["tests_pipelines_tf"] = deps_list()

src/transformers/__init__.py

@@ -18,7 +18,7 @@
 # to defer the actual importing for when the objects are requested. This way `import transformers` provides the names
 # in the namespace without actually importing anything (and especially none of the backends).
 
-__version__ = "4.50.0.dev0"
+__version__ = "4.49.0.dev0"
 
 from typing import TYPE_CHECKING
 
@@ -496,7 +496,6 @@ _import_structure = {
     "models.gptj": ["GPTJConfig"],
     "models.granite": ["GraniteConfig"],
     "models.granitemoe": ["GraniteMoeConfig"],
-    "models.granitemoeshared": ["GraniteMoeSharedConfig"],
     "models.grounding_dino": [
         "GroundingDinoConfig",
         "GroundingDinoProcessor",
@@ -2540,14 +2539,6 @@ else:
             "GraniteMoePreTrainedModel",
         ]
     )
-    _import_structure["models.granitemoeshared"].extend(
-        [
-            "GraniteMoeSharedForCausalLM",
-            "GraniteMoeSharedModel",
-            "GraniteMoeSharedPreTrainedModel",
-        ]
-    )
-
     _import_structure["models.grounding_dino"].extend(
         [
             "GroundingDinoForObjectDetection",
@@ -5614,7 +5605,6 @@ if TYPE_CHECKING:
     from .models.gptj import GPTJConfig
     from .models.granite import GraniteConfig
     from .models.granitemoe import GraniteMoeConfig
-    from .models.granitemoeshared import GraniteMoeSharedConfig
     from .models.grounding_dino import (
         GroundingDinoConfig,
         GroundingDinoProcessor,
@@ -7489,11 +7479,6 @@ if TYPE_CHECKING:
             GraniteMoeModel,
             GraniteMoePreTrainedModel,
         )
-        from .models.granitemoeshared import (
-            GraniteMoeSharedForCausalLM,
-            GraniteMoeSharedModel,
-            GraniteMoeSharedPreTrainedModel,
-        )
         from .models.grounding_dino import (
             GroundingDinoForObjectDetection,
             GroundingDinoModel,

src/transformers/audio_utils.py

@@ -390,7 +390,6 @@ def spectrogram(
     center: bool = True,
     pad_mode: str = "reflect",
     onesided: bool = True,
-    dither: float = 0.0,
     preemphasis: Optional[float] = None,
     mel_filters: Optional[np.ndarray] = None,
     mel_floor: float = 1e-10,
@@ -461,12 +460,6 @@ def spectrogram(
         onesided (`bool`, *optional*, defaults to `True`):
             If True, only computes the positive frequencies and returns a spectrogram containing `fft_length // 2 + 1`
             frequency bins. If False, also computes the negative frequencies and returns `fft_length` frequency bins.
-        dither (`float`, *optional*, defaults to 0.0):
-            Adds dithering. In other words, adds a small Gaussian noise to each frame.
-            E.g. use 4.0 to add dithering with a normal distribution centered
-            around 0.0 with standard deviation 4.0, 0.0 means no dithering.
-            Dithering has similar effect as `mel_floor`. It reduces the high log_mel_fbank
-            values for signals with hard-zero sections, when VAD cutoff is present in the signal.
         preemphasis (`float`, *optional*)
             Coefficient for a low-pass filter that applies pre-emphasis before the DFT.
         mel_filters (`np.ndarray` of shape `(num_freq_bins, num_mel_filters)`, *optional*):
@@ -547,9 +540,6 @@ def spectrogram(
     for frame_idx in range(num_frames):
         buffer[:frame_length] = waveform[timestep : timestep + frame_length]
 
-        if dither != 0.0:
-            buffer[:frame_length] += dither * np.random.randn(frame_length)
-
         if remove_dc_offset:
             buffer[:frame_length] = buffer[:frame_length] - buffer[:frame_length].mean()
 
@@ -601,7 +591,6 @@ def spectrogram_batch(
     center: bool = True,
     pad_mode: str = "reflect",
     onesided: bool = True,
-    dither: float = 0.0,
     preemphasis: Optional[float] = None,
     mel_filters: Optional[np.ndarray] = None,
     mel_floor: float = 1e-10,
@@ -664,10 +653,6 @@ def spectrogram_batch(
             The padding strategy when `center` is `True`.
         onesided (`bool`, *optional*, defaults to `True`):
             If True, returns a one-sided spectrogram for real input signals.
-        dither (`float`, *optional*, defaults to 0.0):
-            Adds dithering. In other words, adds a small Gaussian noise to each frame.
-            E.g. use 4.0 to add dithering with a normal distribution centered
-            around 0.0 with standard deviation 4.0, 0.0 means no dithering.
         preemphasis (`float`, *optional*):
             Applies a pre-emphasis filter to each frame.
         mel_filters (`np.ndarray`, *optional*):
@@ -756,9 +741,6 @@ def spectrogram_batch(
         timestep = frame_idx * hop_length
         buffer[:, :frame_length] = padded_waveform_batch[:, timestep : timestep + frame_length]
 
-        if dither != 0.0:
-            buffer[:, :frame_length] += dither * np.random.randn(*buffer[:, :frame_length].shape)
-
         if remove_dc_offset:
             buffer[:, :frame_length] -= buffer[:, :frame_length].mean(axis=1, keepdims=True)
 

src/transformers/cache_utils.py

@@ -9,7 +9,12 @@ import torch
 from packaging import version
 
 from .configuration_utils import PretrainedConfig
-from .utils import is_hqq_available, is_optimum_quanto_available, logging
+from .utils import (
+    is_hqq_available,
+    is_optimum_quanto_available,
+    is_torchdynamo_compiling,
+    logging,
+)
 from .utils.deprecation import deprecate_kwarg
 
 
@@ -19,7 +24,7 @@ if is_hqq_available():
 logger = logging.get_logger(__name__)
 
 
-class Cache:
+class Cache(torch.nn.Module):
     """
     Base, abstract class for all caches. The actual data structure is specific to each subclass.
     """
@@ -358,7 +363,8 @@ class DynamicCache(Cache):
         ```
     """
 
-    def __init__(self) -> None:
+    @deprecate_kwarg("num_hidden_layers", version="4.47.0")
+    def __init__(self, num_hidden_layers: Optional[int] = None) -> None:
         super().__init__()
         self._seen_tokens = 0  # Used in `generate` to keep tally of how many tokens the cache has seen
         self.key_cache: List[torch.Tensor] = []
@@ -460,7 +466,10 @@ class DynamicCache(Cache):
         return legacy_cache
 
     @classmethod
-    def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "DynamicCache":
+    @deprecate_kwarg("num_hidden_layers", version="4.47.0")
+    def from_legacy_cache(
+        cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, num_hidden_layers: int = None
+    ) -> "DynamicCache":
         """Converts a cache in the legacy cache format into an equivalent `DynamicCache`. Used for
         backward compatibility."""
         cache = cls()
@@ -486,7 +495,10 @@ class DynamicCache(Cache):
                 self.key_cache[idx] = self.key_cache[idx][..., :max_length, :]
                 self.value_cache[idx] = self.value_cache[idx][..., :max_length, :]
 
-    def batch_split(self, full_batch_size: int, split_size: int) -> List["DynamicCache"]:
+    @deprecate_kwarg("num_hidden_layers", version="4.47.0")
+    def batch_split(
+        self, full_batch_size: int, split_size: int, num_hidden_layers: int = None
+    ) -> List["DynamicCache"]:
         """Split the current instance into a list of `DynamicCache` by the batch size. This will be used by
         `_split_model_inputs()` in `generation.utils`"""
         out = []
@@ -499,7 +511,8 @@ class DynamicCache(Cache):
         return out
 
     @classmethod
-    def from_batch_splits(cls, splits: List["DynamicCache"]) -> "DynamicCache":
+    @deprecate_kwarg("num_hidden_layers", version="4.47.0")
+    def from_batch_splits(cls, splits: List["DynamicCache"], num_hidden_layers: int = None) -> "DynamicCache":
         """This is the opposite of the above `batch_split()` method. This will be used by `stack_model_outputs` in
         `generation.utils`"""
         cache = cls()
@@ -1135,10 +1148,18 @@ class StaticCache(Cache):
                 layer_device = self.device
             new_layer_key_cache = torch.zeros(cache_shape, dtype=self.dtype, device=layer_device)
             new_layer_value_cache = torch.zeros(cache_shape, dtype=self.dtype, device=layer_device)
-            # Note: `mark_static_address` is used to tag the cache as a fixed data pointer,
-            # preventing compiled graph breaks when updating the cache.
-            torch._dynamo.mark_static_address(new_layer_key_cache)
-            torch._dynamo.mark_static_address(new_layer_value_cache)
+            # Notes:
+            # 1. `mark_static_address` is used to tag the cache as an fixed data pointer, preventing cuda graph
+            #     breaks when updating the cache. It can't be used if the cache code is being compiled (but in that case
+            #     it is not needed anyway)
+            # 2. `torch.export()` requires mutations to be registered as buffers.
+            if not is_torchdynamo_compiling():
+                self.register_buffer(f"key_cache_{idx}", torch.zeros(cache_shape, dtype=dtype, device=layer_device))
+                self.register_buffer(f"value_cache_{idx}", torch.zeros(cache_shape, dtype=dtype, device=layer_device))
+                new_layer_key_cache = getattr(self, f"key_cache_{idx}")
+                new_layer_value_cache = getattr(self, f"value_cache_{idx}")
+                torch._dynamo.mark_static_address(new_layer_key_cache)
+                torch._dynamo.mark_static_address(new_layer_value_cache)
             self.key_cache.append(new_layer_key_cache)
             self.value_cache.append(new_layer_value_cache)
 
@@ -1506,7 +1527,10 @@ class EncoderDecoderCache(Cache):
         self.check_dynamic_cache(self.crop.__name__)
         self.self_attention_cache.crop(maximum_length)
 
-    def batch_split(self, full_batch_size: int, split_size: int) -> "List[EncoderDecoderCache]":
+    @deprecate_kwarg("num_hidden_layers", version="4.47.0")
+    def batch_split(
+        self, full_batch_size: int, split_size: int, num_hidden_layers: int = None
+    ) -> "List[EncoderDecoderCache]":
         """Split the current instance into a list of `DynamicCache` by the batch size. This will be used by
         `_split_model_inputs()` in `generation.utils`"""
         self.check_dynamic_cache(self.batch_split.__name__)
@@ -1519,7 +1543,10 @@ class EncoderDecoderCache(Cache):
         return out
 
     @classmethod
-    def from_batch_splits(cls, splits: List["EncoderDecoderCache"]) -> "EncoderDecoderCache":
+    @deprecate_kwarg("num_hidden_layers", version="4.47.0")
+    def from_batch_splits(
+        cls, splits: List["EncoderDecoderCache"], num_hidden_layers: int = None
+    ) -> "EncoderDecoderCache":
         """This is the opposite of the above `batch_split()` method. This will be used by `stack_model_outputs` in
         `generation.utils`"""
         self_attention_cache = DynamicCache()

src/transformers/generation/utils.py

@@ -420,7 +420,6 @@ class GenerationMixin:
             model_inputs[input_ids_key] = input_ids.clone(memory_format=torch.contiguous_format)
 
         # 4. Create missing `position_ids` on the fly
-        encoder_attention_mask = attention_mask if self.config.is_encoder_decoder else None
         attention_mask = (
             kwargs.pop("decoder_attention_mask", None) if self.config.is_encoder_decoder else attention_mask
         )
@@ -491,9 +490,6 @@ class GenerationMixin:
         if attention_mask is not None:
             model_inputs[attention_mask_key] = attention_mask
 
-        if encoder_attention_mask is not None:
-            model_inputs["attention_mask"] = encoder_attention_mask
-
         # 7. Forward ALL kwargs that are uninitialized (e.g. `use_cache`).
         for key, value in kwargs.items():
             if key not in model_inputs:
@@ -4524,7 +4520,7 @@ def _ranking_fast(
     return selected_idx
 
 
-def _split(data, full_batch_size: int, split_size: int = None):
+def _split(data, full_batch_size: int, num_hidden_layers: int, split_size: int = None):
     """
     Takes care of three cases:
     1. data is a tensor: e.g. last_hidden_state, pooler_output etc. split them on the batch_size dim
@@ -4542,7 +4538,7 @@ def _split(data, full_batch_size: int, split_size: int = None):
     elif isinstance(data, DynamicCache) or (
         isinstance(data, EncoderDecoderCache) and isinstance(data.self_attention_cache, DynamicCache)
     ):
-        return data.batch_split(full_batch_size, split_size)
+        return data.batch_split(full_batch_size, split_size, num_hidden_layers)
     elif isinstance(data, tuple):
         # If the elements of the tuple are also tuples (e.g., past_key_values in our earlier example)
         if isinstance(data[0], tuple):
@@ -4595,9 +4591,11 @@ def _split_model_inputs(
     keys_to_ignore = ["cache_position", "encoder_outputs", "logits_to_keep"]
     non_bool_keys = [k for k in keys if not isinstance(model_input[k], bool) and k not in keys_to_ignore]
 
+    num_hidden_layers = config.get_text_config().num_hidden_layers
+
     # we split the tensors and tuples of tensors
     data_split_list = [
-        {k: _split(model_input[k], full_batch_size, split_size)[i] for k in non_bool_keys}
+        {k: _split(model_input[k], full_batch_size, num_hidden_layers, split_size)[i] for k in non_bool_keys}
         for i in range(full_batch_size // split_size)
     ]
     # bool values are the same and replicated for each split
@@ -4634,6 +4632,7 @@ def stack_model_outputs(model_outputs: List[ModelOutput], config: PretrainedConf
 
     # Infer the class from the first object in the list
     model_output_cls = type(model_outputs[0])
+    num_hidden_layers = config.get_text_config().num_hidden_layers
 
     # Ensure all objects are of the same type
     if not all(isinstance(obj, model_output_cls) for obj in model_outputs):
@@ -4650,9 +4649,9 @@ def stack_model_outputs(model_outputs: List[ModelOutput], config: PretrainedConf
             return torch.cat(data, dim=0)
         # New cache format
         elif isinstance(data[0], DynamicCache):
-            return DynamicCache.from_batch_splits(data)
+            return DynamicCache.from_batch_splits(data, num_hidden_layers=num_hidden_layers)
         elif isinstance(data[0], EncoderDecoderCache):
-            return EncoderDecoderCache.from_batch_splits(data)
+            return EncoderDecoderCache.from_batch_splits(data, num_hidden_layers=num_hidden_layers)
         elif isinstance(data[0], tuple):
             # If the elements of the tuple are also tuples (e.g., past_key_values in our earlier example)
             if isinstance(data[0][0], tuple):

src/transformers/integrations/executorch.py

@@ -16,7 +16,10 @@ from ..utils.import_utils import is_torch_available
 
 
 if is_torch_available():
-    from transformers import PreTrainedModel, StaticCache
+    from transformers import (
+        PreTrainedModel,
+        StaticCache,
+    )
     from transformers.pytorch_utils import is_torch_greater_or_equal_than_2_3
 
 
@@ -69,13 +72,9 @@ class TorchExportableModuleWithStaticCache(torch.nn.Module):
             config=self.model.config,
             batch_size=self.model.generation_config.cache_config.batch_size,
             max_cache_len=self.model.generation_config.cache_config.max_cache_len,
-            device=self.model.generation_config.cache_config.device,
             dtype=self.model.dtype,
+            device=self.model.generation_config.cache_config.device,
         )
-        for i in range(len(self.static_cache.key_cache)):
-            self.register_buffer(f"key_cache_{i}", self.static_cache.key_cache[i], persistent=False)
-            self.register_buffer(f"value_cache_{i}", self.static_cache.value_cache[i], persistent=False)
-
         self.is_causal = any("CausalLM" in arch for arch in self.model.config.architectures)
         if self.is_causal:
             causal_mask = torch.tril(
@@ -110,15 +109,12 @@ class TorchExportableModuleWithStaticCache(torch.nn.Module):
         """
         _, seqlen = input_ids.shape
         attn_mask = self.mask[cache_position, :seqlen] if self.is_causal else None
-        position_ids = cache_position.unsqueeze(0)
-        past_key_values = self.static_cache
-
         outs = self.model(
             input_ids=input_ids,
             attention_mask=attn_mask,
-            position_ids=position_ids,
+            position_ids=cache_position.unsqueeze(0),
             cache_position=cache_position,
-            past_key_values=past_key_values,
+            past_key_values=self.static_cache,
             use_cache=True,
         )
         return outs.logits
@@ -147,7 +143,7 @@ class TorchExportableModuleWithStaticCache(torch.nn.Module):
         prompt_token_len = prompt_token_ids.shape[-1]
         max_generation_length = prompt_token_len + max_new_tokens
         for buffer_name, buffer in exported_program.named_buffers():
-            if buffer_name.startswith("key_cache"):
+            if buffer_name.startswith("static_cache.key_cache"):
                 max_cache_len = buffer.shape[2]
                 max_generation_length = min(max_generation_length, max_cache_len)
                 break

src/transformers/integrations/higgs.py

@@ -28,12 +28,15 @@ if is_torch_available():
 
 
 if is_flute_available():
-    from flute.integrations.higgs import prepare_data_transposed
-    from flute.tune import TuneMetaData, qgemm_v2
+    import flute.utils
 
 if is_hadamard_available():
     from fast_hadamard_transform import hadamard_transform
 
+if is_flute_available():
+    import flute.utils
+    from flute.integrations.higgs import prepare_data_transposed
+
 
 def pad_to_block(tensor, dims, had_block_size, value=0):
     pad_dims = [0 for _ in range(2 * len(tensor.shape))]
@@ -461,14 +464,14 @@ def quantize_with_higgs(weight, bits: int = 4, p: int = 2, group_size: int = 256
 
     # Quantize
     codes = torch.empty(weight.shape[:-1], device=device, dtype=torch.uint8)
-    for i in range(0, weight.shape[0], 16):
-        codes[i : i + 16] = torch.argmax(2 * weight[i : i + 16] @ grid.T - grid_norm_2, dim=-1).to(torch.uint8)
+    for i in range(0, weight.shape[0], 64):
+        codes[i : i + 64] = torch.argmax(2 * weight[i : i + 64] @ grid.T - grid_norm_2, dim=-1).to(torch.uint8)
     del weight
 
     codes = codes.reshape(codes.shape[0], -1)
     scales = scales / sqrt(hadamard_size)
 
-    weight, scales, tables, tables2, tune_metadata = prepare_data_transposed(
+    weight, scales, tables, tables2 = prepare_data_transposed(
         codes,
         torch.repeat_interleave(scales.to(dtype), hadamard_size // group_size, dim=1),
         grid.to(dtype),
@@ -477,7 +480,6 @@ def quantize_with_higgs(weight, bits: int = 4, p: int = 2, group_size: int = 256
         vector_size=p,
         dtype=dtype,
         device=device,
-        check_correctness=False,
     )
 
     return {
@@ -485,7 +487,6 @@ def quantize_with_higgs(weight, bits: int = 4, p: int = 2, group_size: int = 256
         "scales": scales,
         "tables": tables,
         "tables2": tables2.view(dtype=torch.float16),
-        "tune_metadata": tune_metadata,
     }
 
 
@@ -507,6 +508,7 @@ class HiggsLinear(torch.nn.Module):
         self.num_bits = num_bits
         self.group_size = group_size
         self.hadamard_size = hadamard_size
+        self.num_sms_packed = nn.Parameter(torch.tensor(-1, dtype=torch.int32, device=device), requires_grad=False)
 
         assert in_features % group_size == 0
         assert num_bits in [2, 3, 4]
@@ -529,7 +531,6 @@ class HiggsLinear(torch.nn.Module):
             self.register_parameter("bias", None)
 
         self.workspace = None  # must be set externally to be reused among layers
-        self.tune_metadata: TuneMetaData = None  # must be set externally because architecture dependent
 
     def forward(self, x):
         x = pad_to_block(x, [-1], self.hadamard_size)
@@ -537,15 +538,16 @@ class HiggsLinear(torch.nn.Module):
         if self.workspace is None:
             raise Exception("Workspace must be set before calling forward")
 
-        return qgemm_v2(
+        return flute.qgemm_hadamard(
             x,
             self.weight,
             self.scales,
             self.tables,
             self.tables2.view(dtype=torch.float32),
             self.workspace,
-            self.tune_metadata,
-            hadamard_size=self.hadamard_size,
+            self.num_bits,
+            self.group_size,
+            self.hadamard_size,
         )
 
 

src/transformers/modeling_utils.py

@@ -787,7 +787,6 @@ def _load_state_dict_into_meta_model(
     keep_in_fp32_modules=None,
     unexpected_keys=None,  # passing `unexpected` for cleanup from quantization items
     pretrained_model_name_or_path=None,  # for flagging the user when the model contains renamed keys
-    device_mesh=None,
 ):
     """
     This is somewhat similar to `_load_state_dict_into_model`, but deals with a model that has some or all of its
@@ -797,8 +796,6 @@ def _load_state_dict_into_meta_model(
     `start_prefix` is used for models which insert their name into model keys, e.g. `bert` in
     `bert.pooler.dense.weight`
 
-    It also initialize tensor parallelism for each module if needed.
-
     """
 
     # XXX: remaining features to implement to be fully compatible with _load_state_dict_into_model
@@ -812,12 +809,6 @@ def _load_state_dict_into_meta_model(
 
     is_torch_e4m3fn_available = hasattr(torch, "float8_e4m3fn")
 
-    # we need this later to initialize tensor parallelism
-    if device_mesh is not None:
-        full_tp_plan = model.config.base_model_tp_plan
-        for submodule in model.modules():
-            full_tp_plan.update(getattr(submodule, "_tp_plan", {}))
-
     for param_name, param in state_dict.items():
         if param_name not in expected_keys:
             continue
@@ -921,37 +912,6 @@ def _load_state_dict_into_meta_model(
                 setattr(module, tensor_name, value)
             # TODO: consider removing used param_parts from state_dict before return
 
-        # In this case, let's parallelize the modules!
-        if device_mesh is not None:
-            # Immediate parent
-            split_parent_module_name = param_name.split(".")[:-1]
-            parent_module_name = ".".join(split_parent_module_name)
-            parent_module = model
-            for name in split_parent_module_name:
-                parent_module = getattr(parent_module, name)
-
-            # Check if we are part of the tp_plan
-            current_module_plan = None
-            for param, plan in full_tp_plan.items():
-                # "*" are a placeholder for layer indices, so we replace them by "[0-9]+" in the regex pattern
-                pattern = param.replace("*", "[0-9]+")
-                if re.search(pattern, parent_module_name):
-                    current_module_plan = plan
-                    break
-
-            # We can only apply the tp_plan after all parameters of the current module have been correctly initialized (e.g.
-            # if we have bias, we need both `weights` and `bias` of a nn.Linear to be initialized)
-            process_device = list(device_map.values())[0]
-            all_module_parameters_initialized = all(
-                m.device == process_device for m in parent_module.parameters(recurse=False)
-            ) and all(m.device == process_device for m in parent_module.buffers(recurse=False))
-            if current_module_plan is not None and all_module_parameters_initialized:
-                torch.distributed.tensor.parallel.parallelize_module(
-                    parent_module,
-                    device_mesh=device_mesh,
-                    parallelize_plan=translate_to_torch_parallel_style(current_module_plan),
-                )
-
     return error_msgs, offload_index, state_dict_index
 
 
@@ -3529,11 +3489,12 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             )
 
         # We need to correctly dispatch the model on the current process device. The easiest way for this is to use a simple
-        # `device_map` pointing to the correct device
-        device_mesh = None
+        # `device_map` pointing to the correct device. If we don't, torch will use the default device (index 0) for all
+        # childs processes at parallelization time, resulting in excessive memory usage on device 0 and OOMs.
+        # And temporarily setting the default device to current process rank result in the following error
+        # `torch.distributed.DistBackendError: Attempt to perform collective on tensor not on device passed to init_process_group`
+        tp_device = None
         if tp_plan is not None:
-            if not is_torch_greater_or_equal("2.5"):
-                raise EnvironmentError("tensor parallel is only supported for `torch>=2.5`.")
             if not torch.distributed.is_initialized():
                 raise ValueError("Tensor Parallel requires torch.distributed to be initialized first.")
 
@@ -3545,10 +3506,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             # This is the easiest way to dispatch to the current process device
             device_map = tp_device
 
-            # Assuming sharding the model onto the world
-            world_size = torch.distributed.get_world_size()
-            device_mesh = torch.distributed.init_device_mesh(tp_device.type, (world_size,))
-
         if is_fsdp_enabled():
             low_cpu_mem_usage = True
 
@@ -3643,7 +3600,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             if low_cpu_mem_usage is None:
                 low_cpu_mem_usage = True
             elif not low_cpu_mem_usage:
-                raise ValueError("Passing along a `device_map` or a `tp_plan` requires `low_cpu_mem_usage=True`")
+                raise ValueError("Passing along a `device_map` requires `low_cpu_mem_usage=True`")
 
         if low_cpu_mem_usage:
             if is_deepspeed_zero3_enabled():
@@ -3652,7 +3609,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 )
             elif not is_accelerate_available():
                 raise ImportError(
-                    f"Using `low_cpu_mem_usage=True`, a `device_map` or a `tp_plan` requires Accelerate: `pip install 'accelerate>={ACCELERATE_MIN_VERSION}'`"
+                    f"Using `low_cpu_mem_usage=True` or a `device_map` requires Accelerate: `pip install 'accelerate>={ACCELERATE_MIN_VERSION}'`"
                 )
 
         # handling bnb config from kwargs, remove after `load_in_{4/8}bit` deprecation.
@@ -3749,10 +3706,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             device_map = hf_quantizer.update_device_map(device_map)
 
             # In order to ensure popular quantization methods are supported. Can be disable with `disable_telemetry`
-            if hasattr(hf_quantizer.quantization_config.quant_method, "value"):
-                user_agent["quant"] = hf_quantizer.quantization_config.quant_method.value
-            else:
-                user_agent["quant"] = hf_quantizer.quantization_config.quant_method
+            user_agent["quant"] = hf_quantizer.quantization_config.quant_method.value
+
             # Force-set to `True` for more mem efficiency
             if low_cpu_mem_usage is None:
                 low_cpu_mem_usage = True
@@ -4229,9 +4184,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             # Let's make sure we don't run the init function of buffer modules
             model = cls(config, *model_args, **model_kwargs)
 
-        if device_mesh is not None and not model.supports_tp_plan:
-            raise NotImplementedError("This model does not have a tensor parallel plan.")
-
         # make sure we use the model's config since the __init__ call might have copied it
         config = model.config
 
@@ -4315,12 +4267,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             # check if we don't have tied param in different devices
             check_tied_parameters_on_same_device(tied_params, device_map)
 
-        if gguf_path and device_map is not None and "disk" in device_map.values():
-            raise RuntimeError(
-                "One or more modules is configured to be mapped to disk. Disk offload is not supported for models "
-                "loaded from GGUF files."
-            )
-
         if from_tf:
             if resolved_archive_file.endswith(".index"):
                 # Load from a TensorFlow 1.X checkpoint - provided by original authors
@@ -4382,7 +4328,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 keep_in_fp32_modules=keep_in_fp32_modules,
                 gguf_path=gguf_path,
                 weights_only=weights_only,
-                device_mesh=device_mesh,
             )
 
         # make sure token embedding weights are still tied if needed
@@ -4417,9 +4362,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 )
                 pass
 
-        # Dispatch model with hooks on all devices if necessary (not needed with a tp_plan, so we skip it as it slightly
-        # harm performances)
-        if device_map is not None and device_mesh is None:
+        # Dispatch model with hooks on all devices if necessary
+        if device_map is not None:
             device_map_kwargs = {
                 "device_map": device_map,
                 "offload_dir": offload_folder,
@@ -4446,13 +4390,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             if not is_fsdp_enabled() and not is_deepspeed_zero3_enabled():
                 dispatch_model(model, **device_map_kwargs)
 
-        # This is needed for the RotaryEmbedding, which was not initialized on the correct device as it is
-        # not part of the state_dict (persistent=False)
-        if device_mesh is not None:
-            for buffer in model.buffers():
-                if buffer.device != tp_device:
-                    buffer.data = buffer.to(tp_device)
-
         if hf_quantizer is not None:
             hf_quantizer.postprocess_model(model, config=config)
             model.hf_quantizer = hf_quantizer
@@ -4475,6 +4412,16 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 }
             return model, loading_info
 
+        if tp_plan is not None:
+            assert tp_device is not None, "tp_device not set!"
+            if not model.supports_tp_plan:
+                raise NotImplementedError("This model does not have a tensor parallel plan.")
+            # Assuming sharding the model onto the world
+            world_size = torch.distributed.get_world_size()
+            device_mesh = torch.distributed.init_device_mesh(tp_device.type, (world_size,))
+            # Apply Tensor Parallelism
+            model.tensor_parallel(device_mesh)
+
         return model
 
     @staticmethod
@@ -4568,7 +4515,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         keep_in_fp32_modules=None,
         gguf_path=None,
         weights_only=True,
-        device_mesh=None,
     ):
         is_safetensors = False
         is_quantized = hf_quantizer is not None
@@ -4579,7 +4525,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             archive_file = (
                 resolved_archive_file[0] if isinstance(resolved_archive_file, (list, tuple)) else resolved_archive_file
             )
-            is_safetensors = archive_file is not None and archive_file.endswith(".safetensors")
+            is_safetensors = archive_file.endswith(".safetensors")
             if offload_folder is None and not is_safetensors:
                 raise ValueError(
                     "The current `device_map` had weights offloaded to the disk. Please provide an `offload_folder`"
@@ -4868,7 +4814,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                     is_safetensors=is_safetensors,
                     keep_in_fp32_modules=keep_in_fp32_modules,
                     unexpected_keys=unexpected_keys,
-                    device_mesh=device_mesh,
                 )
             else:
                 # Sharded checkpoint or whole but low_cpu_mem_usage==True
@@ -4958,7 +4903,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                             is_safetensors=is_safetensors,
                             keep_in_fp32_modules=keep_in_fp32_modules,
                             unexpected_keys=unexpected_keys,
-                            device_mesh=device_mesh,
                         )
                         error_msgs += new_error_msgs
                 else:
@@ -5236,12 +5180,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
 
     def tensor_parallel(self, device_mesh):
         """
-        Tensor parallelize the model across the given device mesh. This function is a helper to be called after the model
-        was already loaded in memory, note however that this means that each process will first initialize the whole model,
-        then parallelize it accross devices. Thus there is a huge waste of GPU memory, and this can lead to OOM at loading time.
-
-        Calling `from_pretrained(..., tp_plan="auto")` is prefered, and will parallelize module-by-module during initialization,
-        so that the expected per-device memory spike at loading time is not larger than the final model size on each device.
+        Tensor parallelize the model across the given device mesh.
 
         Args:
             device_mesh (`torch.distributed.DeviceMesh`):

src/transformers/models/__init__.py

@@ -118,7 +118,6 @@ from . import (
     gptj,
     granite,
     granitemoe,
-    granitemoeshared,
     grounding_dino,
     groupvit,
     helium,

src/transformers/models/audio_spectrogram_transformer/feature_extraction_audio_spectrogram_transformer.py

@@ -193,7 +193,7 @@ class ASTFeatureExtractor(SequenceFeatureExtractor):
                 )
         else:
             logger.warning(
-                f"It is strongly recommended to pass the `sampling_rate` argument to `{self.__class__.__name__}()`. "
+                "It is strongly recommended to pass the `sampling_rate` argument to this function. "
                 "Failing to do so can result in silent errors that might be hard to debug."
             )
 

src/transformers/models/auto/configuration_auto.py

@@ -137,7 +137,6 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("gptsan-japanese", "GPTSanJapaneseConfig"),
         ("granite", "GraniteConfig"),
         ("granitemoe", "GraniteMoeConfig"),
-        ("granitemoeshared", "GraniteMoeSharedConfig"),
         ("granitevision", "LlavaNextConfig"),
         ("graphormer", "GraphormerConfig"),
         ("grounding-dino", "GroundingDinoConfig"),
@@ -468,7 +467,6 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("gptsan-japanese", "GPTSAN-japanese"),
         ("granite", "Granite"),
         ("granitemoe", "GraniteMoeMoe"),
-        ("granitemoeshared", "GraniteMoeSharedMoe"),
         ("granitevision", "LLaVA-NeXT"),
         ("graphormer", "Graphormer"),
         ("grounding-dino", "Grounding DINO"),

src/transformers/models/auto/modeling_auto.py

@@ -132,7 +132,6 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("gptsan-japanese", "GPTSanJapaneseForConditionalGeneration"),
         ("granite", "GraniteModel"),
         ("granitemoe", "GraniteMoeModel"),
-        ("granitemoeshared", "GraniteMoeSharedModel"),
         ("graphormer", "GraphormerModel"),
         ("grounding-dino", "GroundingDinoModel"),
         ("groupvit", "GroupViTModel"),
@@ -527,7 +526,6 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("gptj", "GPTJForCausalLM"),
         ("granite", "GraniteForCausalLM"),
         ("granitemoe", "GraniteMoeForCausalLM"),
-        ("granitemoeshared", "GraniteMoeSharedForCausalLM"),
         ("helium", "HeliumForCausalLM"),
         ("jamba", "JambaForCausalLM"),
         ("jetmoe", "JetMoeForCausalLM"),

src/transformers/models/blip_2/modeling_blip_2.py

@@ -2016,9 +2016,6 @@ class Blip2VisionModelWithProjection(Blip2PreTrainedModel):
 class Blip2ForConditionalGeneration(Blip2PreTrainedModel, GenerationMixin):
     config_class = Blip2Config
     main_input_name = "pixel_values"
-    _supports_cache_class = True
-    _supports_static_cache = True
-    _supports_quantized_cache = False  # not all LM bacbones support (e.g. T5)
 
     def __init__(self, config: Blip2Config):
         super().__init__(config)

src/transformers/models/chameleon/modeling_chameleon.py

@@ -1284,13 +1284,13 @@ class ChameleonModel(ChameleonPreTrainedModel):
 
         if pixel_values is not None:
             image_tokens = self.get_image_tokens(pixel_values)
-            special_image_mask = input_ids == self.vocabulary_mapping.image_token_id
-            if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_tokens.numel():
-                n_image_tokens_in_text = (input_ids == self.vocabulary_mapping.image_token_id).sum()
-                n_image_features = image_tokens.shape[0] * image_tokens.shape[1]
+            n_image_tokens_in_text = (input_ids == self.vocabulary_mapping.image_token_id).sum().item()
+            n_image_features = image_tokens.shape[0] * image_tokens.shape[1]
+            if n_image_tokens_in_text != n_image_features:
                 raise ValueError(
                     f"Image features and image tokens do not match: tokens: {n_image_tokens_in_text}, features {n_image_features}"
                 )
+            special_image_mask = input_ids == self.vocabulary_mapping.image_token_id
             image_tokens = image_tokens.to(input_ids.device, input_ids.dtype)
             input_ids = input_ids.masked_scatter(special_image_mask, image_tokens)
 

src/transformers/models/clap/feature_extraction_clap.py

@@ -308,7 +308,7 @@ class ClapFeatureExtractor(SequenceFeatureExtractor):
                 )
         else:
             logger.warning(
-                f"It is strongly recommended to pass the `sampling_rate` argument to `{self.__class__.__name__}()`. "
+                "It is strongly recommended to pass the `sampling_rate` argument to this function. "
                 "Failing to do so can result in silent errors that might be hard to debug."
             )