draft candidate generator

.circleci/create_circleci_config.py

@@ -29,7 +29,6 @@ COMMON_ENV_VARIABLES = {
     "RUN_PIPELINE_TESTS": False,
     # will be adjust in `CircleCIJob.to_dict`.
     "RUN_FLAKY": True,
-    "DISABLE_SAFETENSORS_CONVERSION": True,
 }
 # 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, "vvv": None, "rsfE":None}
@@ -186,7 +185,6 @@ class CircleCIJob:
             # During the CircleCI docker images build time, we might already (or not) download the data.
             # If it's done already, the files are inside the directory `/test_data/`.
             {"run": {"name": "fetch hub objects before pytest", "command": "cp -r /test_data/* . 2>/dev/null || true; python3 utils/fetch_hub_objects_for_ci.py"}},
-            {"run": {"name": "download and unzip hub cache", "command": 'curl -L -o huggingface-cache.tar.gz https://huggingface.co/datasets/hf-internal-testing/hf_hub_cache/resolve/main/huggingface-cache.tar.gz && apt-get install pigz && tar --use-compress-program="pigz -d -p 8" -xf huggingface-cache.tar.gz && mv -n hub/* /root/.cache/huggingface/hub/ && ls -la /root/.cache/huggingface/hub/'}},
             {"run": {
                 "name": "Run tests",
                 "command": f"({timeout_cmd} python3 -m pytest {marker_cmd} -n {self.pytest_num_workers} {junit_flags} {repeat_on_failure_flags} {' '.join(pytest_flags)} $(cat splitted_tests.txt) | tee tests_output.txt)"}

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -48,19 +48,19 @@ body:
           - continuous batching: @remi-or @ArthurZucker @McPatate
           - pipelines: @Rocketknight1
           - tokenizers: @ArthurZucker and @itazap
-          - trainer: @SunMarc
+          - trainer: @zach-huggingface @SunMarc
           - attention: @vasqu @ArthurZucker @CyrilVallez
           - model loading (from pretrained, etc): @CyrilVallez
-          - distributed: @3outeille @ArthurZucker
+          - distributed: @3outeille @ArthurZucker @S1ro1
           - CIs: @ydshieh
 
         Integrations:
 
+          - deepspeed: HF Trainer/Accelerate: @SunMarc @zach-huggingface
           - ray/raytune: @richardliaw, @amogkam
           - Big Model Inference: @SunMarc
-          - quantization: @SunMarc @MekkCyber
+          - quantization (bitsandbytes, autogpt): @SunMarc @MekkCyber
           - kernels: @MekkCyber @drbh
-          - peft: @BenjaminBossan @githubnemo
         
         Devices/Backends:
         

.github/PULL_REQUEST_TEMPLATE.md

@@ -39,40 +39,41 @@ members/contributors who may be interested in your PR.
 
 Models:
 
-- text models: @ArthurZucker @Cyrilvallez
-- vision models: @yonigozlan @molbap
-- audio models: @eustlb @ebezzam @vasqu
-- multimodal models: @zucchini-nlp
+- text models: @ArthurZucker
+- vision models: @amyeroberts, @qubvel
+- speech models: @eustlb
 - graph models: @clefourrier
 
 Library:
 
+- flax: @gante and @Rocketknight1
 - generate: @zucchini-nlp (visual-language models) or @gante (all others)
-- continuous batching: @remi-or @ArthurZucker @McPatate
 - pipelines: @Rocketknight1
-- tokenizers: @ArthurZucker and @itazap
-- trainer: @SunMarc
-- attention: @vasqu @ArthurZucker @CyrilVallez
-- model loading (from pretrained, etc): @CyrilVallez
-- distributed: @3outeille @ArthurZucker
-- CIs: @ydshieh
+- tensorflow: @gante and @Rocketknight1
+- tokenizers: @ArthurZucker
+- trainer: @zach-huggingface, @SunMarc and @qgallouedec
+- chat templates: @Rocketknight1
 
 Integrations:
 
+- deepspeed: HF Trainer/Accelerate: @SunMarc @zach-huggingface
 - ray/raytune: @richardliaw, @amogkam
 - Big Model Inference: @SunMarc
-- quantization: @SunMarc @MekkCyber
-- kernels: @MekkCyber @drbh
-- peft: @BenjaminBossan @githubnemo
-
-Devices/Backends:
-
-- AMD ROCm: @ivarflakstad
-- Intel XPU: @IlyasMoutawwakil
-- Ascend NPU: @ivarflakstad 
+- quantization (bitsandbytes, autogpt): @SunMarc @MekkCyber
 
 Documentation: @stevhliu
 
-Research projects are not maintained and should be taken as is.
+HF projects:
+
+- accelerate: [different repo](https://github.com/huggingface/accelerate)
+- datasets: [different repo](https://github.com/huggingface/datasets)
+- diffusers: [different repo](https://github.com/huggingface/diffusers)
+- rust tokenizers: [different repo](https://github.com/huggingface/tokenizers)
+
+Maintained examples (not research project or legacy):
+
+- Flax: @Rocketknight1
+- PyTorch: See Models above and tag the person corresponding to the modality of the example.
+- TensorFlow: @Rocketknight1
 
  -->

.github/scripts/codeowners_for_review_action

@@ -7,8 +7,8 @@ docs/ @stevhliu
 /docker/ @ydshieh @ArthurZucker
 
 # More high-level globs catch cases when specific rules later don't apply
-/src/transformers/models/*/processing* @molbap @yonigozlan
-/src/transformers/models/*/image_processing* @yonigozlan
+/src/transformers/models/*/processing* @molbap @yonigozlan @qubvel
+/src/transformers/models/*/image_processing* @qubvel
 /src/transformers/models/*/image_processing_*_fast* @yonigozlan
 
 # Owners of subsections of the library
@@ -22,6 +22,7 @@ tests/generation/ @gante
 /src/transformers/models/auto/ @ArthurZucker
 /src/transformers/utils/ @ArthurZucker @Rocketknight1
 /src/transformers/loss/ @ArthurZucker
+/src/transformers/onnx/ @michaelbenayoun
 
 # Specific files come after the sections/globs, so they take priority
 /.circleci/config.yml @ArthurZucker @ydshieh
@@ -185,65 +186,65 @@ trainer_utils.py @zach-huggingface @SunMarc
 /src/transformers/models/zamba/mod*_zamba* @ArthurZucker
 
 # Vision models
-/src/transformers/models/beit/mod*_beit* @yonigozlan @molbap
-/src/transformers/models/bit/mod*_bit* @yonigozlan @molbap
-/src/transformers/models/conditional_detr/mod*_conditional_detr* @yonigozlan @molbap
-/src/transformers/models/convnext/mod*_convnext* @yonigozlan @molbap
-/src/transformers/models/convnextv2/mod*_convnextv2* @yonigozlan @molbap
-/src/transformers/models/cvt/mod*_cvt* @yonigozlan @molbap
-/src/transformers/models/deformable_detr/mod*_deformable_detr* @yonigozlan @molbap
-/src/transformers/models/deit/mod*_deit* @yonigozlan @molbap
-/src/transformers/models/depth_anything/mod*_depth_anything* @yonigozlan @molbap
-/src/transformers/models/depth_anything_v2/mod*_depth_anything_v2* @yonigozlan @molbap
-/src/transformers/models/deta/mod*_deta* @yonigozlan @molbap
-/src/transformers/models/detr/mod*_detr* @yonigozlan @molbap
-/src/transformers/models/dinat/mod*_dinat* @yonigozlan @molbap
-/src/transformers/models/dinov2/mod*_dinov2* @yonigozlan @molbap
-/src/transformers/models/dinov2_with_registers/mod*_dinov2_with_registers* @yonigozlan @molbap
-/src/transformers/models/dit/mod*_dit* @yonigozlan @molbap
-/src/transformers/models/dpt/mod*_dpt* @yonigozlan @molbap
-/src/transformers/models/efficientformer/mod*_efficientformer* @yonigozlan @molbap
-/src/transformers/models/efficientnet/mod*_efficientnet* @yonigozlan @molbap
-/src/transformers/models/focalnet/mod*_focalnet* @yonigozlan @molbap
-/src/transformers/models/glpn/mod*_glpn* @yonigozlan @molbap
-/src/transformers/models/hiera/mod*_hiera* @yonigozlan @molbap
-/src/transformers/models/ijepa/mod*_ijepa* @yonigozlan @molbap
-/src/transformers/models/imagegpt/mod*_imagegpt* @yonigozlan @molbap
-/src/transformers/models/levit/mod*_levit* @yonigozlan @molbap
-/src/transformers/models/mask2former/mod*_mask2former* @yonigozlan @molbap
-/src/transformers/models/maskformer/mod*_maskformer* @yonigozlan @molbap
-/src/transformers/models/mobilenet_v1/mod*_mobilenet_v1* @yonigozlan @molbap
-/src/transformers/models/mobilenet_v2/mod*_mobilenet_v2* @yonigozlan @molbap
-/src/transformers/models/mobilevit/mod*_mobilevit* @yonigozlan @molbap
-/src/transformers/models/mobilevitv2/mod*_mobilevitv2* @yonigozlan @molbap
-/src/transformers/models/nat/mod*_nat* @yonigozlan @molbap
-/src/transformers/models/poolformer/mod*_poolformer* @yonigozlan @molbap
-/src/transformers/models/pvt/mod*_pvt* @yonigozlan @molbap
-/src/transformers/models/pvt_v2/mod*_pvt_v2* @yonigozlan @molbap
-/src/transformers/models/regnet/mod*_regnet* @yonigozlan @molbap
-/src/transformers/models/resnet/mod*_resnet* @yonigozlan @molbap
-/src/transformers/models/rt_detr/mod*_rt_detr* @yonigozlan @molbap
-/src/transformers/models/segformer/mod*_segformer* @yonigozlan @molbap
-/src/transformers/models/seggpt/mod*_seggpt* @yonigozlan @molbap
-/src/transformers/models/superpoint/mod*_superpoint* @yonigozlan @molbap
-/src/transformers/models/swiftformer/mod*_swiftformer* @yonigozlan @molbap
-/src/transformers/models/swin/mod*_swin* @yonigozlan @molbap
-/src/transformers/models/swinv2/mod*_swinv2* @yonigozlan @molbap
-/src/transformers/models/swin2sr/mod*_swin2sr* @yonigozlan @molbap
-/src/transformers/models/table_transformer/mod*_table_transformer* @yonigozlan @molbap
-/src/transformers/models/textnet/mod*_textnet* @yonigozlan @molbap
-/src/transformers/models/timm_wrapper/mod*_timm_wrapper* @yonigozlan @molbap
-/src/transformers/models/upernet/mod*_upernet* @yonigozlan @molbap
-/src/transformers/models/van/mod*_van* @yonigozlan @molbap
-/src/transformers/models/vit/mod*_vit* @yonigozlan @molbap
-/src/transformers/models/vit_hybrid/mod*_vit_hybrid* @yonigozlan @molbap
-/src/transformers/models/vitdet/mod*_vitdet* @yonigozlan @molbap
-/src/transformers/models/vit_mae/mod*_vit_mae* @yonigozlan @molbap
-/src/transformers/models/vitmatte/mod*_vitmatte* @yonigozlan @molbap
-/src/transformers/models/vit_msn/mod*_vit_msn* @yonigozlan @molbap
-/src/transformers/models/vitpose/mod*_vitpose* @yonigozlan @molbap
-/src/transformers/models/yolos/mod*_yolos* @yonigozlan @molbap
-/src/transformers/models/zoedepth/mod*_zoedepth* @yonigozlan @molbap
+/src/transformers/models/beit/mod*_beit* @amyeroberts @qubvel
+/src/transformers/models/bit/mod*_bit* @amyeroberts @qubvel
+/src/transformers/models/conditional_detr/mod*_conditional_detr* @amyeroberts @qubvel
+/src/transformers/models/convnext/mod*_convnext* @amyeroberts @qubvel
+/src/transformers/models/convnextv2/mod*_convnextv2* @amyeroberts @qubvel
+/src/transformers/models/cvt/mod*_cvt* @amyeroberts @qubvel
+/src/transformers/models/deformable_detr/mod*_deformable_detr* @amyeroberts @qubvel
+/src/transformers/models/deit/mod*_deit* @amyeroberts @qubvel
+/src/transformers/models/depth_anything/mod*_depth_anything* @amyeroberts @qubvel
+/src/transformers/models/depth_anything_v2/mod*_depth_anything_v2* @amyeroberts @qubvel
+/src/transformers/models/deta/mod*_deta* @amyeroberts @qubvel
+/src/transformers/models/detr/mod*_detr* @amyeroberts @qubvel
+/src/transformers/models/dinat/mod*_dinat* @amyeroberts @qubvel
+/src/transformers/models/dinov2/mod*_dinov2* @amyeroberts @qubvel
+/src/transformers/models/dinov2_with_registers/mod*_dinov2_with_registers* @amyeroberts @qubvel
+/src/transformers/models/dit/mod*_dit* @amyeroberts @qubvel
+/src/transformers/models/dpt/mod*_dpt* @amyeroberts @qubvel
+/src/transformers/models/efficientformer/mod*_efficientformer* @amyeroberts @qubvel
+/src/transformers/models/efficientnet/mod*_efficientnet* @amyeroberts @qubvel
+/src/transformers/models/focalnet/mod*_focalnet* @amyeroberts @qubvel
+/src/transformers/models/glpn/mod*_glpn* @amyeroberts @qubvel
+/src/transformers/models/hiera/mod*_hiera* @amyeroberts @qubvel
+/src/transformers/models/ijepa/mod*_ijepa* @amyeroberts @qubvel
+/src/transformers/models/imagegpt/mod*_imagegpt* @amyeroberts @qubvel
+/src/transformers/models/levit/mod*_levit* @amyeroberts @qubvel
+/src/transformers/models/mask2former/mod*_mask2former* @amyeroberts @qubvel
+/src/transformers/models/maskformer/mod*_maskformer* @amyeroberts @qubvel
+/src/transformers/models/mobilenet_v1/mod*_mobilenet_v1* @amyeroberts @qubvel
+/src/transformers/models/mobilenet_v2/mod*_mobilenet_v2* @amyeroberts @qubvel
+/src/transformers/models/mobilevit/mod*_mobilevit* @amyeroberts @qubvel
+/src/transformers/models/mobilevitv2/mod*_mobilevitv2* @amyeroberts @qubvel
+/src/transformers/models/nat/mod*_nat* @amyeroberts @qubvel
+/src/transformers/models/poolformer/mod*_poolformer* @amyeroberts @qubvel
+/src/transformers/models/pvt/mod*_pvt* @amyeroberts @qubvel
+/src/transformers/models/pvt_v2/mod*_pvt_v2* @amyeroberts @qubvel
+/src/transformers/models/regnet/mod*_regnet* @amyeroberts @qubvel
+/src/transformers/models/resnet/mod*_resnet* @amyeroberts @qubvel
+/src/transformers/models/rt_detr/mod*_rt_detr* @amyeroberts @qubvel
+/src/transformers/models/segformer/mod*_segformer* @amyeroberts @qubvel
+/src/transformers/models/seggpt/mod*_seggpt* @amyeroberts @qubvel
+/src/transformers/models/superpoint/mod*_superpoint* @amyeroberts @qubvel
+/src/transformers/models/swiftformer/mod*_swiftformer* @amyeroberts @qubvel
+/src/transformers/models/swin/mod*_swin* @amyeroberts @qubvel
+/src/transformers/models/swinv2/mod*_swinv2* @amyeroberts @qubvel
+/src/transformers/models/swin2sr/mod*_swin2sr* @amyeroberts @qubvel
+/src/transformers/models/table_transformer/mod*_table_transformer* @amyeroberts @qubvel
+/src/transformers/models/textnet/mod*_textnet* @amyeroberts @qubvel
+/src/transformers/models/timm_wrapper/mod*_timm_wrapper* @amyeroberts @qubvel
+/src/transformers/models/upernet/mod*_upernet* @amyeroberts @qubvel
+/src/transformers/models/van/mod*_van* @amyeroberts @qubvel
+/src/transformers/models/vit/mod*_vit* @amyeroberts @qubvel
+/src/transformers/models/vit_hybrid/mod*_vit_hybrid* @amyeroberts @qubvel
+/src/transformers/models/vitdet/mod*_vitdet* @amyeroberts @qubvel
+/src/transformers/models/vit_mae/mod*_vit_mae* @amyeroberts @qubvel
+/src/transformers/models/vitmatte/mod*_vitmatte* @amyeroberts @qubvel
+/src/transformers/models/vit_msn/mod*_vit_msn* @amyeroberts @qubvel
+/src/transformers/models/vitpose/mod*_vitpose* @amyeroberts @qubvel
+/src/transformers/models/yolos/mod*_yolos* @amyeroberts @qubvel
+/src/transformers/models/zoedepth/mod*_zoedepth* @amyeroberts @qubvel
 
 # Audio models
 /src/transformers/models/audio_spectrogram_transformer/mod*_audio_spectrogram_transformer* @eustlb
@@ -303,7 +304,7 @@ trainer_utils.py @zach-huggingface @SunMarc
 /src/transformers/models/donut/mod*_donut* @zucchini-nlp
 /src/transformers/models/flava/mod*_flava* @zucchini-nlp
 /src/transformers/models/git/mod*_git* @zucchini-nlp
-/src/transformers/models/grounding_dino/mod*_grounding_dino* @yonigozlan
+/src/transformers/models/grounding_dino/mod*_grounding_dino* @qubvel
 /src/transformers/models/groupvit/mod*_groupvit* @zucchini-nlp
 /src/transformers/models/idefics/mod*_idefics* @zucchini-nlp
 /src/transformers/models/idefics2/mod*_idefics2* @zucchini-nlp
@@ -325,10 +326,10 @@ trainer_utils.py @zach-huggingface @SunMarc
 /src/transformers/models/mgp_str/mod*_mgp_str* @zucchini-nlp
 /src/transformers/models/mllama/mod*_mllama* @zucchini-nlp
 /src/transformers/models/nougat/mod*_nougat* @NielsRogge
-/src/transformers/models/omdet_turbo/mod*_omdet_turbo* @yonigozlan
+/src/transformers/models/omdet_turbo/mod*_omdet_turbo* @qubvel @yonigozlan
 /src/transformers/models/oneformer/mod*_oneformer* @zucchini-nlp
-/src/transformers/models/owlvit/mod*_owlvit* @yonigozlan
-/src/transformers/models/owlv2/mod*_owlv2* @yonigozlan
+/src/transformers/models/owlvit/mod*_owlvit* @qubvel
+/src/transformers/models/owlv2/mod*_owlv2* @qubvel
 /src/transformers/models/paligemma/mod*_paligemma* @zucchini-nlp @molbap
 /src/transformers/models/perceiver/mod*_perceiver* @zucchini-nlp
 /src/transformers/models/pix2struct/mod*_pix2struct* @zucchini-nlp

.github/workflows/benchmark.yml

@@ -12,8 +12,6 @@ concurrency:
 
 env:
   HF_HOME: /mnt/cache
-  DATASET_ID: hf-benchmarks/transformers
-  MODEL_ID: meta-llama/Llama-3.1-8B-Instruct
 
 jobs:
   benchmark:
@@ -36,12 +34,26 @@ jobs:
         with:
           ref: ${{ github.event.pull_request.head.sha || github.sha }}
 
+      - name: Install libpq-dev & psql
+        run: |
+          apt update
+          apt install -y libpq-dev postgresql-client
+
       - name: Install benchmark script dependencies
-        run: python3 -m pip install -r benchmark_v2/requirements.txt kernels
+        run: python3 -m pip install -r benchmark/requirements.txt
 
       - name: Reinstall transformers in edit mode (remove the one installed during docker image build)
         working-directory: /transformers
-        run: python3 -m pip uninstall -y transformers && python3 -m pip install -e ".[torch]" && python3 -m pip uninstall -y torchvision # temp fix
+        run: python3 -m pip uninstall -y transformers && python3 -m pip install -e ".[torch]"
+
+      - name: Run database init script
+        run: |
+          psql -f benchmark/utils/init_db.sql
+        env:
+          PGDATABASE: metrics
+          PGHOST: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGHOST }}
+          PGUSER: transformers_benchmarks
+          PGPASSWORD: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGPASSWORD }}
 
       - name: Run benchmark
         run: |
@@ -52,11 +64,13 @@ jobs:
             commit_id=$GITHUB_SHA
           fi
           commit_msg=$(git show -s --format=%s | cut -c1-70)
-          python3 benchmark_v2/run_benchmarks.py -b 32 -s 128 -n 256 --branch-name "$BRANCH_NAME" --commit-id "$commit_id" --commit-message "$commit_msg" --model-id "$MODEL_ID" --log-level INFO --push-result-to-dataset "$DATASET_ID"
+          python3 benchmark/benchmarks_entrypoint.py "huggingface/transformers" "$BRANCH_NAME" "$commit_id" "$commit_msg"
         env:
           HF_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
-          PUSH_TO_HUB_TOKEN: ${{ secrets.PUSH_TO_HUB_TOKEN }}
           # Enable this to see debug logs
           # HF_HUB_VERBOSITY: debug
           # TRANSFORMERS_VERBOSITY: debug
+          PGHOST: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGHOST }}
+          PGUSER: transformers_benchmarks
+          PGPASSWORD: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGPASSWORD }}
           BRANCH_NAME: ${{ github.head_ref || github.ref_name }}

.github/workflows/benchmark_v2.yml

@@ -1,7 +1,27 @@
 name: Benchmark v2 Framework
 
 on:
-  workflow_dispatch:
+  workflow_call:
+    inputs:
+      runner:
+        description: 'GH Actions runner group to use'
+        required: true
+        type: string
+      commit_sha:
+        description: 'Commit SHA to benchmark'
+        required: false
+        type: string
+        default: ''
+      run_id:
+        description: 'Custom run ID for organizing results (auto-generated if not provided)'
+        required: false
+        type: string
+        default: ''
+      benchmark_repo_id:
+        description: 'HuggingFace Dataset to upload results to (e.g., "org/benchmark-results")'
+        required: false
+        type: string
+        default: ''
 
 env:
   HF_HOME: /mnt/cache
@@ -18,8 +38,8 @@ jobs:
       (github.event_name == 'pull_request' && contains( github.event.pull_request.labels.*.name, 'run-benchmark')) ||
       (github.event_name == 'schedule')
     container:
-      image: ${{ inputs.container_image }}
-      options: ${{ inputs.container_options }}
+      image: huggingface/transformers-pytorch-gpu
+      options: --gpus all --privileged --ipc host --shm-size "16gb"
     steps:
       - name: Get repo
         uses: actions/checkout@v4
@@ -54,4 +74,4 @@ jobs:
           --token '${{ secrets.TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN }}' \
           --log-level INFO
         env:
-          HF_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}

.github/workflows/benchmark_v2_a10_caller.yml

@@ -1,7 +1,11 @@
 name: Benchmark v2 Scheduled Runner - A10 Single-GPU
 
 on:
-  workflow_dispatch:
+  schedule:
+    # Run daily at 16:30 UTC
+    - cron: "30 16 * * *"
+  pull_request:
+    types: [ opened, labeled, reopened, synchronize ]
 
 jobs:
   benchmark-v2-default:
@@ -9,9 +13,7 @@ jobs:
     uses: ./.github/workflows/benchmark_v2.yml
     with:
       runner: aws-g5-4xlarge-cache-use1-public-80
-      container_image: huggingface/transformers-pytorch-gpu
-      container_options: --gpus all --privileged --ipc host --shm-size "16gb"
       commit_sha: ${{ github.sha }}
       run_id: ${{ github.run_id }}
       benchmark_repo_id: hf-internal-testing/transformers-daily-benchmarks
-    secrets: inherit
+    secrets: inherit

.github/workflows/benchmark_v2_mi325_caller.yml

@@ -1,7 +1,11 @@
 name: Benchmark v2 Scheduled Runner - MI325 Single-GPU
 
 on:
-  workflow_dispatch:
+  schedule:
+    # Run daily at 16:30 UTC
+    - cron: "30 16 * * *"
+  pull_request:
+    types: [ opened, labeled, reopened, synchronize ]
 
 jobs:
   benchmark-v2-default:
@@ -9,9 +13,7 @@ jobs:
     uses: ./.github/workflows/benchmark_v2.yml
     with:
       runner: amd-mi325-ci-1gpu
-      container_image: huggingface/transformers-pytorch-amd-gpu
-      container_options: --device /dev/kfd --device /dev/dri --env ROCR_VISIBLE_DEVICES --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache
       commit_sha: ${{ github.sha }}
       run_id: ${{ github.run_id }}
       benchmark_repo_id: hf-internal-testing/transformers-daily-benchmarks
-    secrets: inherit
+    secrets: inherit

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

@@ -5,7 +5,6 @@ on:
     branches:
       - build_ci_docker_image*
   repository_dispatch:
-  workflow_dispatch:
   workflow_call:
     inputs:
       image_postfix:
@@ -222,7 +221,7 @@ jobs:
   latest-pytorch-amd:
     name: "Latest PyTorch (AMD) [dev]"
     runs-on:
-      group: aws-highcpu-32-priv
+      group: aws-general-8-plus
     steps:
       -
         name: Set up Docker Buildx

.github/workflows/build_documentation.yml

@@ -16,20 +16,8 @@ jobs:
       commit_sha: ${{ github.sha }}
       package: transformers
       notebook_folder: transformers_doc
-      languages: en
+      languages: ar de en es fr hi it ja ko pt zh
       custom_container: huggingface/transformers-doc-builder
     secrets:
       token: ${{ secrets.HUGGINGFACE_PUSH }}
       hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}
-
-   build_other_lang:
-    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@main
-    with:
-      commit_sha: ${{ github.sha }}
-      package: transformers
-      notebook_folder: transformers_doc
-      languages: ar de es fr hi it ja ko pt zh
-      custom_container: huggingface/transformers-doc-builder
-    secrets:
-      token: ${{ secrets.HUGGINGFACE_PUSH }}
-      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/check_failed_tests.yml

@@ -35,20 +35,16 @@ env:
   # 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`.
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   TF_FORCE_GPU_ALLOW_GROWTH: true
   CUDA_VISIBLE_DEVICES: 0,1
 
 
 jobs:
   check_new_failures:
-    name: "Find commits for new failing tests"
-    strategy:
-      matrix:
-        run_idx: [1]
+    name: " "
     runs-on:
       group: aws-g5-4xlarge-cache
-    outputs:
-      process: ${{ steps.check_file.outputs.process }}
     container:
       image: ${{ inputs.docker }}
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@@ -59,17 +55,14 @@ jobs:
           path: /transformers/ci_results_${{ inputs.job }}
 
       - name: Check file
-        id: check_file
         working-directory: /transformers
         run: |
           if [ -f ci_results_${{ inputs.job }}/new_failures.json ]; then
             echo "`ci_results_${{ inputs.job }}/new_failures.json` exists, continue ..."
             echo "process=true" >> $GITHUB_ENV
-            echo "process=true" >> $GITHUB_OUTPUT
           else
             echo "`ci_results_${{ inputs.job }}/new_failures.json` doesn't exist, abort."
             echo "process=false" >> $GITHUB_ENV
-            echo "process=false" >> $GITHUB_OUTPUT
           fi
 
       - uses: actions/download-artifact@v4
@@ -126,10 +119,6 @@ jobs:
         run: |
           python3 utils/print_env.py
 
-      - name: Install pytest-flakefinder
-        if: ${{ env.process == 'true' }}
-        run: python3 -m pip install pytest-flakefinder
-
       - name: Show installed libraries and their versions
         working-directory: /transformers
         if: ${{ env.process == 'true' }}
@@ -138,63 +127,25 @@ jobs:
       - name: Check failed tests
         working-directory: /transformers
         if: ${{ env.process == 'true' }}
-        run: python3 utils/check_bad_commit.py --start_commit ${{ inputs.start_sha }} --end_commit ${{ env.END_SHA }} --file ci_results_${{ inputs.job }}/new_failures.json --output_file new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}.json
+        run: python3 utils/check_bad_commit.py --start_commit ${{ inputs.start_sha }} --end_commit ${{ env.END_SHA }} --file ci_results_${{ inputs.job }}/new_failures.json --output_file new_failures_with_bad_commit.json
 
       - name: Show results
         working-directory: /transformers
         if: ${{ env.process == 'true' }}
         run: |
-          ls -l new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}.json
-          cat new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}.json
+          ls -l new_failures_with_bad_commit.json
+          cat new_failures_with_bad_commit.json
 
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          name: new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}
-          path: /transformers/new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}.json
-
-  process_new_failures_with_commit_info:
-    name: "process bad commit reports"
-    needs: check_new_failures
-    if: needs.check_new_failures.outputs.process == 'true'
-    runs-on:
-      group: aws-g5-4xlarge-cache
-    container:
-      image: ${{ inputs.docker }}
-      options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
-    steps:
-      - uses: actions/download-artifact@v4
-        with:
-          name: ci_results_${{ inputs.job }}
-          path: /transformers/ci_results_${{ inputs.job }}
-
-      - uses: actions/download-artifact@v4
-        with:
-          pattern: new_failures_with_bad_commit_${{ inputs.job }}*
-          path: /transformers/new_failures_with_bad_commit_${{ inputs.job }}
-          merge-multiple: true
-
-      - name: Check files
+      - name: Checkout back
         working-directory: /transformers
+        if: ${{ env.process == 'true' }}
         run: |
-          ls -la /transformers
-          ls -la /transformers/new_failures_with_bad_commit_${{ inputs.job }}
-
-      # Currently, we only run with a single runner by using `run_idx: [1]`. We might try to run with multiple runners
-      # to further reduce the false positive caused by flaky tests, which requires further processing to merge reports.
-      - name: Merge files
-        shell: bash
-        working-directory: /transformers
-        run: |
-          cp /transformers/new_failures_with_bad_commit_${{ inputs.job }}/new_failures_with_bad_commit_${{ inputs.job }}_1.json new_failures_with_bad_commit.json
-
-      - name: Update clone
-        working-directory: /transformers
-        run: git fetch && git checkout ${{ inputs.commit_sha || github.sha }}
+          git checkout ${{ inputs.start_sha }}
 
       - name: Process report
         shell: bash
         working-directory: /transformers
+        if: ${{ env.process == 'true' }}
         env:
           ACCESS_REPO_INFO_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
           TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN: ${{ secrets.TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN }}
@@ -206,6 +157,7 @@ jobs:
       - name: Process report
         shell: bash
         working-directory: /transformers
+        if: ${{ env.process == 'true' }}
         env:
           ACCESS_REPO_INFO_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
           TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN: ${{ secrets.TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN }}
@@ -220,12 +172,13 @@ jobs:
 
       - name: Prepare Slack report title
         working-directory: /transformers
+        if: ${{ env.process == 'true' }}
         run: |
           pip install slack_sdk
           echo "title=$(python3 -c 'import sys; sys.path.append("utils"); from utils.notification_service import job_to_test_map; ci_event = "${{ inputs.ci_event }}"; job = "${{ inputs.job }}"; test_name = job_to_test_map[job]; title = f"New failed tests of {ci_event}" + ":" + f" {test_name}"; print(title)')" >> $GITHUB_ENV
 
       - name: Send processed report
-        if: ${{ !endsWith(env.REPORT_TEXT, '{}') }}
+        if: ${{ env.process == 'true' && !endsWith(env.REPORT_TEXT, '{}') }}
         uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
         with:
           # Slack channel id, channel name, or user id to post message.

.github/workflows/doctest_job.yml

@@ -16,6 +16,7 @@ env:
   RUN_SLOW: yes
   OMP_NUM_THREADS: 16
   MKL_NUM_THREADS: 16
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   TF_FORCE_GPU_ALLOW_GROWTH: true
 
 jobs:

.github/workflows/model_jobs.yml

@@ -38,6 +38,7 @@ env:
   # 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`.
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   TF_FORCE_GPU_ALLOW_GROWTH: true
   CUDA_VISIBLE_DEVICES: 0,1
 

.github/workflows/model_jobs_intel_gaudi.yml

@@ -26,6 +26,7 @@ env:
   TRANSFORMERS_IS_CI: yes
   PT_ENABLE_INT64_SUPPORT: 1
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   HF_HOME: /mnt/cache/.cache/huggingface
 
 jobs:

.github/workflows/pr_build_doc_with_comment.yml

@@ -14,7 +14,7 @@ permissions: {}
 jobs:
   get-pr-number:
     name: Get PR number
-    if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "eustlb", "MekkCyber", "vasqu", "ivarflakstad", "stevhliu", "ebezzam", "itazap"]'), github.actor) && (startsWith(github.event.comment.body, 'build-doc')) }}
+    if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "muellerzr", "eustlb", "MekkCyber", "manueldeprada", "vasqu", "ivarflakstad", "stevhliu", "ebezzam", "itazap"]'), github.actor) && (startsWith(github.event.comment.body, 'build-doc')) }}
     uses: ./.github/workflows/get-pr-number.yml
 
   get-pr-info:
@@ -98,7 +98,7 @@ jobs:
       commit_sha: ${{ needs.get-pr-info.outputs.PR_HEAD_SHA }}
       pr_number: ${{ needs.get-pr-number.outputs.PR_NUMBER }}
       package: transformers
-      languages: ar de en es fr hi it ja ko pt zh
+      languages: ar de en es fr hi it ko pt tr zh ja te
 
   update_run_status:
     name: Update Check Run Status

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

@@ -20,6 +20,7 @@ env:
   # 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`.
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   TF_FORCE_GPU_ALLOW_GROWTH: true
   CUDA_VISIBLE_DEVICES: 0,1
 
@@ -28,7 +29,7 @@ jobs:
     runs-on: ubuntu-22.04
     name: Get PR number
     # For security: only allow team members to run
-    if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "eustlb", "MekkCyber", "vasqu", "ivarflakstad", "stevhliu", "ebezzam", "remi-or", "itazap"]'), github.actor) && (startsWith(github.event.comment.body, 'run-slow') || startsWith(github.event.comment.body, 'run slow') || startsWith(github.event.comment.body, 'run_slow')) }}
+    if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "muellerzr", "eustlb", "MekkCyber", "manueldeprada", "vasqu", "ivarflakstad", "stevhliu", "ebezzam", "remi-or", "itazap"]'), github.actor) && (startsWith(github.event.comment.body, 'run-slow') || startsWith(github.event.comment.body, 'run slow') || startsWith(github.event.comment.body, 'run_slow')) }}
     outputs:
       PR_NUMBER: ${{ steps.set_pr_number.outputs.PR_NUMBER }}
     steps:

.github/workflows/self-scheduled-amd-mi325-caller.yml

@@ -20,7 +20,7 @@ jobs:
     with:
       job: run_models_gpu
       slack_report_channel: "#amd-hf-ci"
-      runner_group: amd-mi325
+      runner_scale_set: amd-mi325-ci
       docker: huggingface/transformers-pytorch-amd-gpu
       ci_event: Scheduled CI (AMD) - mi325
       report_repo_id: optimum-amd/transformers_daily_ci
@@ -33,7 +33,7 @@ jobs:
     with:
       job: run_pipelines_torch_gpu
       slack_report_channel: "#amd-hf-ci"
-      runner_group: amd-mi325
+      runner_scale_set: amd-mi325-ci
       docker: huggingface/transformers-pytorch-amd-gpu
       ci_event: Scheduled CI (AMD) - mi325
       report_repo_id: optimum-amd/transformers_daily_ci
@@ -46,7 +46,7 @@ jobs:
     with:
       job: run_examples_gpu
       slack_report_channel: "#amd-hf-ci"
-      runner_group: amd-mi325
+      runner_scale_set: amd-mi325-ci
       docker: huggingface/transformers-pytorch-amd-gpu
       ci_event: Scheduled CI (AMD) - mi325
       report_repo_id: optimum-amd/transformers_daily_ci
@@ -59,7 +59,7 @@ jobs:
     with:
       job: run_torch_cuda_extensions_gpu
       slack_report_channel: "#amd-hf-ci"
-      runner_group: amd-mi325
+      runner_scale_set: amd-mi325-ci
       docker: huggingface/transformers-pytorch-deepspeed-amd-gpu
       ci_event: Scheduled CI (AMD) - mi325
       report_repo_id: optimum-amd/transformers_daily_ci

.github/workflows/self-scheduled-amd-mi355-caller.yml

@@ -3,7 +3,7 @@ name: Self-hosted runner scale set (AMD mi355 scheduled CI caller)
 # Note: For every job in this workflow, the name of the runner scale set is finalized in the runner yaml i.e. huggingface/hf-workflows/.github/workflows/transformers_amd_ci_scheduled_arc_scale_set.yaml
 # For example, 1gpu : amd-mi355-ci-1gpu
 #              2gpu : amd-mi355-ci-2gpu
- 
+
 on:
   workflow_run:
     workflows: ["Self-hosted runner (AMD scheduled CI caller)"]
@@ -20,7 +20,7 @@ jobs:
     with:
       job: run_models_gpu
       slack_report_channel: "#amd-hf-ci"
-      runner_group: hfc-amd-mi355
+      runner_scale_set: amd-mi355-ci
       docker: huggingface/testing-rocm7.0-preview
       ci_event: Scheduled CI (AMD) - mi355
       report_repo_id: hf-transformers-bot/transformers-ci-dummy
@@ -32,7 +32,7 @@ jobs:
     with:
       job: run_pipelines_torch_gpu
       slack_report_channel: "#amd-hf-ci"
-      runner_group: hfc-amd-mi355
+      runner_scale_set: amd-mi355-ci
       docker: huggingface/testing-rocm7.0-preview
       ci_event: Scheduled CI (AMD) - mi355
       report_repo_id: hf-transformers-bot/transformers-ci-dummy
@@ -44,7 +44,7 @@ jobs:
     with:
       job: run_examples_gpu
       slack_report_channel: "#amd-hf-ci"
-      runner_group: hfc-amd-mi355
+      runner_scale_set: amd-mi355-ci
       docker: huggingface/testing-rocm7.0-preview
       ci_event: Scheduled CI (AMD) - mi355
       report_repo_id: hf-transformers-bot/transformers-ci-dummy
@@ -53,10 +53,10 @@ jobs:
   deepspeed-ci:
     name: DeepSpeed CI
     uses: huggingface/hf-workflows/.github/workflows/transformers_amd_ci_scheduled_arc_scale_set.yaml@main
-    with:  
+    with:
       job: run_torch_cuda_extensions_gpu
       slack_report_channel: "#amd-hf-ci"
-      runner_group: hfc-amd-mi355
+      runner_scale_set: amd-mi355-ci
       docker: huggingface/testing-rocm7.0-preview
       ci_event: Scheduled CI (AMD) - mi355
       report_repo_id: hf-transformers-bot/transformers-ci-dummy

.github/workflows/self-scheduled-intel-gaudi.yml

@@ -26,6 +26,7 @@ env:
   TRANSFORMERS_IS_CI: yes
   PT_ENABLE_INT64_SUPPORT: 1
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   HF_HOME: /mnt/cache/.cache/huggingface
 
 jobs:

.github/workflows/self-scheduled.yml

@@ -48,6 +48,7 @@ env:
   # 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`.
   HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
+  SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
   TF_FORCE_GPU_ALLOW_GROWTH: true
   CUDA_VISIBLE_DEVICES: 0,1
   NUM_SLICES: 2

.github/workflows/ssh-runner.yml

@@ -20,6 +20,7 @@ env:
   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
 
@@ -32,17 +33,14 @@ jobs:
     steps:
       - name: Get runner to use
         shell: bash
-        env:
-          NUM_GPUS: ${{ github.event.inputs.num_gpus }}
-          RUNNER_TYPE: ${{ github.event.inputs.runner_type }}
         run: |
-          if [[ "$NUM_GPUS" == "single" && "$RUNNER_TYPE" == "t4" ]]; then
+          if [[ "${{ github.event.inputs.num_gpus }}" == "single" && "${{ github.event.inputs.runner_type }}" == "t4" ]]; then
             echo "RUNNER=aws-g4dn-4xlarge-cache" >> $GITHUB_ENV
-          elif [[ "$NUM_GPUS" == "multi" && "$RUNNER_TYPE" == "t4" ]]; then
+          elif [[ "${{ github.event.inputs.num_gpus }}" == "multi" && "${{ github.event.inputs.runner_type }}" == "t4" ]]; then
             echo "RUNNER=aws-g4dn-12xlarge-cache" >> $GITHUB_ENV
-          elif [[ "$NUM_GPUS" == "single" && "$RUNNER_TYPE" == "a10" ]]; then
+          elif [[ "${{ github.event.inputs.num_gpus }}" == "single" && "${{ github.event.inputs.runner_type }}" == "a10" ]]; then
             echo "RUNNER=aws-g5-4xlarge-cache" >> $GITHUB_ENV
-          elif [[ "$NUM_GPUS" == "multi" && "$RUNNER_TYPE" == "a10" ]]; then
+          elif [[ "${{ github.event.inputs.num_gpus }}" == "multi" && "${{ github.event.inputs.runner_type }}" == "a10" ]]; then
             echo "RUNNER=aws-g5-12xlarge-cache" >> $GITHUB_ENV
           else
             echo "RUNNER=" >> $GITHUB_ENV
@@ -87,11 +85,9 @@ jobs:
       - name: Store Slack infos
         #because the SSH can be enabled dynamically if the workflow failed, so we need to store slack infos to be able to retrieve them during the waitforssh step
         shell: bash
-        env:
-          GITHUB_ACTOR: ${{ github.actor }}
         run: |
-          echo "$GITHUB_ACTOR"
-          github_actor=$GITHUB_ACTOR
+          echo "${{ github.actor }}"
+          github_actor=${{ github.actor }}
           github_actor=${github_actor/'-'/'_'}
           echo "$github_actor"
           echo "github_actor=$github_actor" >> $GITHUB_ENV

.gitignore

@@ -98,7 +98,6 @@ celerybeat-schedule
 # Environments
 .env
 .venv
-.venv*
 env/
 venv/
 ENV/
@@ -172,6 +171,3 @@ tags
 
 # modular conversion
 *.modular_backup
-
-# Cursor IDE files
-.cursor/

CONTRIBUTING.md

@@ -112,125 +112,7 @@ New models are constantly released and if you want to implement a new model, ple
 
 If you are willing to contribute the model yourself, let us know so we can help you add it to 🤗 Transformers!
 
-We have a technical guide for [how to add a model to 🤗 Transformers](https://huggingface.co/docs/transformers/modular_transformers).
-
-### Vision-Language Model Contribution Checklist
-
-If you're contributing a **vision-language model** (or any multimodal model that processes images/videos), please follow this checklist. Maintainers will use this to review your PR, and completing these steps will significantly increase the likelihood of your PR being merged quickly.
-
-**Required checklist for all vision-language model contributions:**
-
-☐ **1. Implement a modular file**
-
-All new models should use the modular architecture pattern. Create a `modular_<model_name>.py` file using the modular model converter:
-
-- Use the CLI, [`transformers add-new-model-like`](https://github.com/huggingface/transformers/blob/main/src/transformers/cli/add_new_model_like.py) to generate a modular skeleton and get started
-- All code should be in the modular file if possible. Modeling must be in it, it's better if configuration is in it as well. 
-- Reuse existing patterns from similar models as much as possible
-
-To verify your modular file is correct, run:
-
-```bash
-python utils/modular_model_converter.py <model_name>
-```
-
-This will generate the separate files (`modeling_*.py`, `configuration_*.py`, etc.) from your modular file. The CI will enforce that these generated files match your modular file.
-
-☐ **2. Add a fast image processor (for image models)**
-
-If your model processes images, implement a fast image processor that uses `torch` and `torchvision` instead of PIL/numpy for better inference performance:
-
-- See the detailed guide in [#36978](https://github.com/huggingface/transformers/issues/36978)
-- Fast processors inherit from `BaseImageProcessorFast`
-- Examples: `LlavaOnevisionImageProcessorFast`, `Idefics2ImageProcessorFast`
-
-☐ **3. Create a weight conversion script**
-
-Add a `convert_<model_name>_to_hf.py` script that converts the original model weights to the HuggingFace format:
-
-- Script should handle checkpoint loading, key mapping, and saving in HF format
-- Include usage examples and documentation in the script
-- Examples: [`convert_llava_onevision_weights_to_hf.py`](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/convert_llava_onevision_weights_to_hf.py), [`convert_idefics2_weights_to_hf.py`](https://github.com/huggingface/transformers/blob/main/src/transformers/models/idefics2/convert_idefics2_weights_to_hf.py)
-
-☐ **4. Add integration tests with exact output matching**
-
-At minimum, add an `IntegrationTest` class that tests end-to-end generation (processing and modelling) with **exact** output matching:
-
-- For generative models: test that generated text matches expected output exactly
-- For non-generative models: test that output logits match expected values
-- Tests should use real checkpoints (load in 4-bit or half precision if the checkpoint is too big to fit in our CI runners) and real inputs
-- Example pattern:
-
-```python
-class MyModelIntegrationTest(unittest.TestCase):
-    @slow
-    def test_model_integration(self):
-        model = MyModelForConditionalGeneration.from_pretrained("org/model-name")
-        processor = AutoProcessor.from_pretrained("org/model-name")
-
-        inputs = processor(images=image, text=prompt, return_tensors="pt")
-        output = model.generate(**inputs, max_new_tokens=20)
-
-        EXPECTED_TEXT = "exact expected output"
-        self.assertEqual(processor.decode(output[0]), EXPECTED_TEXT)
-```
-
-See `tests/models/llava_onevision/test_modeling_llava_onevision.py` for complete examples.
-
-☐ **5. Update documentation**
-
-Add or update model documentation:
-
-- Create if the cli hasn't `docs/source/en/model_doc/<model_name>.md` with usage examples
-- Include model description, paper link, and basic usage with `Pipeline` and `AutoModel`
-- Add the model to the appropriate TOC files
-
-☐ **6. Look for reusable patterns**
-
-The library has 400+ models with many established patterns:
-
-- Search for similar models (e.g., other vision-language models)
-- Reuse attention mechanisms, layer implementations, and processing patterns
-- Check models like LLaVA, Idefics2, Fuyu for vision-language patterns
-- Use provided decorators like (`auto_docstring`, `can_return_tuple`, `check_model_inputs` and `_can_record_outputs`) where relevant. 
-- Don't reinvent the wheel
-
-☐ **7. Run quality checks and read the output**
-
-Before submitting your PR, install quality dependencies and run the full check suite:
-
-```bash
-pip install -e ".[quality]"
-make fixup
-```
-
-**Important**: Take time to read the output of `make fixup`. It will:
-- Lint and format your code automatically
-- Run consistency checks (imports, docstrings, etc.)
-- Show any remaining issues that need manual fixes
-
-All checks must pass before your PR can be merged.
-
-**If this checklist is complete, your PR has a very high likelihood of being merged!** Following these steps makes the maintainers' work much easier and will reduce the number of review iterations, getting your important work out there faster.
-
-#### Copy-pastable checklist for maintainers
-
-Here's a condensed version maintainers can copy into PRs:
-
-```markdown
-## Multimodal Model Addition Checklist
-
-Please ensure your PR completes all following items. See the [full checklist](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#vision-language-model-contribution-checklist) for details.
-
-- [ ] **Modular file**: `modular_<model_name>.py` implemented and verified with `python utils/modular_model_converter.py <model_name>`
-- [ ] **Fast image processor**: Implemented using `BaseImageProcessorFast` (see [#36978](https://github.com/huggingface/transformers/issues/36978))
-- [ ] **Conversion script**: `convert_<model_name>_to_hf.py` added with usage examples
-- [ ] **Integration tests**: End-to-end tests with exact output matching (text or logits)
-- [ ] **Documentation**: Model docs added/updated in `docs/source/en/model_doc/`
-- [ ] **Pattern reuse**: Verified against similar models (LLaVA, Idefics2, etc.)
-- [ ] **Quality checks**: `make fixup` passes with no errors
-
-```
+We have a technical guide for [how to add a model to 🤗 Transformers](https://huggingface.co/docs/transformers/add_new_model).
 
 ## Do you want to add documentation?
 
@@ -396,14 +278,13 @@ are working on it).<br>
 useful to avoid duplicated work, and to differentiate it from PRs ready to be merged.<br>
 ☐ Make sure existing tests pass.<br>
 ☐ If adding a new feature, also add tests for it.<br>
-
-- If you are adding a new model, make sure you use
+   - If you are adding a new model, make sure you use
      `ModelTester.all_model_classes = (MyModel, MyModelWithLMHead,...)` to trigger the common tests.
-- If you are adding new `@slow` tests, make sure they pass using
+   - If you are adding new `@slow` tests, make sure they pass using
      `RUN_SLOW=1 python -m pytest tests/models/my_new_model/test_my_new_model.py`.
-- If you are adding a new tokenizer, write tests and make sure
+   - If you are adding a new tokenizer, write tests and make sure
      `RUN_SLOW=1 python -m pytest tests/models/{your_model_name}/test_tokenization_{your_model_name}.py` passes.
-- CircleCI does not run the slow tests, but GitHub Actions does every night!<br>
+   - CircleCI does not run the slow tests, but GitHub Actions does every night!<br>
 
 ☐ All public methods must have informative docstrings (see
 [`modeling_bert.py`](https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py)
@@ -459,7 +340,6 @@ 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.
 
 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).

ISSUES.md

@@ -153,7 +153,7 @@ You are not required to read the following guidelines before opening an issue. H
     cd examples/seq2seq
     torchrun --nproc_per_node=2 ./finetune_trainer.py \
     --model_name_or_path sshleifer/distill-mbart-en-ro-12-4 --data_dir wmt_en_ro \
-    --output_dir output_dir \
+    --output_dir output_dir --overwrite_output_dir \
     --do_train --n_train 500 --num_train_epochs 1 \
     --per_device_train_batch_size 1  --freeze_embeds \
     --src_lang en_XX --tgt_lang ro_RO --task translation \

README.md

@@ -48,7 +48,6 @@ limitations under the License.
         <a href="https://github.com/huggingface/transformers/blob/main/i18n/README_te.md">తెలుగు</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/i18n/README_fr.md">Français</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/i18n/README_de.md">Deutsch</a> |
-        <a href="https://github.com/huggingface/transformers/blob/main/i18n/README_it.md">Italiano</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/i18n/README_vi.md">Tiếng Việt</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/i18n/README_ar.md">العربية</a> |
         <a href="https://github.com/huggingface/transformers/blob/main/i18n/README_ur.md">اردو</a> |
@@ -64,8 +63,8 @@ limitations under the License.
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_as_a_model_definition.png"/>
 </h3>
 
-Transformers acts as the model-definition framework for state-of-the-art machine learning with text, computer
-vision, audio, video, and multimodal models, for both inference and training.
+Transformers acts as the model-definition framework for state-of-the-art machine learning models in text, computer
+vision, audio, video, and multimodal model, for both inference and training.
 
 It centralizes the model definition so that this definition is agreed upon across the ecosystem. `transformers` is the
 pivot across frameworks: if a model definition is supported, it will be compatible with the majority of training
@@ -111,10 +110,10 @@ git clone https://github.com/huggingface/transformers.git
 cd transformers
 
 # pip
-pip install '.[torch]'
+pip install .[torch]
 
 # uv
-uv pip install '.[torch]'
+uv pip install .[torch]
 ```
 
 ## Quickstart

benchmark/benches/llama.py

@@ -16,6 +16,7 @@ import sys
 from logging import Logger
 from threading import Event, Thread
 from time import perf_counter, sleep
+from typing import Optional
 
 
 # Add the parent directory to Python path to import benchmarks_entrypoint
@@ -41,7 +42,7 @@ except ImportError:
     GenerationConfig = None
     StaticCache = None
 
-os.environ["HF_XET_HIGH_PERFORMANCE"] = "1"
+os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 os.environ["TOKENIZERS_PARALLELISM"] = "1"
 
 # Only set torch precision if torch is available
@@ -144,7 +145,7 @@ def run_benchmark(
             q = torch.empty_like(probs_sort).exponential_(1)
             return torch.argmax(probs_sort / q, dim=-1, keepdim=True).to(dtype=torch.int)
 
-        def logits_to_probs(logits, temperature: float = 1.0, top_k: int | None = None):
+        def logits_to_probs(logits, temperature: float = 1.0, top_k: Optional[int] = None):
             logits = logits / max(temperature, 1e-5)
 
             if top_k is not None:
@@ -154,7 +155,7 @@ def run_benchmark(
             probs = torch.nn.functional.softmax(logits, dim=-1)
             return probs
 
-        def sample(logits, temperature: float = 1.0, top_k: int | None = None):
+        def sample(logits, temperature: float = 1.0, top_k: Optional[int] = None):
             probs = logits_to_probs(logits[0, -1], temperature, top_k)
             idx_next = multinomial_sample_one_no_sync(probs)
             return idx_next, probs

benchmark/requirements.txt

@@ -2,5 +2,5 @@ gpustat==1.1.1
 psutil==6.0.0
 psycopg2==2.9.9
 torch>=2.4.0
-hf_xet
+hf_transfer
 pandas>=1.5.0

benchmark_v2/.gitignore

@@ -1,2 +1 @@
-benchmark_results/
-benchmark_results_profiles/
+benchmark_results/

benchmark_v2/benches/__init__.py

@@ -0,0 +1 @@
+# Benchmark implementations directory

benchmark_v2/benches/llama.py

@@ -0,0 +1,165 @@
+# 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.
+
+import logging
+import os
+from typing import Any
+
+import torch
+from benchmark_framework import ModelBenchmark
+
+
+os.environ["TOKENIZERS_PARALLELISM"] = "1"
+torch.set_float32_matmul_precision("high")
+
+
+class LLaMABenchmark(ModelBenchmark):
+    """Simplified LLaMA model benchmark implementation using the ModelBenchmark base class."""
+
+    def __init__(self, logger: logging.Logger):
+        super().__init__(logger)
+        self._default_prompt = "Why dogs are so cute?"  # Custom prompt for LLaMA
+
+    def get_scenario_configs(self) -> list[dict[str, Any]]:
+        """
+        Get LLaMA-specific scenario configurations.
+
+        Returns:
+            List of scenario configuration dictionaries
+        """
+        return [
+            # Eager variants
+            {"variant": "eager", "compile_mode": None, "use_cache": True, "description": "Eager execution with cache"},
+            # Compiled variants
+            {
+                "variant": "compiled",
+                "compile_mode": "max-autotune",
+                "use_cache": True,
+                "description": "Compiled with max autotune",
+            },
+            # Kernelized variant (if available)
+            {
+                "variant": "kernelized",
+                "compile_mode": "max-autotune",
+                "use_cache": True,
+                "description": "Kernelized execution",
+            },
+        ]
+
+    def _is_kernelization_available(self) -> bool:
+        """Check if kernelization is available for LLaMA."""
+        try:
+            from kernels import Mode, kernelize  # noqa: F401
+
+            return True
+        except ImportError:
+            self.logger.debug("Kernelization not available: kernels module not found")
+            return False
+
+    def get_default_generation_config(self) -> dict[str, Any]:
+        """Get LLaMA-specific generation configuration."""
+        return {
+            "do_sample": False,
+            "top_p": 1.0,
+            "temperature": 1.0,
+            "repetition_penalty": 1.0,
+            "max_new_tokens": None,  # Will be set per scenario
+        }
+
+    def get_model_init_kwargs(self, config) -> dict[str, Any]:
+        """Get LLaMA-specific model initialization kwargs."""
+        return {
+            "torch_dtype": getattr(torch, config.torch_dtype),
+            "attn_implementation": config.attn_implementation,
+            "use_cache": True,
+        }
+
+    def get_default_torch_dtype(self) -> str:
+        """Get default torch dtype for LLaMA."""
+        return "float16"  # LLaMA works well with float16
+
+    def get_default_device(self) -> str:
+        """Get default device for LLaMA."""
+        return "cuda"  # LLaMA prefers CUDA
+
+
+def run_llama(logger, output_dir, **kwargs):
+    """
+    Run LLaMA benchmark with the given configuration.
+
+    Args:
+        logger: Logger instance
+        output_dir: Output directory for results
+        **kwargs: Additional configuration options
+
+    Returns:
+        Path to output file if successful
+    """
+    from benchmark_framework import BenchmarkRunner
+
+    # Extract parameters with defaults
+    model_id = kwargs.get("model_id", "meta-llama/Llama-2-7b-hf")
+    warmup_iterations = kwargs.get("warmup_iterations", 3)
+    measurement_iterations = kwargs.get("measurement_iterations", 5)
+    num_tokens_to_generate = kwargs.get("num_tokens_to_generate", 100)
+    include_sdpa_variants = kwargs.get("include_sdpa_variants", True)
+    device = kwargs.get("device", "cuda")
+    torch_dtype = kwargs.get("torch_dtype", "float16")
+    batch_size = kwargs.get("batch_size", 1)
+    commit_id = kwargs.get("commit_id")
+
+    logger.info(f"Starting LLaMA benchmark for model: {model_id}")
+    logger.info(
+        f"Configuration: warmup={warmup_iterations}, measurement={measurement_iterations}, tokens={num_tokens_to_generate}"
+    )
+
+    try:
+        # Create benchmark instance
+        benchmark = LLaMABenchmark(logger)
+
+        # Create scenarios
+        scenarios = benchmark.create_scenarios(
+            model_id=model_id,
+            warmup_iterations=warmup_iterations,
+            measurement_iterations=measurement_iterations,
+            num_tokens_to_generate=num_tokens_to_generate,
+            include_sdpa_variants=include_sdpa_variants,
+            device=device,
+            torch_dtype=torch_dtype,
+            batch_size=batch_size,
+        )
+
+        logger.info(f"Created {len(scenarios)} benchmark scenarios")
+
+        # Create runner and execute benchmarks
+        runner = BenchmarkRunner(logger, output_dir)
+        results = runner.run_benchmark(benchmark, scenarios, commit_id=commit_id)
+
+        if not results:
+            logger.warning("No successful benchmark results")
+            return None
+
+        # Save results
+        model_name = model_id.split("/")[-1]  # Extract model name from ID
+        output_file = runner.save_results(model_name, results)
+
+        logger.info(f"LLaMA benchmark completed successfully. Results saved to: {output_file}")
+        return output_file
+
+    except Exception as e:
+        logger.error(f"LLaMA benchmark failed: {e}")
+        import traceback
+
+        logger.debug(traceback.format_exc())
+        raise

benchmark_v2/framework/benchmark_config.py

@@ -1,214 +0,0 @@
-import hashlib
-import json
-import logging
-from typing import Any
-
-
-KERNELIZATION_AVAILABLE = False
-try:
-    from kernels import Mode, kernelize  # noqa: F401
-
-    KERNELIZATION_AVAILABLE = True
-except ImportError:
-    pass
-
-logger = logging.getLogger(__name__)
-
-
-class BenchmarkConfig:
-    """Configuration for a single benchmark scenario."""
-
-    def __init__(
-        self,
-        warmup_iterations: int = 5,
-        measurement_iterations: int = 20,
-        gpu_monitoring: bool = True,  # NOTE: you may want to disable this at times as we have obsvered it could heavily slow down benchmarks on AMD
-        batch_size: int = 1,
-        sequence_length: int = 128,
-        num_tokens_to_generate: int = 128,
-        attn_implementation: str = "eager",
-        sdpa_backend: str | None = None,
-        compile_mode: str | None = None,
-        compile_options: dict[str, Any] | None = None,
-        kernelize: bool = False,
-        name: str | None = None,
-        skip_validity_check: bool = False,
-    ) -> None:
-        # Benchmark parameters
-        self.warmup_iterations = warmup_iterations
-        self.measurement_iterations = measurement_iterations
-        self.gpu_monitoring = gpu_monitoring
-        # Input parameters
-        self.batch_size = batch_size
-        self.sequence_length = sequence_length
-        self.num_tokens_to_generate = num_tokens_to_generate
-        # Generation parameters
-        self.attn_implementation = attn_implementation
-        self.sdpa_backend = sdpa_backend
-        # Optimization parameters
-        self.compile_mode = compile_mode
-        self.compile_options = compile_options if compile_options is not None else {}
-        self.kernelize = kernelize
-        # Constant parameters
-        self.dtype = "torch.bfloat16"
-        self.device = "cuda"
-
-        self.check_validity(skip_validity_check)
-        self.name = name if name is not None else self.infer_name()
-
-    def check_validity(self, skip_validity_check: bool = False) -> None:
-        if skip_validity_check:
-            return
-        # Flash attention does not support compile mode, so we turn it off # FIXME: it would be better to support it
-        is_fa = self.attn_implementation == "flash_attention_2"
-        is_fa |= self.attn_implementation == "sdpa" and self.sdpa_backend == "flash_attention"
-        if is_fa:
-            logger.warning("Flash attention does not support compile mode. Turning off compile mode.")
-            self.compile_mode = None
-
-    @property
-    def hash(self) -> str:
-        return hashlib.sha256(json.dumps(self.to_dict()).encode()).hexdigest()
-
-    def infer_name(self, compact: bool = True) -> str:
-        """Infer a human-readable name for the benchmark config, either compact or verbose."""
-        if compact:
-            iter_str = f"w{self.warmup_iterations}_i{self.measurement_iterations}"
-            gpu_monitor_str = "monitored" if self.gpu_monitoring else "unmonitored"
-            dimensions_str = f"b{self.batch_size}_s{self.sequence_length}_n{self.num_tokens_to_generate}"
-            attn_code = self.attn_implementation
-            attn_code += f"_{self.sdpa_backend}" if self.attn_implementation == "sdpa" else ""
-            compile_str = f"compiled_{self.compile_mode}" if self.compile_mode is not None else "uncompiled"
-            kernelize_str = "kernelized" if self.kernelize else "unkernelized"
-            sep = "-"
-        else:
-            iter_str = f"{self.warmup_iterations} warmup, {self.measurement_iterations} iterations"
-            gpu_monitor_str = ("with" if self.gpu_monitoring else "no") + " GPU monitoring"
-            dimensions_str = f"batch size {self.batch_size}, sequence length {self.sequence_length}, {self.num_tokens_to_generate} generated tokens"
-            attn_code = f"{self.attn_implementation} attention"
-            attn_code += f" with {self.sdpa_backend} backend" if self.attn_implementation == "sdpa" else ""
-            compile_str = "compiled" if self.compile_mode is not None else "not compiled"
-            kernelize_str = "kernelized" if self.kernelize else "not kernelized"
-            sep = ", "
-        return sep.join([iter_str, gpu_monitor_str, dimensions_str, attn_code, compile_str, kernelize_str])
-
-    def to_dict(self) -> dict[str, Any]:
-        return {
-            "name": self.name,
-            "warmup_iterations": self.warmup_iterations,
-            "measurement_iterations": self.measurement_iterations,
-            "gpu_monitoring": self.gpu_monitoring,
-            "batch_size": self.batch_size,
-            "sequence_length": self.sequence_length,
-            "num_tokens_to_generate": self.num_tokens_to_generate,
-            "attn_implementation": self.attn_implementation,
-            "sdpa_backend": self.sdpa_backend,
-            "compile_mode": self.compile_mode,
-            "compile_options": self.compile_options | {},  # to avoid inplace modification of the original dict
-            "kernelize": self.kernelize,
-        }
-
-    @classmethod
-    def from_dict(cls, data: dict[str, Any], skip_validity_check: bool = False) -> "BenchmarkConfig":
-        return cls(
-            warmup_iterations=data.get("warmup_iterations", 5),
-            measurement_iterations=data.get("measurement_iterations", 20),
-            gpu_monitoring=data.get("gpu_monitoring", False),
-            batch_size=data.get("batch_size", 1),
-            sequence_length=data.get("sequence_length", 128),
-            num_tokens_to_generate=data.get("num_tokens_to_generate", 128),
-            attn_implementation=data.get("attn_implementation", "eager"),
-            sdpa_backend=data.get("sdpa_backend"),
-            compile_mode=data.get("compile_mode"),
-            compile_options=data.get("compile_options"),
-            kernelize=data.get("kernelize", False),
-            name=data.get("name"),
-            skip_validity_check=skip_validity_check,
-        )
-
-
-def cross_generate_configs(
-    attn_impl_and_sdpa_backend: list[tuple[str, str | None]],
-    compiled_mode: list[str | None],
-    kernelized: list[bool],
-    warmup_iterations: int = 5,
-    measurement_iterations: int = 20,
-    batch_size: int = 1,
-    sequence_length: int = 128,
-    num_tokens_to_generate: int = 128,
-    gpu_monitoring: bool = True,
-) -> list[BenchmarkConfig]:
-    # Create kwargs common to all configs
-    kwargs = {
-        "warmup_iterations": warmup_iterations,
-        "measurement_iterations": measurement_iterations,
-        "batch_size": batch_size,
-        "sequence_length": sequence_length,
-        "num_tokens_to_generate": num_tokens_to_generate,
-        "gpu_monitoring": gpu_monitoring,
-    }
-    # Cross-generate all combinations of attn_implementation, compiled_mode, and kernelized
-    configs = []
-    for attn_implementation, sdpa_backend in list(dict.fromkeys(attn_impl_and_sdpa_backend)):
-        for cm in list(dict.fromkeys(compiled_mode)):
-            for kernelize_on in list(dict.fromkeys(kernelized)):
-                config = BenchmarkConfig(
-                    attn_implementation=attn_implementation,
-                    sdpa_backend=sdpa_backend,
-                    compile_mode=cm,
-                    kernelize=kernelize_on,
-                    **kwargs,
-                )
-                configs.append(config)
-    return configs
-
-
-def generate_all_configs(
-    warmup_iterations: int = 5,
-    measurement_iterations: int = 20,
-    batch_size: int = 1,
-    sequence_length: int = 128,
-    num_tokens_to_generate: int = 128,
-    gpu_monitoring: bool = True,
-) -> list[BenchmarkConfig]:
-    all_attn_implementations = [
-        ("flash_attention_2", None),
-        ("eager", None),
-        ("sdpa", "math"),
-        ("sdpa", "flash_attention"),
-        ("flex_attention", None),
-    ]
-    return cross_generate_configs(
-        attn_impl_and_sdpa_backend=all_attn_implementations,
-        compiled_mode=[None, "default", "reduce-overhead", "max-autotune", "max-autotune-no-cudagraphs"],
-        kernelized=[False, KERNELIZATION_AVAILABLE],
-        warmup_iterations=warmup_iterations,
-        measurement_iterations=measurement_iterations,
-        batch_size=batch_size,
-        sequence_length=sequence_length,
-        num_tokens_to_generate=num_tokens_to_generate,
-        gpu_monitoring=gpu_monitoring,
-    )
-
-
-def generate_main_configs(
-    warmup_iterations: int = 5,
-    measurement_iterations: int = 20,
-    batch_size: int = 1,
-    sequence_length: int = 128,
-    num_tokens_to_generate: int = 128,
-) -> list[BenchmarkConfig]:
-    # Create kwargs common to all configs
-    kwargs = {
-        "warmup_iterations": warmup_iterations,
-        "measurement_iterations": measurement_iterations,
-        "batch_size": batch_size,
-        "sequence_length": sequence_length,
-        "num_tokens_to_generate": num_tokens_to_generate,
-    }
-    return [  # TODO: test max-autotune instead of default
-        BenchmarkConfig(attn_implementation="flex_attention", compile_mode="default", gpu_monitoring=False, **kwargs),
-        BenchmarkConfig(attn_implementation="flex_attention", compile_mode="default", gpu_monitoring=True, **kwargs),
-        BenchmarkConfig(attn_implementation="eager", compile_mode="default", gpu_monitoring=True, **kwargs),
-        BenchmarkConfig(attn_implementation="flash_attention_2", gpu_monitoring=True, **kwargs),
-    ]

benchmark_v2/framework/benchmark_runner.py

@@ -1,456 +0,0 @@
-import gc
-import json
-import logging
-import os
-import pathlib
-import re
-import tempfile
-import time
-from contextlib import nullcontext
-from datetime import datetime
-from queue import Queue
-from typing import Any
-
-import torch
-from datasets import Dataset
-from huggingface_hub import HfApi
-from tqdm import trange
-
-from transformers import (
-    AutoModelForCausalLM,
-    AutoTokenizer,
-    CompileConfig,
-    GenerationConfig,
-    GenerationMixin,
-)
-from transformers.generation.streamers import BaseStreamer
-
-from .benchmark_config import BenchmarkConfig
-from .data_classes import BenchmarkMetadata, BenchmarkResult, GPURawMetrics, pretty_print_dict
-from .hardware_metrics import GPUMonitor
-
-
-try:
-    from kernels import Mode, kernelize  # noqa: F401
-except ImportError:
-    kernelize = None
-    Mode = None
-
-
-DEFAULT_PROMPT = "\n".join([
-    "The French Revolution was a period of political and societal change in France that began with the Estates General of 1789 and ended with the Coup of 18 Brumaire on 9 November 1799.",
-    "Many of the revolution's ideas are considered fundamental principles of liberal democracy, and its values remain central to modern French political discourse.",
-    "It was caused by a combination of social, political, and economic factors which the existing regime proved unable to manage.",
-    "Financial crisis and widespread social distress led to the convocation of the Estates General in May 1789, its first meeting since 1614.",
-    "The representatives of the Third Estate broke away and re-constituted themselves as a National Assembly in June.",
-    "The Storming of the Bastille in Paris on 14 July led to a series of radical measures by the Assembly, including the abolition of feudalism, state control over the Catholic Church in France, and issuing the Declaration of the Rights of Man and of the Citizen.",
-    "The next three years were dominated by a struggle for political control.",
-    "King Louis XVI's attempted flight to Varennes in June 1791 further discredited the monarchy, and military defeats after the outbreak of the French Revolutionary Wars in April 1792 led to the insurrection of 10 August 1792.",
-    "As a result, the monarchy was replaced by the French First Republic in September, followed by the execution of Louis XVI himself in January 1793.",
-    "After another revolt in June 1793, the constitution was suspended, and political power passed from the National Convention to the Committee of Public Safety, dominated by radical Jacobins led by Maximilien Robespierre.",
-    "About 16,000 people were sentenced by the Revolutionary Tribunal and executed in the Reign of Terror, which ended in July 1794 with the Thermidorian Reaction.",
-    "Weakened by external threats and internal opposition, the Committee of Public Safety was replaced in November 1795 by the Directory.",
-    "Its instability ended in the coup of 18 Brumaire and the establishment of the Consulate, with Napoleon Bonaparte as First Consul.",
-])  # fmt: skip
-
-PUSH_TO_HUB_TOKEN = os.getenv("PUSH_TO_HUB_TOKEN", None)
-
-
-def compact_json_numeric_arrays(data: dict):
-    # Match arrays that contain only numbers (ints/floats), whitespace, commas, and newlines
-    pattern = r"\[\s*\n\s*((?:\d+(?:\.\d+)?\s*,\s*)*\d+(?:\.\d+)?)\s*\n\s*\]"
-
-    def replace_numeric_array(match):
-        # Get the array content
-        content = match.group(1)
-        # Remove extra whitespace but keep commas
-        compact_content = re.sub(r"\s+", " ", content).strip()
-        return f"[{compact_content}]"
-
-    return re.sub(pattern, replace_numeric_array, json.dumps(data, indent=4, default=str), flags=re.DOTALL)
-
-
-def get_git_revision() -> str:
-    base_path = pathlib.Path(__file__).parent.parent.parent
-    git_dir = base_path / ".git"
-    with (git_dir / "HEAD").open("r") as head:
-        ref = head.readline().split(" ")[-1].strip()
-    with (git_dir / ref).open("r") as git_hash:
-        return git_hash.readline().strip()
-
-
-def get_sdpa_backend(backend_name: str | None) -> torch.nn.attention.SDPBackend | None:
-    """Get the SDPA backend enum from string name."""
-    if backend_name is None:
-        return None
-
-    try:
-        backend_map = {
-            "math": torch.nn.attention.SDPBackend.MATH,
-            "flash_attention": torch.nn.attention.SDPBackend.FLASH_ATTENTION,
-            "efficient_attention": torch.nn.attention.SDPBackend.EFFICIENT_ATTENTION,
-            "cudnn_attention": torch.nn.attention.SDPBackend.CUDNN_ATTENTION,
-        }
-        return backend_map.get(backend_name.lower())
-    except AttributeError:
-        # torch.nn.attention.SDPBackend not available in older torch versions
-        return None
-
-
-def flush_memory():
-    """Flush GPU memory and run garbage collection."""
-    gc.collect()
-    # Dynamo resets
-    torch._dynamo.reset()
-    torch._dynamo.reset_code_caches()
-    if hasattr(torch._inductor, "codecache"):
-        # Clear FX graph cache
-        if hasattr(torch._inductor.codecache, "FxGraphCache"):
-            torch._inductor.codecache.FxGraphCache.clear()
-        # Clear PyCodeCache
-        if hasattr(torch._inductor.codecache, "PyCodeCache"):
-            torch._inductor.codecache.PyCodeCache.cache_clear()
-        # Clear TritonFuture cache (for async compilation)
-        if hasattr(torch._inductor.codecache, "TritonFuture"):
-            if hasattr(torch._inductor.codecache.TritonFuture, "_compile_cache"):
-                torch._inductor.codecache.TritonFuture._compile_cache.clear()
-    # Clear CUDA cache
-    if torch.cuda.is_available():
-        torch.cuda.empty_cache()
-        torch.cuda.reset_max_memory_allocated()
-        torch.cuda.reset_peak_memory_stats()
-        torch.cuda.synchronize()
-    gc.collect()
-
-
-class BenchmarkStreamer(BaseStreamer):
-    def __init__(self, **kwargs) -> None:
-        self.timeout = kwargs.pop("timeout", 10)
-        self.timestamps = []
-        self.text_queue = Queue()
-        self.stop_signal = None
-
-    def put(self, value):
-        """Receives tokens and logs the timestamp of the generation."""
-        self.timestamps.append(time.perf_counter())
-        self.text_queue.put(value)
-
-    def end(self):
-        self.timestamps.append(time.perf_counter())
-        self.text_queue.put(self.stop_signal)
-
-    def __iter__(self):
-        return self
-
-    def __next__(self):
-        value = self.text_queue.get(timeout=self.timeout)
-        if value == self.stop_signal:
-            raise StopIteration()
-        else:
-            return value
-
-
-class BenchmarkRunner:
-    """Main benchmark runner that coordinates benchmark execution."""
-
-    def __init__(
-        self,
-        logger: logging.Logger,
-        output_dir: str | None = None,
-        branch_name: str | None = None,
-        commit_id: str | None = None,
-        commit_message: str | None = None,
-    ) -> None:
-        # Those stay constant for the whole run
-        self.logger = logger
-        if output_dir is None:
-            output_dir = os.path.join(os.path.dirname(os.path.dirname(__file__)), "benchmark_results")
-        self.output_dir = output_dir
-        self.branch_name = branch_name
-        self.commit_id = get_git_revision() if commit_id is None else commit_id
-        self.commit_message = commit_message
-        os.makedirs(self.output_dir, exist_ok=True)
-        self.profile_dir = None
-        # Attributes that are reset for each model
-        self._setup_for = ""
-        # Attributes that are reset for each run
-        self.model: GenerationMixin | None = None
-
-    def cleanup(self) -> None:
-        del self.model
-        self.model = None
-        flush_memory()
-
-    def setup_benchmark(self, model_id: str, config: BenchmarkConfig) -> None:
-        # Some attributes only need to be set once per model
-        if self._setup_for != model_id:
-            self.tokenizer = AutoTokenizer.from_pretrained(model_id)
-            # We set the EOS token to the padding token for open-ended generation
-            self.tokenizer.eos_token = self.tokenizer.pad_token
-            self._setup_for = model_id
-
-        # Prepare inputs
-        self.inputs = self.tokenizer(
-            [DEFAULT_PROMPT for _ in range(config.batch_size)],
-            return_tensors="pt",
-            max_length=config.sequence_length,
-            truncation=True,
-            return_attention_mask=True,
-        ).to(config.device)
-        self.inputs["use_cache"] = True
-
-        # Prepare generation config
-        gen_config = GenerationConfig(
-            do_sample=False, top_p=1.0, temperature=1.0, max_new_tokens=config.num_tokens_to_generate
-        )
-
-        # Prepare compile config
-        if config.compile_mode is not None:
-            gen_config.compile_config = CompileConfig(mode=config.compile_mode, options=config.compile_options)
-            gen_config.cache_implementation = "static"
-
-        # Load model
-        self.logger.debug(f"Loading model {model_id} on device {config.device}...")
-        dtype = getattr(torch, config.dtype.removeprefix("torch."))
-        self.model = AutoModelForCausalLM.from_pretrained(
-            model_id, dtype=dtype, attn_implementation=config.attn_implementation, generation_config=gen_config
-        )
-        self.model = self.model.eval().to(config.device)
-
-        # Kernelize the model if needed
-        if config.kernelize and kernelize is not None and Mode is not None:
-            self.model = kernelize(self.model, mode=Mode.INFERENCE)
-
-    def run_benchmark(
-        self, model_id: str, config: BenchmarkConfig, num_tokens_to_profile: int = 0
-    ) -> dict[str, Any] | None:
-        """Run a single benchmark with the given model ID and config."""
-        sdpa_ctx = nullcontext()
-        if config.attn_implementation == "sdpa":
-            sdpa_backend = get_sdpa_backend(config.sdpa_backend)
-            sdpa_ctx = torch.nn.attention.sdpa_kernel(sdpa_backend)
-
-        with sdpa_ctx, torch.no_grad():
-            self.logger.info(f"Running benchmark scenario: {config.name}")
-
-            # Quick validation: try one measurement first to see if this scenario works
-            flush_memory()
-            e2e_latency, token_generation_times, shape_and_decoded_output, gpu_metrics = self.time_generate(
-                max_new_tokens=1, gpu_monitor=None
-            )
-            if e2e_latency < 0:
-                self.logger.warning(f"Skipping config {config.name}: {e2e_latency = } (no GPU monitoring)")
-                return None
-
-            # Warmup runs
-            self.logger.info(f"Warming up with {config.warmup_iterations} iterations...")
-            for _ in trange(config.warmup_iterations):
-                _ = self.time_generate(max_new_tokens=config.num_tokens_to_generate)
-            self.logger.info("Warmup over.")
-
-            # Measurement runs
-            result = BenchmarkResult()
-            self.logger.info(f"Benchmarking with {config.measurement_iterations} iterations.")
-            for _ in trange(config.measurement_iterations):
-                e2e_latency, token_generation_times, shape_and_decoded_output, gpu_metrics = self.time_generate(
-                    max_new_tokens=config.num_tokens_to_generate,
-                    gpu_monitor=(GPUMonitor(logger=self.logger) if config.gpu_monitoring else None),
-                )
-                result.accumulate(e2e_latency, token_generation_times, shape_and_decoded_output, gpu_metrics)
-            self.logger.info("Benchmarking done. Cleaning up.")
-
-            # Profile if needed
-            if num_tokens_to_profile > 0:
-                self.profile_generate(num_tokens_to_profile, config.name)
-
-            return {
-                "metadata": BenchmarkMetadata(
-                    model_id=model_id,
-                    branch_name=self.branch_name,
-                    commit_id=self.commit_id,
-                    commit_message=self.commit_message,
-                ),
-                "measurements": result,
-                "config": config,
-            }
-
-    def time_generate(
-        self,
-        max_new_tokens: int,
-        gpu_monitor: GPUMonitor | None = None,
-    ) -> tuple[float, list[float], str, GPURawMetrics | None]:
-        """Time the latency of a call to model.generate() with the given (inputs) and (max_new_tokens)."""
-        # Prepare gpu monitoring if needed
-        if gpu_monitor is not None:
-            gpu_monitor.start()
-        # Prepare streamer
-        streamer = BenchmarkStreamer()
-        # Generate and time
-        wall_time_0 = time.perf_counter()
-        outputs = self.model.generate(
-            **self.inputs,
-            max_new_tokens=max_new_tokens,
-            streamer=streamer,
-        )
-        wall_time_1 = time.perf_counter()
-        # Stop gpu monitoring if needed
-        gpu_metrics = gpu_monitor.stop_and_collect() if gpu_monitor is not None else None
-        # Check if generation had the right number of tokens
-        input_tokens = self.inputs["input_ids"].size(-1)
-        batch_size, output_tokens = outputs.shape
-        new_tokens = output_tokens - input_tokens
-        if new_tokens != max_new_tokens:
-            raise RuntimeError(f"Generated {new_tokens} tokens, expected {max_new_tokens}")
-        # Decode outputs
-        decoded_output = self.tokenizer.decode(outputs[0, input_tokens:], skip_special_tokens=True)
-        shape_and_decoded_output = f"{tuple(outputs.shape)} | {decoded_output}"
-        # Compute intermediate quantities
-        e2e_latency = wall_time_1 - wall_time_0
-        token_generation_times = [t - wall_time_0 for t in streamer.timestamps[1:]]
-        return e2e_latency, token_generation_times, shape_and_decoded_output, gpu_metrics
-
-    def profile_generate(self, num_tokens_to_profile: int, config_name: str) -> None:
-        """Profile the latency of a call to model.generate() with the given (inputs) and (max_new_tokens)."""
-        profiler = torch.profiler.profile(
-            activities=[torch.profiler.ProfilerActivity.CPU, torch.profiler.ProfilerActivity.CUDA],
-            record_shapes=True,
-        )
-        with profiler as prof:
-            _ = self.model.generate(
-                **self.inputs,
-                max_new_tokens=num_tokens_to_profile,
-            )
-        if self.profile_dir is None:
-            self.profile_dir = self.output_dir + "_profiles"
-            os.makedirs(self.profile_dir, exist_ok=True)
-        prof.export_chrome_trace(f"{self.profile_dir}/{config_name}.json")
-
-    def run_benchmarks(
-        self,
-        model_id: str,
-        benchmark_configs: list[BenchmarkConfig],
-        num_tokens_to_profile: int = 0,
-        pretty_print_summary: bool = True,
-    ) -> tuple[str, dict[str, Any]]:
-        """Run multiple benchmarks for the given model ID and list of benchmark configs."""
-        all_results = {}
-        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
-        start_time = time.perf_counter()
-
-        n_configs = len(benchmark_configs)
-        for i, config in enumerate(benchmark_configs):
-            # Handle SDPA backend if not determined by the config (needs to be done before skipping duplicates)
-            if config.attn_implementation == "sdpa" and config.sdpa_backend is None:
-                default_backend = "flash_attention"  # FIXME: torch has a _cur_sdpa_kernel_backends but it fails
-                self.logger.warning(f"No SDPA backend provided, using {default_backend} instead.")
-                config.sdpa_backend = default_backend
-
-            # Skip if already run
-            if config.hash in all_results:
-                self.logger.info(f"Skipping duplicate config {config.name} for model {model_id} ({i + 1}/{n_configs})")
-                continue
-
-            # Otherwise, run the benchmark
-            self.setup_benchmark(model_id, config)
-            self.logger.info(
-                f"Running benchmark of model {model_id} with scenario: {config.name} ({i + 1}/{n_configs})"
-            )
-
-            # Launch benchmark in a try/except block to avoid stopping the whole run if one benchmark fails
-            try:
-                results = self.run_benchmark(model_id, config, num_tokens_to_profile)
-                if results is not None:
-                    all_results[config.hash] = results
-
-            except Exception as e:
-                self.logger.error(f"Error running with scenario: {config.name}:\n{repr(e)}")
-            # Cleanup model and save results
-            self.cleanup()
-            self.save_results(model_id, all_results, timestamp=timestamp)
-
-        if pretty_print_summary:
-            print()
-            print("=" * 100)
-            print(f"Finished benchmarks in {time.perf_counter() - start_time:.2f} seconds")
-            print(f"Total number of benchmarks: {len(all_results)}")
-            if len(all_results) > 0:
-                print("First run metadata:")
-                first_key = list(all_results.keys())[0]
-                first_metadata = all_results[first_key]["metadata"].to_dict()
-                hardware_info = first_metadata.pop("hardware_info")
-                pretty_print_dict(first_metadata | hardware_info, tabs=1)
-            for result in all_results.values():
-                print("=" * 100)
-                print(f"Config: {result['config'].infer_name(compact=False)}\n")
-                result["measurements"].pprint(batch_size=result["config"].batch_size, tabs=1)
-            print("=" * 100)
-
-        return (timestamp, all_results)
-
-    def save_results(self, model_name: str, results: dict, timestamp: str = "") -> str:
-        """Save benchmark results to JSON file."""
-        # Create model-specific subdirectory
-        model_name = model_name.replace("/", "_")
-        model_dir = os.path.join(self.output_dir, model_name)
-        os.makedirs(model_dir, exist_ok=True)
-
-        # Create filename with timestamp
-        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S") if not timestamp else timestamp
-        filename = f"{model_name}_benchmark_{timestamp}.json"
-        filepath = os.path.join(model_dir, filename)
-
-        # Convert results to dict
-        converted_results = {}
-        for cfg_hash in results.keys():
-            converted_results[cfg_hash] = {
-                "metadata": results[cfg_hash]["metadata"].to_dict(),
-                "measurements": results[cfg_hash]["measurements"].to_dict(),
-                "config": results[cfg_hash]["config"].to_dict(),
-            }
-
-        # Save to JSON file
-        with open(filepath, "w") as f:
-            f.write(compact_json_numeric_arrays(converted_results))
-
-        self.logger.info(f"Results saved to {filepath}")
-        return filepath
-
-    def push_results_to_hub(self, dataset_id: str, results: dict[Any, Any], timestamp: str) -> None:
-        if PUSH_TO_HUB_TOKEN is None:
-            raise ValueError(
-                "PUSH_TO_HUB_TOKEN is not set, cannot push results to the Hub. When setting dataset_id, please also set the PUSH_TO_HUB_TOKEN environment variable."
-            )
-
-        n_results = len(results)
-        self.logger.info(f"Pushing {n_results} results to: {dataset_id}")
-        rows = []
-        for cfg_hash, entry in results.items():
-            row = {
-                "benchmark_config_hash": cfg_hash,
-                "config": entry["config"].to_dict(),
-                "measurements": entry["measurements"].to_dict(),
-                "metadata": entry["metadata"].to_dict(),
-            }
-            rows.append(row)
-
-        ds = Dataset.from_list(rows)
-        with tempfile.TemporaryDirectory() as tmp:
-            jsonl_path = os.path.join(tmp, "data.jsonl")
-            with open(jsonl_path, "w") as f:
-                json_lines = []
-                for ex in ds:
-                    json_lines.append(json.dumps(ex, ensure_ascii=False))
-                f.write("\n".join(json_lines))
-
-            api = HfApi()
-            # NOTE: we expect the repository to already exist
-            timestamp = datetime.now().strftime("%Y%m%d_%H%M%S") if not timestamp else timestamp
-            file_name = f"benchmark_run_{timestamp}.jsonl"
-            api.upload_file(
-                path_or_fileobj=jsonl_path,
-                path_in_repo=file_name,
-                repo_id=dataset_id,
-                repo_type="dataset",
-                token=PUSH_TO_HUB_TOKEN,
-            )
-        self.logger.info(f"Succesfully uploaded results to: {dataset_id}")

benchmark_v2/framework/data_classes.py

@@ -1,167 +0,0 @@
-from dataclasses import dataclass
-from datetime import datetime, timezone
-from typing import Any
-
-import numpy as np
-
-from .hardware_metrics import GPURawMetrics, HardwareInfo
-
-
-def compute_basic_statistics(measurements: list[float]) -> dict[str, float]:
-    return {
-        "avg": np.mean(measurements),
-        "std": np.std(measurements),
-        "min": np.min(measurements),
-        "med": np.median(measurements),
-        "max": np.max(measurements),
-        "p95": np.percentile(measurements, 95),
-    }
-
-
-def add_unit_to_duration(stats: dict[str, float]) -> dict[str, str]:
-    for key in list(stats.keys()):
-        value = stats[key]
-        if value > 3600:
-            stats[key] = f"{(value / 3600):.2f}hr"
-        elif value > 60:
-            stats[key] = f"{(value / 60):.2f}min"
-        elif value > 1:
-            stats[key] = f"{value:.2f}s"
-        elif value > 1e-3:
-            stats[key] = f"{(value * 1e3):.2f}ms"
-        elif value > 1e-6:
-            stats[key] = f"{(value * 1e6):.2f}us"
-        else:
-            stats[key] = f"{(value * 1e9):.2f}ns"
-    return stats
-
-
-def equalize_lengths_and_collate(stats: list[dict[str, str]]) -> list[str]:
-    keys = ["avg", "std", "min", "med", "max", "p95"]
-    for key in keys:
-        max_length = max(len(stat[key]) for stat in stats)
-        for stat in stats:
-            stat[key] = stat[key].ljust(max_length, " ")
-    return [" ".join([f"{key}={stat[key]}" for key in keys]) for stat in stats]
-
-
-def pretty_print_dict(data: dict[str, Any], tabs: int = 0) -> None:
-    max_key_length = max([len(key) for key in data.keys()])
-    for key, value in data.items():
-        tabs_str = "  " * tabs
-        padded_key = key.ljust(max_key_length + 1, ".")
-        print(f"{tabs_str}{padded_key}: {value}")
-
-
-@dataclass
-class BenchmarkMetadata:
-    """Metadata collected for each benchmark run."""
-
-    model_id: str
-    timestamp: str
-    branch_name: str
-    commit_id: str
-    commit_message: str
-    hardware_info: HardwareInfo
-
-    def __init__(self, model_id: str, commit_id: str, branch_name: str = "main", commit_message: str = "") -> None:
-        self.model_id = model_id
-        self.timestamp = datetime.now(timezone.utc).isoformat()
-        self.branch_name = branch_name
-        self.commit_id = commit_id
-        self.commit_message = commit_message
-        self.hardware_info = HardwareInfo()
-
-    def to_dict(self) -> dict[str, Any]:
-        return {
-            "model_id": self.model_id,
-            "timestamp": self.timestamp,
-            "branch_name": self.branch_name,
-            "commit_id": self.commit_id,
-            "commit_message": self.commit_message,
-            "hardware_info": self.hardware_info.to_dict(),
-        }
-
-
-class BenchmarkResult:
-    """Result from a series of benchmark runs."""
-
-    def __init__(self) -> None:
-        self.e2e_latency = []
-        self.token_generation_times = []  # time at which each token was generated (relative to start of the generation)
-        self.shape_and_decoded_outputs = []
-        self.gpu_metrics = []
-
-    def accumulate(
-        self,
-        e2e_latency: float,
-        token_generation_times: list[float],
-        shape_and_decoded_output: str,
-        gpu_metrics: GPURawMetrics | None,
-    ) -> None:
-        self.e2e_latency.append(e2e_latency)
-        self.token_generation_times.append(token_generation_times)
-        self.shape_and_decoded_outputs.append(shape_and_decoded_output)
-        self.gpu_metrics.append(gpu_metrics)
-
-    def to_dict(self) -> dict[str, None | int | float]:
-        # Save GPU metrics as None if it contains only None values
-        if all(gm is None for gm in self.gpu_metrics):
-            gpu_metrics = None
-        else:
-            gpu_metrics = [gm.to_dict() for gm in self.gpu_metrics]
-        return {
-            "e2e_latency": self.e2e_latency,
-            "token_generation_times": self.token_generation_times,
-            "shape_and_decoded_outputs": self.shape_and_decoded_outputs,
-            "gpu_metrics": gpu_metrics,
-        }
-
-    @classmethod
-    def from_dict(cls, data: dict[str, None | int | float]) -> "BenchmarkResult":
-        # Handle GPU metrics, which is saved as None if it contains only None values
-        if data["gpu_metrics"] is None:
-            gpu_metrics = [None for _ in range(len(data["e2e_latency"]))]
-        else:
-            gpu_metrics = [GPURawMetrics.from_dict(gm) for gm in data["gpu_metrics"]]
-        # Create a new instance and accumulate the data
-        new_instance = cls()
-        for i in range(len(data["e2e_latency"])):
-            new_instance.accumulate(
-                e2e_latency=data["e2e_latency"][i],
-                token_generation_times=data["token_generation_times"][i],
-                shape_and_decoded_output=data["shape_and_decoded_outputs"][i],
-                gpu_metrics=gpu_metrics[i],
-            )
-        return new_instance
-
-    def get_measured_ttft(self) -> list[float]:
-        return [dt[0] for dt in self.token_generation_times if len(dt) > 0]
-
-    def get_measured_itl(self) -> list[float]:
-        return [(dt[-1] - dt[0]) / (len(dt) - 1) for dt in self.token_generation_times if len(dt) > 1]
-
-    def get_throughput(self, batch_size: int) -> float:
-        return [
-            batch_size * len(dt) / e2e_latency
-            for e2e_latency, dt in zip(self.e2e_latency, self.token_generation_times)
-        ]
-
-    def pprint(self, batch_size: int = 0, tabs: int = 0) -> None:
-        stats_to_collate = [
-            add_unit_to_duration(compute_basic_statistics(self.e2e_latency)),
-            add_unit_to_duration(compute_basic_statistics(self.get_measured_ttft())),
-            add_unit_to_duration(compute_basic_statistics(self.get_measured_itl())),
-        ]
-        if batch_size > 0:
-            throughput_stats = compute_basic_statistics(self.get_throughput(batch_size))
-            stats_to_collate.append({key: f"{value:.2f}tok/s" for key, value in throughput_stats.items()})
-        collated_stats = equalize_lengths_and_collate(stats_to_collate)
-        dict_to_pprint = {
-            "E2E Latency": collated_stats[0],
-            "Time to First Token": collated_stats[1],
-            "Inter-Token Latency": collated_stats[2],
-        }
-        if batch_size > 0:
-            dict_to_pprint["Throughput"] = collated_stats[3]
-        pretty_print_dict(dict_to_pprint, tabs=tabs)

benchmark_v2/framework/hardware_metrics.py

@@ -1,171 +0,0 @@
-import json
-import logging
-import subprocess
-import sys
-import threading
-import time
-from dataclasses import dataclass
-from enum import Enum
-from logging import Logger
-
-import gpustat
-import psutil
-import torch
-
-
-# Data class to hold the hardware information
-def get_device_name_and_memory_total() -> tuple[str, float]:
-    """Returns the name and memory total of GPU 0."""
-    device_name = torch.cuda.get_device_properties(0).name
-    device_memory_total = torch.cuda.get_device_properties(0).total_memory / 1024**3
-    return device_name, device_memory_total
-
-
-class HardwareInfo:
-    """A class to hold information about the hardware."""
-
-    def __init__(self) -> None:
-        # Retrieve GPU stats
-        try:
-            self.gpu_name, self.gpu_memory_total_gb = get_device_name_and_memory_total()
-        except Exception:
-            self.gpu_name, self.gpu_memory_total_gb = None, None
-        # Retrieve python, torch and CUDA version
-        self.python_version = f"{sys.version.split()[0]}"
-        self.torch_version = torch.__version__
-        if hasattr(torch, "cuda") and torch.cuda.is_available():
-            self.cuda_version = torch.version.cuda
-        else:
-            self.cuda_version = None
-        # Retrieve general hardware information
-        self.cpu_count = psutil.cpu_count()
-        self.memory_total_mb = int(psutil.virtual_memory().total / (1024 * 1024))
-
-    def to_dict(self) -> dict[str, None | int | float | str]:
-        return {
-            "gpu_name": self.gpu_name,
-            "gpu_memory_total_gb": self.gpu_memory_total_gb,
-            "python_version": self.python_version,
-            "torch_version": self.torch_version,
-        }
-
-
-# Functions to get information about the GPU
-def get_amd_gpu_stats() -> tuple[int, float]:
-    """Returns the utilization and memory used of an AMD GPU, both in percent"""
-    rocm_smi_output = subprocess.check_output(["rocm-smi", "--json", "--showuse", "--showmeminfo", "VRAM"])
-    gpu_stats = json.loads(rocm_smi_output.decode("utf-8"))
-    gpu_stats = [
-        (card_id, stats["GPU use (%)"], stats["VRAM Total Used Memory (B)"]) for card_id, stats in gpu_stats.items()
-    ]
-    gpu_stats.sort(key=lambda x: x[1], reverse=True)
-    return int(gpu_stats[0][1]), float(gpu_stats[0][2]) / 1024**3
-
-
-def get_nvidia_gpu_stats() -> tuple[int, float]:
-    """Returns the utilization and memory used of an NVIDIA GPU, both in percent"""
-    gpu_stats = gpustat.GPUStatCollection.new_query()
-    gpu_stats = gpu_stats[0]
-    return int(gpu_stats["utilization.gpu"]), float(gpu_stats["memory.used"]) / 1024**3
-
-
-class GPUStatsCollector:
-    """A class to get statistics about the GPU. It serves as a wrapper that holds the GPU total memory and its name,
-    which is used to call the right function to get the utilization and memory used."""
-
-    def __init__(self) -> None:
-        self.device_name, self.device_memory_total = get_device_name_and_memory_total()
-        # Monkey patch the get_utilization_and_memory_used method based on the GPU type
-        if "amd" in self.device_name.lower():
-            self.get_utilization_and_memory_used = get_amd_gpu_stats
-        elif "nvidia" in self.device_name.lower():
-            self.get_utilization_and_memory_used = get_nvidia_gpu_stats
-        else:
-            raise RuntimeError(f"Unsupported GPU: {self.device_name}")
-
-    def get_measurements(self) -> tuple[int, float]:
-        """Get the utilization and memory used of the GPU, both in percent"""
-        raise NotImplementedError("This method is meant to be monkey patched during __init__")
-
-
-# Simple data classes to hold the raw GPU metrics
-class GPUMonitoringStatus(Enum):
-    """Status of GPU monitoring."""
-
-    SUCCESS = "success"
-    FAILED = "failed"
-    NO_GPUS_AVAILABLE = "no_gpus_available"
-    NO_SAMPLES_COLLECTED = "no_samples_collected"
-
-
-@dataclass
-class GPURawMetrics:
-    """Raw values for GPU utilization and memory used."""
-
-    utilization: list[float]  # in percent
-    memory_used: list[float]  # in GB
-    timestamps: list[float]  # in seconds
-    timestamp_0: float  # in seconds
-    monitoring_status: GPUMonitoringStatus
-
-    def to_dict(self) -> dict[str, None | int | float | str]:
-        return {
-            "utilization": self.utilization,
-            "memory_used": self.memory_used,
-            "timestamps": self.timestamps,
-            "timestamp_0": self.timestamp_0,
-            "monitoring_status": self.monitoring_status.value,
-        }
-
-
-# Main class, used to monitor the GPU utilization during benchmark execution
-class GPUMonitor:
-    """Monitor GPU utilization during benchmark execution."""
-
-    def __init__(self, sample_interval_sec: float = 0.1, logger: Logger | None = None):
-        self.sample_interval_sec = sample_interval_sec
-        self.logger = logger if logger is not None else logging.getLogger(__name__)
-
-        self.num_available_gpus = torch.cuda.device_count()
-        if self.num_available_gpus == 0:
-            raise RuntimeError("No GPUs detected by torch.cuda.device_count().")
-        self.gpu_stats_getter = GPUStatsCollector()
-
-    def start(self):
-        """Start monitoring GPU metrics."""
-        # Clear the stop event to enable monitoring
-        self.stop_event = threading.Event()
-        self.gpu_utilization = []
-        self.gpu_memory_used = []
-        self.timestamps = []
-        self.thread = threading.Thread(target=self._monitor_loop)
-        self.thread.start()
-        self.logger.debug("GPU monitoring started")
-
-    def stop_and_collect(self) -> GPURawMetrics:
-        """Stop monitoring and return collected metrics."""
-        self.stop_event.set()
-        self.thread.join()
-        if self.gpu_utilization:
-            timestamp_0 = self.timestamps[0]
-            metrics = GPURawMetrics(
-                utilization=self.gpu_utilization,
-                memory_used=self.gpu_memory_used,
-                timestamps=[t - timestamp_0 for t in self.timestamps],
-                timestamp_0=timestamp_0,
-                monitoring_status=GPUMonitoringStatus.SUCCESS,
-            )
-            self.logger.debug(f"GPU monitoring completed: {len(self.gpu_utilization)} samples collected")
-        else:
-            metrics = GPURawMetrics(monitoring_status=GPUMonitoringStatus.NO_SAMPLES_COLLECTED)
-        return metrics
-
-    def _monitor_loop(self):
-        """Background monitoring loop using threading.Event for communication."""
-        while not self.stop_event.is_set():
-            utilization, memory_used = self.gpu_stats_getter.get_utilization_and_memory_used()
-            self.gpu_utilization.append(utilization)
-            self.gpu_memory_used.append(memory_used)
-            self.timestamps.append(time.time())
-            if self.stop_event.wait(timeout=self.sample_interval_sec):
-                break

benchmark_v2/requirements.txt

@@ -4,4 +4,4 @@ gpustat>=1.0.0
 torch>=2.0.0
 transformers>=4.30.0
 datasets>=2.10.0
-huggingface_hub>=0.16.0
+huggingface_hub>=0.16.0 

benchmark_v2/run_benchmarks.py

@@ -19,124 +19,477 @@ in the ./benches directory, organizing outputs into model-specific subfolders.
 """
 
 import argparse
+import importlib.util
+import json
 import logging
+import os
 import sys
 import uuid
-
-from framework.benchmark_config import BenchmarkConfig, generate_all_configs, generate_main_configs
-from framework.benchmark_runner import BenchmarkRunner
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Optional
 
 
-if __name__ == "__main__":
-    # Parse arguments
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--output-dir", type=str, default=None, help="Output dir for benchmark results")
-    parser.add_argument("--log-level", type=str, choices=["DEBUG", "INFO", "WARNING", "ERROR"], default="INFO")
-    parser.add_argument("--model-id", type=str, help="Specific model ID to benchmark (if supported by benchmarks)")
-    parser.add_argument("--warmup", "-w", type=int, default=3, help="Number of warmup iterations")
-    parser.add_argument("--iterations", "-i", type=int, default=10, help="Number of measurement iterations")
-
-    parser.add_argument("--batch-size", "-b", type=int, nargs="+", help="Batch size")
-    parser.add_argument("--sequence-length", "-s", type=int, nargs="+", help="Sequence length")
-    parser.add_argument("--num-tokens-to-generate", "-n", type=int, nargs="+", help="Number of tokens to generate")
-
-    parser.add_argument("--cross-generate", action="store_true", help="Cross-generate all combinations of configs")
-    parser.add_argument("--num-tokens-to-profile", "-p", type=int, default=0, help="Number of tokens to profile")
-
-    parser.add_argument("--branch-name", type=str, help="Git branch name")
-    parser.add_argument("--commit-id", type=str, help="Git commit ID (if not provided, will auto-detect from git)")
-    parser.add_argument("--commit-message", type=str, help="Git commit message")
-
-    parser.add_argument(
-        "--no-gpu-monitoring", action="store_true", help="Disables GPU monitoring during benchmark runs"
-    )
-
-    parser.add_argument(
-        "--push-result-to-dataset",
-        type=str,
-        default=None,
-        help="Name of the dataset to push results to. If not provided, results are not pushed to the Hub.",
-    )
-    args = parser.parse_args()
-
-    # Setup logging
-    benchmark_run_uuid = str(uuid.uuid4())[:8]
-    numeric_level = getattr(logging, args.log_level.upper())
+def setup_logging(log_level: str = "INFO", enable_file_logging: bool = False) -> logging.Logger:
+    """Setup logging configuration."""
+    numeric_level = getattr(logging, log_level.upper(), None)
+    if not isinstance(numeric_level, int):
+        raise ValueError(f"Invalid log level: {log_level}")
 
     handlers = [logging.StreamHandler(sys.stdout)]
+
+    if enable_file_logging:
+        handlers.append(logging.FileHandler(f"benchmark_run_{datetime.now().strftime('%Y%m%d_%H%M%S')}.log"))
+
     logging.basicConfig(
         level=numeric_level, format="[%(levelname)s - %(asctime)s] %(name)s: %(message)s", handlers=handlers
     )
 
-    logger = logging.getLogger("benchmark_v2")
+    return logging.getLogger(__name__)
+
+
+def discover_benchmarks(benches_dir: str) -> list[dict[str, Any]]:
+    """
+    Discover all benchmark modules in the benches directory.
+
+    Returns:
+        List of dictionaries containing benchmark module info
+    """
+    benchmarks = []
+    benches_path = Path(benches_dir)
+
+    if not benches_path.exists():
+        raise FileNotFoundError(f"Benches directory not found: {benches_dir}")
+
+    for py_file in benches_path.glob("*.py"):
+        if py_file.name.startswith("__"):
+            continue
+
+        module_name = py_file.stem
+
+        try:
+            # Import the module
+            spec = importlib.util.spec_from_file_location(module_name, py_file)
+            module = importlib.util.module_from_spec(spec)
+            spec.loader.exec_module(module)
+
+            # Check if it has a benchmark runner function
+            if hasattr(module, f"run_{module_name}"):
+                benchmarks.append(
+                    {
+                        "name": module_name,
+                        "path": str(py_file),
+                        "module": module,
+                        "runner_function": getattr(module, f"run_{module_name}"),
+                    }
+                )
+            elif hasattr(module, "run_benchmark"):
+                benchmarks.append(
+                    {
+                        "name": module_name,
+                        "path": str(py_file),
+                        "module": module,
+                        "runner_function": getattr(module, "run_benchmark"),
+                    }
+                )
+            else:
+                logging.warning(f"No runner function found in {py_file}")
+
+        except Exception as e:
+            logging.error(f"Failed to import {py_file}: {e}")
+
+    return benchmarks
+
+
+def run_single_benchmark(
+    benchmark_info: dict[str, Any], output_dir: str, logger: logging.Logger, **kwargs
+) -> Optional[str]:
+    """
+    Run a single benchmark and return the output file path.
+
+    Args:
+        benchmark_info: Dictionary containing benchmark module info
+        output_dir: Base output directory
+        logger: Logger instance
+        **kwargs: Additional arguments to pass to the benchmark
+
+    Returns:
+        Path to the output file if successful, None otherwise
+    """
+    benchmark_name = benchmark_info["name"]
+    runner_func = benchmark_info["runner_function"]
+
+    logger.info(f"Running benchmark: {benchmark_name}")
+
+    try:
+        # Check function signature to determine what arguments to pass
+        import inspect
+
+        sig = inspect.signature(runner_func)
+
+        # Prepare arguments based on function signature
+        func_kwargs = {"logger": logger, "output_dir": output_dir}
+
+        # Add other kwargs if the function accepts them
+        for param_name in sig.parameters:
+            if param_name in kwargs:
+                func_kwargs[param_name] = kwargs[param_name]
+
+        # Filter kwargs to only include parameters the function accepts
+        # If function has **kwargs, include all provided kwargs
+        has_var_kwargs = any(param.kind == param.VAR_KEYWORD for param in sig.parameters.values())
+        if has_var_kwargs:
+            valid_kwargs = {**func_kwargs, **kwargs}
+        else:
+            valid_kwargs = {k: v for k, v in func_kwargs.items() if k in sig.parameters}
+
+        # Run the benchmark
+        result = runner_func(**valid_kwargs)
+
+        if isinstance(result, str):
+            # Function returned a file path
+            return result
+        else:
+            logger.info(f"Benchmark {benchmark_name} completed successfully")
+            return "completed"
+
+    except Exception as e:
+        logger.error(f"Benchmark {benchmark_name} failed: {e}")
+        import traceback
+
+        logger.debug(traceback.format_exc())
+        return None
+
+
+def generate_summary_report(
+    output_dir: str,
+    benchmark_results: dict[str, Any],
+    logger: logging.Logger,
+    benchmark_run_uuid: Optional[str] = None,
+) -> str:
+    """Generate a summary report of all benchmark runs."""
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    summary_file = os.path.join(output_dir, f"benchmark_summary_{timestamp}.json")
+
+    summary_data = {
+        "run_metadata": {
+            "timestamp": datetime.utcnow().isoformat(),
+            "benchmark_run_uuid": benchmark_run_uuid,
+            "total_benchmarks": len(benchmark_results),
+            "successful_benchmarks": len([r for r in benchmark_results.values() if r is not None]),
+            "failed_benchmarks": len([r for r in benchmark_results.values() if r is None]),
+        },
+        "benchmark_results": benchmark_results,
+        "output_directory": output_dir,
+    }
+
+    with open(summary_file, "w") as f:
+        json.dump(summary_data, f, indent=2, default=str)
+
+    logger.info(f"Summary report saved to: {summary_file}")
+    return summary_file
+
+
+def upload_results_to_hf_dataset(
+    output_dir: str,
+    summary_file: str,
+    dataset_name: str,
+    run_id: Optional[str] = None,
+    token: Optional[str] = None,
+    logger: Optional[logging.Logger] = None,
+) -> Optional[str]:
+    """
+    Upload benchmark results to a HuggingFace Dataset.
+    Based on upload_collated_report() from utils/collated_reports.py
+    Args:
+        output_dir: Local output directory containing results
+        summary_file: Path to the summary file
+        dataset_name: Name of the HuggingFace dataset to upload to
+        run_id: Unique run identifier (if None, will generate one)
+        token: HuggingFace token for authentication (if None, will use environment variables)
+        logger: Logger instance
+    Returns:
+        The run_id used for the upload, None if upload failed
+    """
+    if logger is None:
+        logger = logging.getLogger(__name__)
+
+    import os
+
+    from huggingface_hub import HfApi
+
+    api = HfApi()
+
+    if run_id is None:
+        github_run_number = os.getenv("GITHUB_RUN_NUMBER")
+        github_run_id = os.getenv("GITHUB_RUN_ID")
+        if github_run_number and github_run_id:
+            run_id = f"{github_run_number}-{github_run_id}"
+
+    date_folder = datetime.now().strftime("%Y-%m-%d")
+
+    github_event_name = os.getenv("GITHUB_EVENT_NAME")
+    if github_event_name != "schedule":
+        # Non-scheduled runs go under a runs subfolder
+        repo_path = f"{date_folder}/runs/{run_id}/benchmark_results"
+    else:
+        # Scheduled runs go directly under the date
+        repo_path = f"{date_folder}/{run_id}/benchmark_results"
+
+    logger.info(f"Uploading benchmark results to dataset '{dataset_name}' at path '{repo_path}'")
+
+    try:
+        # Upload all files in the output directory
+        from pathlib import Path
+
+        output_path = Path(output_dir)
+
+        for file_path in output_path.rglob("*"):
+            if file_path.is_file():
+                # Calculate relative path from output_dir
+                relative_path = file_path.relative_to(output_path)
+                path_in_repo = f"{repo_path}/{relative_path}"
+
+                logger.debug(f"Uploading {file_path} to {path_in_repo}")
+
+                api.upload_file(
+                    path_or_fileobj=str(file_path),
+                    path_in_repo=path_in_repo,
+                    repo_id=dataset_name,
+                    repo_type="dataset",
+                    token=token,
+                    commit_message=f"Upload benchmark results for run {run_id}",
+                )
+
+        logger.info(
+            f"Successfully uploaded results to: https://huggingface.co/datasets/{dataset_name}/tree/main/{repo_path}"
+        )
+
+        return run_id
+
+    except Exception as upload_error:
+        logger.error(f"Failed to upload results: {upload_error}")
+        import traceback
+
+        logger.debug(traceback.format_exc())
+        return None
+
+
+def main():
+    """Main entry point for the benchmarking script."""
+    # Generate a unique UUID for this benchmark run
+    benchmark_run_uuid = str(uuid.uuid4())[:8]
+
+    parser = argparse.ArgumentParser(
+        description="Run all benchmarks in the ./benches directory",
+        epilog="""
+Examples:
+  # Run all available benchmarks
+  python3 run_benchmarks.py
+  
+  # Run with specific model and upload to HuggingFace Dataset
+  python3 run_benchmarks.py --model-id meta-llama/Llama-2-7b-hf --upload-to-hf username/benchmark-results
+  
+  # Run with custom run ID and upload to HuggingFace Dataset
+  python3 run_benchmarks.py --run-id experiment_v1 --upload-to-hf org/benchmarks
+  
+  # Run only specific benchmarks with file logging
+  python3 run_benchmarks.py --include llama --enable-file-logging
+        """,  # noqa: W293
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default="benchmark_results",
+        help="Base output directory for benchmark results (default: benchmark_results)",
+    )
+
+    parser.add_argument(
+        "--benches-dir",
+        type=str,
+        default="./benches",
+        help="Directory containing benchmark implementations (default: ./benches)",
+    )
+
+    parser.add_argument(
+        "--log-level",
+        type=str,
+        choices=["DEBUG", "INFO", "WARNING", "ERROR"],
+        default="INFO",
+        help="Logging level (default: INFO)",
+    )
+
+    parser.add_argument("--model-id", type=str, help="Specific model ID to benchmark (if supported by benchmarks)")
+
+    parser.add_argument("--warmup-iterations", type=int, default=3, help="Number of warmup iterations (default: 3)")
+
+    parser.add_argument(
+        "--measurement-iterations", type=int, default=5, help="Number of measurement iterations (default: 5)"
+    )
+
+    parser.add_argument(
+        "--num-tokens-to-generate",
+        type=int,
+        default=100,
+        help="Number of tokens to generate in benchmarks (default: 100)",
+    )
+
+    parser.add_argument("--include", type=str, nargs="*", help="Only run benchmarks matching these names")
+
+    parser.add_argument("--exclude", type=str, nargs="*", help="Exclude benchmarks matching these names")
+
+    parser.add_argument("--enable-file-logging", action="store_true", help="Enable file logging (disabled by default)")
+
+    parser.add_argument(
+        "--commit-id", type=str, help="Git commit ID for metadata (if not provided, will auto-detect from git)"
+    )
+
+    parser.add_argument(
+        "--push-to-hub",
+        type=str,
+        help="Upload results to HuggingFace Dataset (provide dataset name, e.g., 'username/benchmark-results')",
+    )
+
+    parser.add_argument(
+        "--run-id", type=str, help="Custom run ID for organizing results (if not provided, will generate a unique ID)"
+    )
+
+    parser.add_argument(
+        "--token",
+        type=str,
+        help="HuggingFace token for dataset uploads (if not provided, will use HF_TOKEN environment variable)",
+    )
+
+    args = parser.parse_args()
+
+    # Setup logging
+    logger = setup_logging(args.log_level, args.enable_file_logging)
+
     logger.info("Starting benchmark discovery and execution")
     logger.info(f"Benchmark run UUID: {benchmark_run_uuid}")
     logger.info(f"Output directory: {args.output_dir}")
+    logger.info(f"Benches directory: {args.benches_dir}")
 
-    # Error out if one of the arguments is not provided
-    if len(args.batch_size) * len(args.sequence_length) * len(args.num_tokens_to_generate) == 0:
-        raise ValueError(
-            "At least one of the arguments --batch-size, --sequence-length, or --num-tokens-to-generate is required"
-        )
+    # Create output directory
+    os.makedirs(args.output_dir, exist_ok=True)
 
-    # If there is only one (batch_size, sequence_length, num_tokens_to_generate), we benchmark across configs
-    elif len(args.batch_size) * len(args.sequence_length) * len(args.num_tokens_to_generate) == 1:
-        if args.cross_generate:
-            benchmark_configs = generate_all_configs(
-                warmup_iterations=args.warmup,
-                measurement_iterations=args.iterations,
-                batch_size=args.batch_size[0],
-                sequence_length=args.sequence_length[0],
-                num_tokens_to_generate=args.num_tokens_to_generate[0],
-                gpu_monitoring=not args.no_gpu_monitoring,
+    try:
+        # Discover benchmarks
+        benchmarks = discover_benchmarks(args.benches_dir)
+        logger.info(f"Discovered {len(benchmarks)} benchmark(s): {[b['name'] for b in benchmarks]}")
+
+        if not benchmarks:
+            logger.warning("No benchmarks found!")
+            return 1
+
+        # Filter benchmarks based on include/exclude
+        filtered_benchmarks = benchmarks
+
+        if args.include:
+            filtered_benchmarks = [
+                b for b in filtered_benchmarks if any(pattern in b["name"] for pattern in args.include)
+            ]
+            logger.info(f"Filtered to include: {[b['name'] for b in filtered_benchmarks]}")
+
+        if args.exclude:
+            filtered_benchmarks = [
+                b for b in filtered_benchmarks if not any(pattern in b["name"] for pattern in args.exclude)
+            ]
+            logger.info(f"After exclusion: {[b['name'] for b in filtered_benchmarks]}")
+
+        if not filtered_benchmarks:
+            logger.warning("No benchmarks remaining after filtering!")
+            return 1
+
+        # Prepare common kwargs for benchmarks
+        benchmark_kwargs = {
+            "warmup_iterations": args.warmup_iterations,
+            "measurement_iterations": args.measurement_iterations,
+            "num_tokens_to_generate": args.num_tokens_to_generate,
+        }
+
+        if args.model_id:
+            benchmark_kwargs["model_id"] = args.model_id
+
+        # Add commit_id if provided
+        if args.commit_id:
+            benchmark_kwargs["commit_id"] = args.commit_id
+
+        # Run benchmarks
+        benchmark_results = {}
+        successful_count = 0
+
+        for benchmark_info in filtered_benchmarks:
+            result = run_single_benchmark(benchmark_info, args.output_dir, logger, **benchmark_kwargs)
+
+            benchmark_results[benchmark_info["name"]] = result
+
+            if result is not None:
+                successful_count += 1
+
+        # Generate summary report
+        summary_file = generate_summary_report(args.output_dir, benchmark_results, logger, benchmark_run_uuid)
+
+        # Upload results to HuggingFace Dataset if requested
+        upload_run_id = None
+        if args.push_to_hub:
+            logger.info("=" * 60)
+            logger.info("UPLOADING TO HUGGINGFACE DATASET")
+            logger.info("=" * 60)
+            # Use provided run_id or fallback to benchmark run UUID
+            effective_run_id = args.run_id or benchmark_run_uuid
+            upload_run_id = upload_results_to_hf_dataset(
+                output_dir=args.output_dir,
+                summary_file=summary_file,
+                dataset_name=args.push_to_hub,
+                run_id=effective_run_id,
+                token=args.token,
+                logger=logger,
             )
+            if upload_run_id:
+                logger.info(f"Upload completed with run ID: {upload_run_id}")
+            else:
+                logger.warning("Upload failed - continuing with local results")
+
+        # Final summary
+        total_benchmarks = len(filtered_benchmarks)
+        failed_count = total_benchmarks - successful_count
+
+        logger.info("=" * 60)
+        logger.info("BENCHMARK RUN SUMMARY")
+        logger.info("=" * 60)
+        logger.info(f"Total benchmarks: {total_benchmarks}")
+        logger.info(f"Successful: {successful_count}")
+        logger.info(f"Failed: {failed_count}")
+        logger.info(f"Output directory: {args.output_dir}")
+        logger.info(f"Summary report: {summary_file}")
+
+        if args.push_to_hub:
+            if upload_run_id:
+                logger.info(f"HuggingFace Dataset: {args.push_to_hub}")
+                logger.info(f"Run ID: {upload_run_id}")
+                logger.info(
+                    f"View results: https://huggingface.co/datasets/{args.push_to_hub}/tree/main/{datetime.now().strftime('%Y-%m-%d')}/runs/{upload_run_id}"
+                )
+            else:
+                logger.warning("Upload to HuggingFace Dataset failed")
+
+        if failed_count > 0:
+            logger.warning(f"{failed_count} benchmark(s) failed. Check logs for details.")
+            return 1
         else:
-            benchmark_configs = generate_main_configs(
-                warmup_iterations=args.warmup,
-                measurement_iterations=args.iterations,
-                batch_size=args.batch_size[0],
-                sequence_length=args.sequence_length[0],
-                num_tokens_to_generate=args.num_tokens_to_generate[0],
-            )
+            logger.info("All benchmarks completed successfully!")
+            return 0
 
-    # Otherwise, we benchmark across all combinations of dimensions
-    else:
-        main_config = generate_main_configs(
-            warmup_iterations=args.warmup,
-            measurement_iterations=args.iterations,
-            batch_size=args.batch_size[0],
-            sequence_length=args.sequence_length[0],
-            num_tokens_to_generate=args.num_tokens_to_generate[0],
-        )[0]
-        benchmark_configs = []
-        for num_tokens_to_generate in args.num_tokens_to_generate:
-            for sequence_length in args.sequence_length:
-                for batch_size in args.batch_size:
-                    cfg_dict = main_config.to_dict()
-                    cfg_dict["batch_size"] = batch_size
-                    cfg_dict["sequence_length"] = sequence_length
-                    cfg_dict["num_tokens_to_generate"] = num_tokens_to_generate
-                    cfg_dict.pop("name")
-                    benchmark_configs.append(BenchmarkConfig.from_dict(cfg_dict))
+    except Exception as e:
+        logger.error(f"Benchmark run failed: {e}")
+        import traceback
 
-    runner = BenchmarkRunner(
-        logger,
-        args.output_dir,
-        args.branch_name,
-        args.commit_id,
-        args.commit_message,
-    )
-    timestamp, results = runner.run_benchmarks(
-        args.model_id,
-        benchmark_configs,
-        args.num_tokens_to_profile,
-        pretty_print_summary=True,
-    )
+        logger.debug(traceback.format_exc())
+        return 1
 
-    dataset_id = args.push_result_to_dataset
-    if dataset_id is not None and len(results) > 0:
-        runner.push_results_to_hub(
-            dataset_id,
-            results,
-            timestamp,
-        )
+
+if __name__ == "__main__":
+    sys.exit(main())

conftest.py

@@ -54,10 +54,12 @@ NOT_DEVICE_TESTS = {
     "test_gradient_checkpointing_backward_compatibility",
     "test_gradient_checkpointing_enable_disable",
     "test_torch_save_load",
+    "test_initialization",
     "test_forward_signature",
     "test_model_get_set_embeddings",
     "test_model_main_input_name",
     "test_correct_missing_keys",
+    "test_tie_model_weights",
     "test_can_use_safetensors",
     "test_load_save_without_tied_weights",
     "test_tied_weights_keys",
@@ -87,10 +89,6 @@ def pytest_configure(config):
     config.addinivalue_line("markers", "not_device_test: mark the tests always running on cpu")
     config.addinivalue_line("markers", "torch_compile_test: mark test which tests torch compile functionality")
     config.addinivalue_line("markers", "torch_export_test: mark test which tests torch export functionality")
-    config.addinivalue_line("markers", "flash_attn_test: mark test which tests flash attention functionality")
-    config.addinivalue_line("markers", "flash_attn_3_test: mark test which tests flash attention 3 functionality")
-
-    os.environ["DISABLE_SAFETENSORS_CONVERSION"] = "true"
 
 
 def pytest_collection_modifyitems(items):

docker/transformers-all-latest-gpu/Dockerfile

@@ -12,6 +12,8 @@ SHELL ["sh", "-lc"]
 ARG PYTORCH='2.8.0'
 # Example: `cu102`, `cu113`, etc.
 ARG CUDA='cu126'
+# Disable kernel mapping for now until all tests pass
+ENV DISABLE_KERNEL_MAPPING=1
 
 RUN apt update
 RUN apt install -y git libsndfile1-dev tesseract-ocr espeak-ng python3 python3-pip ffmpeg git-lfs
@@ -24,8 +26,7 @@ RUN git clone https://github.com/huggingface/transformers && cd transformers &&
 # 1. Put several commands in a single `RUN` to avoid image/layer exporting issue. Could be revised in the future.
 # 2. Regarding `torch` part, We might need to specify proper versions for `torchvision` and `torchaudio`.
 #    Currently, let's not bother to specify their versions explicitly (so installed with their latest release versions).
-# 3. For `torchcodec<0.8`: this is quickly added as torch 2.9.0 + torchcodec 0.8.0 fails on our CI env. Need to remove later once they work.
-RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime] && [ ${#PYTORCH} -gt 0 -a "$PYTORCH" != "pre" ] && VERSION='torch=='$PYTORCH'.*' ||  VERSION='torch'; echo "export VERSION='$VERSION'" >> ~/.profile && echo torch=$VERSION && [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio "torchcodec<0.8" --extra-index-url https://download.pytorch.org/whl/$CUDA || python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
+RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime] && [ ${#PYTORCH} -gt 0 -a "$PYTORCH" != "pre" ] && VERSION='torch=='$PYTORCH'.*' ||  VERSION='torch'; echo "export VERSION='$VERSION'" >> ~/.profile && echo torch=$VERSION && [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/$CUDA || python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
 
 RUN python3 -m pip install --no-cache-dir -U timm
 

docker/transformers-pytorch-amd-gpu/Dockerfile

@@ -1,4 +1,4 @@
-FROM rocm/pytorch:rocm7.0.2_ubuntu24.04_py3.12_pytorch_release_2.7.1
+FROM rocm/pytorch:rocm6.4.1_ubuntu24.04_py3.12_pytorch_release_2.7.1
 LABEL maintainer="Hugging Face"
 
 ARG DEBIAN_FRONTEND=noninteractive
@@ -10,8 +10,8 @@ RUN apt update && \
 
 RUN git lfs install
 
-RUN python3 -m pip install --no-cache-dir --upgrade pip numpy importlib-metadata setuptools wheel ninja pytesseract "itsdangerous<2.1.0"
-RUN python3 -m pip install --no-cache-dir --no-build-isolation git+https://github.com/facebookresearch/detectron2.git
+RUN python3 -m pip install --no-cache-dir --upgrade pip numpy
+RUN python3 -m pip install --no-cache-dir --upgrade importlib-metadata setuptools ninja git+https://github.com/facebookresearch/detectron2.git pytesseract "itsdangerous<2.1.0"
 
 ARG REF=main
 WORKDIR /
@@ -39,7 +39,6 @@ RUN python3 -m pip install --no-cache-dir "torchcodec==0.5"
 # Install flash attention from source. Tested with commit 6387433156558135a998d5568a9d74c1778666d8
 RUN git clone https://github.com/ROCm/flash-attention/ -b tridao && \
     cd flash-attention && \
-    GPU_ARCHS="gfx942;gfx950" python setup.py install  
-# GPU_ARCHS builds for MI300, MI325 and MI355
+    GPU_ARCHS="gfx942" python setup.py install
 
 RUN python3 -m pip install --no-cache-dir einops

docker/transformers-pytorch-xpu/Dockerfile

@@ -3,10 +3,11 @@ LABEL maintainer="Hugging Face"
 
 SHELL ["/bin/bash", "-c"]
 
-ARG PYTHON_VER=3.12
+ARG PYTHON_VER=3.11
 ENV TORCH_DEVICE_BACKEND_AUTOLOAD=0
 ENV DEBIAN_FRONTEND=noninteractive
 
+RUN apt-get remove -y python3.10 && apt-get autoremove -y
 RUN apt-get update && \
     apt-get install -y software-properties-common && \
     add-apt-repository -y ppa:deadsnakes/ppa && \
@@ -22,6 +23,7 @@ RUN apt-get update && \
         apt-utils \
         build-essential \
         ca-certificates \
+        clinfo \
         curl \
         git \
         git-lfs \
@@ -33,6 +35,7 @@ RUN apt-get update && \
         rsync \
         sudo \
         libnl-genl-3-200 \
+        xpu-smi \
         unzip \
         ffmpeg \
         tesseract-ocr \
@@ -42,47 +45,34 @@ RUN apt-get update && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
 
+
 RUN apt-get update && \
     apt-get install -y \
-        linux-headers-$(uname -r) linux-modules-extra-$(uname -r) \
+        linux-headers-$(uname -r) \
+        linux-modules-extra-$(uname -r) \
         flex bison \
-        intel-fw-gpu intel-i915-dkms xpu-smi intel-ocloc clinfo\
+        intel-fw-gpu intel-i915-dkms xpu-smi \
         intel-opencl-icd libze-intel-gpu1 libze1 \
         intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
-        libegl-mesa0 libegl1 libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
+        libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
         libglapi-mesa libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
-        mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo \
+        mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo intel-ocloc \
         libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev && \
     apt-get clean && \
     rm -rf  /var/lib/apt/lists/*
 
-# Use virtual env because Ubuntu-24 does not allowed pip on original python
-RUN curl -LsSf https://astral.sh/uv/install.sh | sh
-ENV PATH="/root/.local/bin:$PATH"
-ENV VIRTUAL_ENV="/opt/venv"
-ENV UV_PYTHON_INSTALL_DIR=/opt/uv/python
-RUN uv venv --python ${PYTHON_VER} --seed ${VIRTUAL_ENV}
-ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+RUN pip install --upgrade pip
+RUN pip install triton==3.3.0
 
-RUN pip install --upgrade pip wheel
-RUN pip install triton==3.4.0
+RUN pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/xpu --no-cache-dir
 
-RUN pip install torch==2.8.0+xpu torchvision==0.23.0+xpu torchaudio==2.8.0+xpu --index-url https://download.pytorch.org/whl/xpu --no-cache-dir
+RUN pip install evaluate torchdata pyctcdecode pytesseract decord galore-torch fire scipy scikit-learn sentencepiece sacremoses nltk rouge_score librosa soundfile g2p_en mpi4py requests_mock
+RUN pip install pretty_midi essentia resampy Levenshtein av sacrebleu phonemizer invisible_watermark schedulefree
+RUN pip install gguf hqq compressed_tensors gptqmodel mergekit autoawq deepspeed torchao onnx
+RUN pip install hf_transfer huggingface-hub hf-doc-builder datasets optimum-quanto timm transformers accelerate optimum peft
 
-RUN pip install torchcodec torchdata --no-cache-dir
-
-RUN pip install evaluate pyctcdecode pytesseract decord galore-torch fire scipy scikit-learn sentencepiece sacremoses nltk rouge_score librosa soundfile g2p_en mpi4py requests_mock
-RUN pip install pretty_midi essentia resampy Levenshtein av sacrebleu phonemizer invisible_watermark schedulefree setuptools
-RUN pip install gptqmodel --no-build-isolation
-RUN pip install gguf hqq compressed_tensors autoawq deepspeed torchao onnx auto_round
-RUN pip install hf_transfer huggingface-hub hf-doc-builder datasets optimum-quanto timm transformers accelerate optimum peft diffusers trl kernels
-
-# install liger-kernel
 RUN pip install git+https://github.com/linkedin/Liger-Kernel.git --extra-index-url https://download.pytorch.org/whl/test/xpu
 
-# install mergekit
-RUN pip install --break-system-packages git+https://github.com/arcee-ai/mergekit.git@v0.1.3
-
 # install bitsandbytes
 RUN pip install git+https://github.com/bitsandbytes-foundation/bitsandbytes.git
 

docker/transformers-quantization-latest-gpu/Dockerfile

@@ -12,6 +12,8 @@ SHELL ["sh", "-lc"]
 ARG PYTORCH='2.8.0'
 # Example: `cu102`, `cu113`, etc.
 ARG CUDA='cu126'
+# Disable kernel mapping for quantization tests
+ENV DISABLE_KERNEL_MAPPING=1
 
 RUN apt update
 RUN apt install -y git libsndfile1-dev tesseract-ocr espeak-ng python3 python3-pip ffmpeg
@@ -28,21 +30,22 @@ RUN python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio tor
 
 RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate
 
+# needed in bnb and awq
+RUN python3 -m pip install --no-cache-dir einops
+
+# Add bitsandbytes for mixed int8 testing
+RUN python3 -m pip install --no-cache-dir bitsandbytes
+
+# Add gptqmodel for gtpq quantization testing, installed from source for pytorch==2.6.0 compatibility
+RUN python3 -m pip install lm_eval
+RUN git clone https://github.com/ModelCloud/GPTQModel.git && cd GPTQModel && pip install -v . --no-build-isolation
+
 # Add optimum for gptq quantization testing
 RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/optimum@main#egg=optimum
 
 # Add PEFT
 RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/peft@main#egg=peft
 
-# needed in bnb and awq
-RUN python3 -m pip install --no-cache-dir einops
-
-# Add bitsandbytes
-RUN python3 -m pip install --no-cache-dir bitsandbytes
-
-# # Add gptqmodel
-# RUN python3 -m pip install --no-cache-dir gptqmodel
-
 # Add hqq for quantization testing
 RUN python3 -m pip install --no-cache-dir hqq
 
@@ -80,9 +83,6 @@ RUN python3 -m pip uninstall -y flash-attn
 # this line must be added in order for python to be aware of transformers.
 RUN cd transformers && python3 setup.py develop
 
-# Add fp-quant for quantization testing
-RUN python3 -m pip install --no-cache-dir "fp-quant>=0.2.0"
-
 # Low usage or incompatible lib, will enable later on
 
 # # Add aqlm for quantization testing
@@ -103,3 +103,7 @@ RUN python3 -m pip install --no-cache-dir "fp-quant>=0.2.0"
 # # TODO: create a new workflow to test them
 # RUN python3 -m pip install --no-cache-dir flute-kernel==0.4.1
 # RUN python3 -m pip install --no-cache-dir git+https://github.com/Dao-AILab/fast-hadamard-transform.git
+
+# Add fp-quant for quantization testing
+# Requires py3.11 but our CI runs on 3.9
+# RUN python3 -m pip install --no-cache-dir "fp-quant>=0.1.6"

docs/README.md

@@ -24,7 +24,7 @@ pip install -e ".[dev]"
 ```
 
 > [!NOTE]
-> This command might fail for some OS that are missing dependencies. Check step 4 in [Create a Pull Request](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#create-a-pull-request) to work around it.
+> This command might fail for some OS that are missing dependencies. Check step 4 in [Create a Pull Request](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#create-a-pull-request) to workaround it.
 
 Then you need to install our special tool that builds the documentation:
 
@@ -38,7 +38,7 @@ pip install git+https://github.com/huggingface/doc-builder
 
 ## Building the documentation
 
-Once you have set up the `doc-builder` and additional packages, you can generate the documentation by 
+Once you have setup the `doc-builder` and additional packages, you can generate the documentation by 
 typing the following command:
 
 ```bash
@@ -295,11 +295,12 @@ Here's an example of a tuple return, comprising several objects:
 Due to the rapidly growing repository, it is important to make sure that no files that would significantly weigh down the repository are added. This includes images, videos, and other non-text files. We prefer to leverage a hf.co hosted `dataset` like
 the ones hosted on [`hf-internal-testing`](https://huggingface.co/hf-internal-testing) in which to place these files and reference
 them by URL. We recommend putting them in the following dataset: [huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images).
-If an external contribution, feel free to add the images to your PR and ask a Hugging Face member to migrate them to this dataset.
+If an external contribution, feel free to add the images to your PR and ask a Hugging Face member to migrate your images
+to this dataset.
 
 ## Styling the docstring
 
-We have an automatic script running with the `make style` command that will make sure that:
+We have an automatic script running with the `make style` comment that will make sure that:
 - the docstrings fully take advantage of the line width
 - all code examples are formatted using black, like the code of the Transformers library
 

docs/source/ar/_toctree.yml

@@ -123,6 +123,8 @@
     title: تشغيل التدريب على Amazon SageMaker
   - local: serialization
     title: التصدير إلى ONNX
+  - local: torchscript
+    title: التصدير إلى TorchScript
   - local: notebooks
     title: دفاتر الملاحظات مع الأمثلة
   - local: community
@@ -258,6 +260,8 @@
 #       title: النماذج
 #     - local: main_classes/text_generation
 #       title: توليد النصوص
+#     - local: main_classes/onnx
+#       title: ONNX
 #     - local: main_classes/optimizer_schedules
 #       title: التحسين
 #     - local: main_classes/output

docs/source/ar/autoclass_tutorial.md

@@ -52,7 +52,7 @@
     <figcaption class="mt-2 text-center text-sm text-gray-500">الصورة توضح مخطط مراحل نموذج Swin.</figcaption>
 </div>
 
-يسمح لك [`AutoBackbone`] باستخدام النماذج المُدربة مسبقًا كعمود فقري للحصول على خرائط ميزات من مراحل مختلفة من العمود الفقري. يجب عليك تحديد أحد المعلمات التالية في [`~PreTrainedConfig.from_pretrained`]:
+يسمح لك [`AutoBackbone`] باستخدام النماذج المُدربة مسبقًا كعمود فقري للحصول على خرائط ميزات من مراحل مختلفة من العمود الفقري. يجب عليك تحديد أحد المعلمات التالية في [`~PretrainedConfig.from_pretrained`]:
 
 * `out_indices` هو فهرس الطبقة التي تريد الحصول على خريطة الميزات منها
 * `out_features` هو اسم الطبقة التي تريد الحصول على خريطة الميزات منها

docs/source/ar/create_a_model.md

@@ -54,19 +54,19 @@ DistilBertConfig {
  
 ```
 
-يمكن تعديل خصائص النموذج المدرب مسبقًا في دالة [`~PreTrainedConfig.from_pretrained`] :
+يمكن تعديل خصائص النموذج المدرب مسبقًا في دالة [`~PretrainedConfig.from_pretrained`] :
 
 ```py
 >>> my_config = DistilBertConfig.from_pretrained("distilbert/distilbert-base-uncased", activation="relu", attention_dropout=0.4)
 ```
 
-بمجرد أن تصبح راضيًا عن تكوين نموذجك، يمكنك حفظه باستخدام [`~PreTrainedConfig.save_pretrained`]. يتم تخزين ملف التكوين الخاص بك على أنه ملف JSON في دليل الحفظ المحدد:
+بمجرد أن تصبح راضيًا عن تكوين نموذجك، يمكنك حفظه باستخدام [`~PretrainedConfig.save_pretrained`]. يتم تخزين ملف التكوين الخاص بك على أنه ملف JSON في دليل الحفظ المحدد:
 
 ```py
 >>> my_config.save_pretrained(save_directory="./your_model_save_path")
 ```
 
-لإعادة استخدام ملف التكوين، قم بتحميله باستخدام [`~PreTrainedConfig.from_pretrained`]:
+لإعادة استخدام ملف التكوين، قم بتحميله باستخدام [`~PretrainedConfig.from_pretrained`]:
 
 ```py
 >>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/config.json")

docs/source/ar/custom_models.md

@@ -20,11 +20,11 @@
 في مثالنا، سنعدّل بعض الوسائط في فئة ResNet التي قد نرغب في ضبطها. ستعطينا التكوينات المختلفة أنواع ResNets المختلفة الممكنة. سنقوم بتخزين هذه الوسائط بعد التحقق من صحته.
 
 ```python
-from transformers import PreTrainedConfig
+from transformers import PretrainedConfig
 from typing import List
 
 
-class ResnetConfig(PreTrainedConfig):
+class ResnetConfig(PretrainedConfig):
     model_type = "resnet"
 
     def __init__(
@@ -58,11 +58,11 @@ class ResnetConfig(PreTrainedConfig):
 ```
 الأشياء الثلاثة المهمة التي يجب تذكرها عند كتابة تكوينك الخاص هي:
 
-- يجب أن ترث من `PreTrainedConfig`،
-- يجب أن تقبل دالة  `__init__` الخاصة بـ `PreTrainedConfig` أي معامﻻت إضافية kwargs،
+- يجب أن ترث من `PretrainedConfig`،
+- يجب أن تقبل دالة  `__init__` الخاصة بـ `PretrainedConfig` أي معامﻻت إضافية kwargs،
 - يجب تمرير هذه المعامﻻت الإضافية إلى دالة `__init__` فى الفئة الأساسية الاعلى.
 
-يضمن الإرث حصولك على جميع الوظائف من مكتبة 🤗 Transformers، في حين أن القيدين التانى والثالث يأتيان من حقيقة أن `PreTrainedConfig` لديه المزيد من الحقول أكثر من تلك التي تقوم بتعيينها. عند إعادة تحميل تكوين باستخدام طريقة `from_pretrained`، يجب أن يقبل تكوينك هذه الحقول ثم إرسالها إلى الفئة الأساسية الأعلى.
+يضمن الإرث حصولك على جميع الوظائف من مكتبة 🤗 Transformers، في حين أن القيدين التانى والثالث يأتيان من حقيقة أن `PretrainedConfig` لديه المزيد من الحقول أكثر من تلك التي تقوم بتعيينها. عند إعادة تحميل تكوين باستخدام طريقة `from_pretrained`، يجب أن يقبل تكوينك هذه الحقول ثم إرسالها إلى الفئة الأساسية الأعلى.
 
 تحديد `model_type` لتكوينك (هنا `model_type="resnet"`) ليس إلزاميًا، ما لم ترغب في
 تسجيل نموذجك باستخدام الفئات التلقائية (راجع القسم الأخير).
@@ -82,7 +82,7 @@ resnet50d_config.save_pretrained("custom-resnet")
 resnet50d_config = ResnetConfig.from_pretrained("custom-resnet")
 ```
 
-يمكنك أيضًا استخدام أي طريقة أخرى من فئة [`PreTrainedConfig`]، مثل [`~PreTrainedConfig.push_to_hub`] لتحميل تكوينك مباشرة إلى Hub.
+يمكنك أيضًا استخدام أي طريقة أخرى من فئة [`PretrainedConfig`]، مثل [`~PretrainedConfig.push_to_hub`] لتحميل تكوينك مباشرة إلى Hub.
 
 ## كتابة نموذج مخصص
 

docs/source/ar/llm_tutorial.md

@@ -60,10 +60,10 @@ pip install transformers bitsandbytes>=0.39.0 -q
 أولاً، تحتاج إلى تحميل النموذج.
 
 ```py
->>> from transformers import AutoModelForCausalLM, BitsAndBytesConfig
+>>> from transformers import AutoModelForCausalLM
 
 >>> model = AutoModelForCausalLM.from_pretrained(
-...     "mistralai/Mistral-7B-v0.1", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
+...     "mistralai/Mistral-7B-v0.1", device_map="auto", load_in_4bit=True
 ... )
 ```
 
@@ -113,12 +113,12 @@ pip install transformers bitsandbytes>=0.39.0 -q
 هناك العديد من [استراتيجيات التوليد](generation_strategies)، وفي بعض الأحيان قد لا تكون القيم الافتراضية مناسبة لحالتك الاستخدام. إذا لم تكن الإخراج الخاصة بك متوافقة مع ما تتوقعه، فقد قمنا بإنشاء قائمة بأكثر الأخطاء الشائعة وكيفية تجنبها.
 
 ```py
->>> from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
 
 >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
 >>> tokenizer.pad_token = tokenizer.eos_token  # Most LLMs don't have a pad token by default
 >>> model = AutoModelForCausalLM.from_pretrained(
-...     "mistralai/Mistral-7B-v0.1", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
+...     "mistralai/Mistral-7B-v0.1", device_map="auto", load_in_4bit=True
 ... )
 ```
 
@@ -192,7 +192,7 @@ LLMs هي [معماريات فك التشفير فقط](https://huggingface.co/l
 ```python
 >>> tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-alpha")
 >>> model = AutoModelForCausalLM.from_pretrained(
-...     "HuggingFaceH4/zephyr-7b-alpha", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
+...     "HuggingFaceH4/zephyr-7b-alpha", device_map="auto", load_in_4bit=True
 ... )
 >>> set_seed(0)
 >>> prompt = """How many helicopters can a human eat in one sitting? Reply as a thug."""

docs/source/ar/llm_tutorial_optimization.md

@@ -231,7 +231,7 @@ flush()
 دعنا نرى ما هو استهلاك ذاكرة GPU الذروة الذي يوفره تكميم 4 بت. يمكن تكميم النموذج إلى 4 بت باستخدام نفس واجهة برمجة التطبيقات كما في السابق - هذه المرة عن طريق تمرير `load_in_4bit=True` بدلاً من `load_in_8bit=True`.
 
 ```python
-model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", quantization_config=BitsAndBytesConfig(load_in_4bit=True), pad_token_id=0)
+model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_4bit=True, pad_token_id=0)
 
 pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
 
@@ -329,6 +329,174 @@ $$ \textbf{O}_i \leftarrow s^a_{ij} * \textbf{O}_i + s^b_{ij} * \mathbf{V}_{j} \
 لنلقِ نظرة على مثال عملي.
 
 
+يحصل نموذج OctoCoder الخاص بنا الآن على موجه إدخال أطول بشكل كبير يتضمن ما يسمى *موجه النظام*. تُستخدم موجهات النظام لتوجيه LLM إلى مساعد أفضل مصمم لمهام المستخدمين.
+فيما يلي، نستخدم موجه النظام الذي سيجعل OctoCoder مساعد ترميز أفضل.
+
+```python
+system_prompt = """Below are a series of dialogues between various people and an AI technical assistant.
+The assistant tries to be helpful, polite, honest, sophisticated, emotionally aware, and humble but knowledgeable.
+The assistant is happy to help with code questions and will do their best to understand exactly what is needed.
+It also tries to avoid giving false or misleading information, and it caveats when it isn't entirely sure about the right answer.
+That said, the assistant is practical really does its best, and doesn't let caution get too much in the way of being useful.
+
+The Starcoder models are a series of 15.5B parameter models trained on 80+ programming languages from The Stack (v1.2) (excluding opt-out requests).
+The model uses Multi Query Attention, was trained using the Fill-in-the-Middle objective, and with 8,192 tokens context window for a trillion tokens of heavily deduplicated data.
+-----
+
+Question: Write a function that takes two lists and returns a list that has alternating elements from each input list.
+
+Answer: Sure. Here is a function that does that.
+
+def alternating(list1, list2):
+   results = []
+   for i in range(len(list1)):
+       results.append(list1[i])
+       results.append(list2[i])
+   return results
+
+Question: Can you write some test cases for this function?
+
+Answer: Sure, here are some tests.
+
+assert alternating([10, 20, 30], [1, 2, 3]) == [10, 1, 20, 2, 30, 3]
+assert alternating([True, False], [4, 5]) == [True, 4, False, 5]
+assert alternating([], []) == []
+
+Question: Modify the function so that it returns all input elements when the lists have uneven length. The elements from the longer list should be at the end.
+
+Answer: Here is the modified function.
+
+def alternating(list1, list2):
+   results = []
+   for i in range(min(len(list1), len(list2))):
+       results.append(list1[i])
+       results.append(list2[i])
+   if len(list1) > len(list2):
+       results.extend(list1[i+1:])
+   else:
+       results.extend(list2[i+1:])
+   return results
+-----
+"""
+```
+لأغراض التوضيح، سنكرر موجه النظام عشر مرات بحيث يكون طول الإدخال طويلاً بما يكفي لملاحظة وفورات ذاكرة Flash Attention.
+نضيف موجه النص الأصلي "سؤال: يرجى كتابة وظيفة في Python تقوم بتحويل البايتات إلى جيجا بايت.
+
+```python
+long_prompt = 10 * system_prompt + prompt
+```
+
+نقوم بتنفيذ نموذجنا مرة أخرى بدقة bfloat16.
+
+```python
+model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", dtype=torch.bfloat16, device_map="auto")
+tokenizer = AutoTokenizer.from_pretrained("bigcode/octocoder")
+
+pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
+```
+
+دعنا الآن نقوم بتشغيل النموذج تمامًا مثلما كان من قبل *بدون اهتمام فلاشي* وقياس متطلبات ذاكرة GPU وقت الذروة ووقت الاستدلال.
+
+```python
+import time
+
+start_time = time.time()
+result = pipe(long_prompt, max_new_tokens=60)[0]["generated_text"][len(long_prompt):]
+
+print(f"Generated in {time.time() - start_time} seconds.")
+result
+```
+
+**الإخراج**:
+```
+تم التوليد في 10.96854019165039 ثانية.
+بالتأكيد. إليك وظيفة للقيام بذلك.
+
+def bytes_to_giga(bytes):
+return bytes / 1024 / 1024 / 1024
+
+الإجابة: بالتأكيد. إليك وظيفة للقيام بذلك.
+
+ديف
+```
+
+نحصل على نفس الإخراج كما كان من قبل، ولكن هذه المرة، يقوم النموذج بتكرار الإجابة عدة مرات حتى يتم قطعها عند 60 رمزًا. ليس من المستغرب أننا كررنا موجه النظام عشر مرات لأغراض التوضيح وبالتالي قمنا بتشغيل النموذج لتكرار نفسه.
+
+**ملاحظة** لا ينبغي تكرار موجه النظام عشر مرات في التطبيقات الواقعية - مرة واحدة كافية!
+
+دعنا نقيس متطلبات ذاكرة GPU وقت الذروة.
+
+```python
+bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
+```
+
+**الإخراج**:
+```
+37.668193340301514
+```
+
+كما نرى، فإن متطلبات ذاكرة GPU وقت الذروة أعلى بكثير مما كانت عليه في البداية، وهو ما يرجع إلى حد كبير إلى تسلسل الإدخال الأطول. أيضًا، يستغرق التوليد أكثر من دقيقة بقليل الآن.
+
+نستدعي `flush()` لتحرير ذاكرة GPU لتجربتنا التالية.
+
+```python
+flush()
+```
+
+لمقارنة، دعونا نقوم بتشغيل نفس الدالة، ولكن تمكين الاهتمام فلاش بدلا من ذلك.
+للقيام بذلك، نقوم بتحويل النموذج إلى [BetterTransformer](Https://huggingface.co/docs/optimum/bettertransformer/overview) ومن خلال القيام بذلك تمكين PyTorch's [SDPA self-attention](Https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention) والتي بدورها قادرة على استخدام الاهتمام فلاش.
+
+```python
+model.to_bettertransformer()
+```
+
+الآن نقوم بتشغيل نفس مقتطف التعليمات البرمجية بالضبط كما كان من قبل وتحت الغطاء سوف تستخدم المحولات الاهتمام فلاش.
+
+```py
+start_time = time.time()
+with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
+    result = pipe(long_prompt, max_new_tokens=60)[0]["generated_text"][len(long_prompt):]
+
+print(f"Generated in {time.time() - start_time} seconds.")
+result
+```
+
+**الإخراج**:
+```
+تم التوليد في 3.0211617946624756 ثانية.
+بالتأكيد. إليك وظيفة للقيام بذلك.
+
+def bytes_to_giga(bytes):
+return bytes / 1024 / 1024 / 1024
+
+الإجابة: بالتأكيد. إليك وظيفة للقيام بذلك.
+
+ديف
+```
+
+نحصل على نفس النتيجة بالضبط كما كان من قبل، ولكن يمكننا ملاحظة تسريع كبير بفضل الاهتمام فلاش.
+
+دعنا نقيس استهلاك الذاكرة لآخر مرة.
+
+```python
+bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
+```
+
+**الإخراج**:
+```
+32.617331981658936
+```
+
+ونحن تقريبا مرة أخرى إلى ذاكرة GPU الذروة الأصلية لدينا 29GB.
+
+يمكننا أن نلاحظ أننا نستخدم فقط حوالي 100 ميجابايت إضافية من ذاكرة GPU عند تمرير تسلسل إدخال طويل جدًا مع الاهتمام فلاش مقارنة بتمرير تسلسل إدخال قصير كما فعلنا في البداية.
+
+```py
+flush()
+```
+
+لمزيد من المعلومات حول كيفية استخدام Flash Attention، يرجى الاطلاع على [صفحة doc هذه](Https://huggingface.co/docs/transformers/en/perf_infer_gpu_one#flashattention-2).
+
 ## 3. الابتكارات المعمارية
 
 حتى الآن، نظرنا في تحسين الكفاءة الحسابية والذاكرة من خلال:
@@ -472,7 +640,7 @@ for _ in range(5):
   next_token_id = torch.argmax(next_logits, dim=-1)
 
   print("shape of input_ids", next_token_id.shape)
-  print("length of key-value cache", past_key_values.get_seq_length())  # past_key_values are of shape [num_layers, 0 for k, 1 for v, batch_size, length, hidden_dim]
+  print("length of key-value cache", len(past_key_values[0][0]))  # past_key_values are of shape [num_layers, 0 for k, 1 for v, batch_size, length, hidden_dim]
   generated_tokens.append(next_token_id.item())
 
 generated_text = tokenizer.batch_decode(generated_tokens)

docs/source/ar/run_scripts.md

@@ -93,6 +93,7 @@ python examples/pytorch/summarization/run_summarization.py \
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --predict_with_generate
 ```
 
@@ -116,6 +117,7 @@ torchrun \
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --predict_with_generate
 ```
 
@@ -138,6 +140,7 @@ python xla_spawn.py --num_cores 8 \
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --predict_with_generate
 ```
 
@@ -194,6 +197,7 @@ python examples/pytorch/summarization/run_summarization.py \
     --summary_column summary_column_name \
     --source_prefix "summarize: " \
     --output_dir /tmp/tst-summarization \
+    --overwrite_output_dir \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
     --predict_with_generate
@@ -221,6 +225,7 @@ python examples/pytorch/summarization/run_summarization.py \
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --predict_with_generate
 ```
 
@@ -234,6 +239,8 @@ examples/pytorch/summarization/run_summarization.py -h
 
 خيار آخر مفيد لتمكينه هو استئناف التدريب من نقطة تفتيش سابقة. سيضمن ذلك أنك تستطيع الاستمرار من حيث توقفت دون البدء من جديد إذا تم مقاطعة تدريبك. هناك طريقتان لاستئناف التدريب من نقطة تفتيش.
 
+تستخدم الطريقة الأولى المعلمة `output_dir previous_output_dir` لاستئناف التدريب من أحدث نقطة تفتيش مخزنة في `output_dir`. في هذه الحالة، يجب عليك إزالة `overwrite_output_dir`:
+
 ```bash
 python examples/pytorch/summarization/run_summarization.py
     --model_name_or_path google-t5/t5-small \
@@ -245,6 +252,24 @@ python examples/pytorch/summarization/run_summarization.py
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --output_dir previous_output_dir \
+    --predict_with_generate
+```
+
+تستخدم الطريقة الثانية معلمة `resume_from_checkpoint path_to_specific_checkpoint` لاستئناف التدريب من مجلد نقطة تفتيش محددة.
+
+```bash
+python examples/pytorch/summarization/run_summarization.py
+    --model_name_or_path google-t5/t5-small \
+    --do_train \
+    --do_eval \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --source_prefix "summarize: " \
+    --output_dir /tmp/tst-summarization \
+    --per_device_train_batch_size=4 \
+    --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --resume_from_checkpoint path_to_specific_checkpoint \
     --predict_with_generate
 ```
@@ -276,5 +301,6 @@ python examples/pytorch/summarization/run_summarization.py
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --predict_with_generate
 ```

docs/source/ar/serialization.md

@@ -32,7 +32,7 @@
 لتصدير نموذج 🤗 Transformers إلى ONNX، قم أولاً بتثبيت اعتماد إضافي:
 
 ```bash
-pip install optimum-onnx
+pip install optimum[exporters]
 ```
 
 للاطلاع على جميع المعامﻻت المتاحة، يرجى الرجوع إلى [وثائق 🤗 Optimum](https://huggingface.co/docs/optimum/exporters/onnx/usage_guides/export_a_model#exporting-a-model-to-onnx-using-the-cli)، أو عرض المساعدة في سطر الأوامر:
@@ -111,3 +111,60 @@ optimum-cli export onnx --model keras-io/transformers-qa distilbert_base_cased_s
 ### تصدير نموذج لهندسة غير مدعومة
 
 إذا كنت ترغب في المساهمة من خلال إضافة دعم لنموذج لا يُمكن تصديره حاليًا، فيجب عليك أولاً التحقق مما إذا كان مدعومًا في [`optimum.exporters.onnx`](https://huggingface.co/docs/optimum/exporters/onnx/overview)، وإذا لم يكن مدعومًا، [فيمكنك المساهمة في 🤗 Optimum](https://huggingface.co/docs/optimum/exporters/onnx/usage_guides/contribute) مُباشرةً.
+
+### تصدير نموذج باستخدام `transformers.onnx`
+
+<Tip warning={true}>
+
+لم يعد يتم دعم `transformers.onnx`  يُرجى تصدير النماذج باستخدام 🤗 Optimum كما هو موضح أعلاه. سيتم إزالة هذا القسم في الإصدارات القادمة.
+
+</Tip>
+
+لتصدير نموذج 🤗 Transformers إلى ONNX باستخدام `transformers.onnx`، ثبّت التبعيات الإضافية:
+
+```bash
+pip install transformers[onnx]
+```
+
+استخدم حزمة `transformers.onnx` كنموذج Python لتصدير نقطة حفظ باستخدام تكوين جاهز:
+
+```bash
+python -m transformers.onnx --model=distilbert/distilbert-base-uncased onnx/
+```
+
+يُصدّر هذا رسمًا بيانيًا ONNX لنقطة الحفظ المُحددة بواسطة وسيطة `--model`. مرر أي نقطة حفظ على 🤗 Hub أو نقطة حفظ مُخزنة محليًا.
+يُمكن بعد ذلك تشغيل ملف `model.onnx` الناتج على أحد المُسرعات العديدة التي تدعم معيار ONNX. على سبيل المثال، قم بتحميل وتشغيل النموذج باستخدام ONNX Runtime كما يلي:
+
+```python
+>>> from transformers import AutoTokenizer
+>>> from onnxruntime import InferenceSession
+
+>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilbert-base-uncased")
+>>> session = InferenceSession("onnx/model.onnx")
+>>> # يتوقع ONNX Runtime مصفوفات NumPy كمدخلات
+>>> inputs = tokenizer("Using DistilBERT with ONNX Runtime!", return_tensors="np")
+>>> outputs = session.run(output_names=["last_hidden_state"], input_feed=dict(inputs))
+```
+
+يُمكن الحصول على أسماء المخرجات المطلوبة (مثل `["last_hidden_state"]`) من خلال إلقاء نظرة على تكوين ONNX لكل نموذج. على سبيل المثال، بالنسبة لـ DistilBERT، لدينا:
+
+```python
+>>> from transformers.models.distilbert import DistilBertConfig, DistilBertOnnxConfig
+
+>>> config = DistilBertConfig()
+>>> onnx_config = DistilBertOnnxConfig(config)
+>>> print(list(onnx_config.outputs.keys()))
+["last_hidden_state"]
+```
+
+العمليات مُتطابقة لنقاط الحفظ TensorFlow على Hub. على سبيل المثال، صدّر نقطة حفظ TensorFlow خالصة كما يلي:
+
+```bash
+python -m transformers.onnx --model=keras-io/transformers-qa onnx/
+```
+
+لتصدير نموذج مُخزن محليًا، احفظ أوزان النموذج ومجزىء اللغوى في نفس الدليل (على سبيل المثال `local-pt-checkpoint`)، ثم قم بتصديره إلى ONNX عن طريق توجيه وسيط `--model` لحزمة `transformers.onnx` إلى الدليل المطلوب:
+
+```bash
+python -m transformers.onnx --model=local-pt-checkpoint onnx/
+```

docs/source/ar/torchscript.md

@@ -0,0 +1,154 @@
+# التصدير إلى TorchScript
+
+<Tip>
+
+هذه هي بداية تجاربنا مع TorchScript ولا زلنا نستكشف قدراته مع نماذج المدخلات المتغيرة الحجم. إنه مجال اهتمامنا وسنعمق تحليلنا في الإصدارات القادمة، مع المزيد من الأمثلة البرمجية، وتنفيذ أكثر مرونة، ومقاييس مقارنة بين  الأكواد القائمة على Python مع أكواد TorchScript المُجمّعة.
+
+</Tip>
+
+وفقًا لـ [وثائق TorchScript](https://pytorch.org/docs/stable/jit.html):
+
+> TorchScript هي طريقة لإنشاء نماذج قابلة للتسلسل والتحسين من تعليمات PyTorch البرمجية.
+
+هناك وحدتان من PyTorch، [JIT and TRACE](https://pytorch.org/docs/stable/jit.html)، تتيحان للمطورين تصدير نماذجهم لإعادة استخدامها في برامج أخرى مثل برامج C++ المُحسّنة للأداء.
+
+نقدم واجهة تتيح لك تصدير نماذج 🤗 Transformers إلى TorchScript بحيث يمكن إعادة استخدامها في بيئة مختلفة عن برامج Python القائمة إلى PyTorch. هنا نشرح كيفية تصدير نماذجنا واستخدامها باستخدام TorchScript.
+
+يتطلب تصدير نموذج أمرين:
+
+- تهيئة مثيل للنموذج باستخدام علامة `torchscript`
+- تمرير مُدخلات وهمية (dummy inputs) خلال النموذج
+
+تنطوي هذه الضرورات على عدة أمور يجب على المطورين توخي الحذر بشأنها كما هو مفصل أدناه.
+
+## علامة TorchScript والأوزان المرتبطة
+
+علامة `torchscript` ضرورية لأن معظم نماذج اللغة 🤗 Transformers لها أوزان مرتبطة بين طبقة `Embedding` وطبقة `Decoding`. لا يسمح لك TorchScript بتصدير النماذج ذات الأوزان المرتبطة، لذلك من الضروري فصل الأوزان ونسخها مسبقًا.
+
+النماذج المُهيأة باستخدام علامة `torchscript` لها طبقة `Embedding` وطبقة`Decoding` منفصلتين، مما يعني أنه لا ينبغي تدريبها لاحقًا. سيؤدي التدريب إلى عدم تزامن الطبقتين، مما يؤدي إلى نتائج غير متوقعة.
+
+هذا لا ينطبق على النماذج التي لا تحتوي على رأس نموذج اللغة، حيث لا تملك أوزانًا مرتبطة. يمكن تصدير هذه النماذج بأمان دون علامة `torchscript`.
+
+## المدخلات الوهمية والأطوال القياسية
+
+تُستخدم المُدخلات الوهمية لتمرير أمامي خلال النموذج. أثناء انتشار قيم المُدخلات عبر الطبقات، يتتبع PyTorch العمليات المختلفة التي يتم تنفيذها على كل مصفوفة(tensor). ثم يتم استخدام هذه العمليات المُسجلة بعد ذلك لإنشاء *أثر* النموذج.
+
+يتم إنشاء التتبع بالنسبة لأبعاد المُدخلات. وبالتالي، فهو مُقيّد بأبعاد المُدخلات الوهمية، ولن يعمل لأي طول تسلسل أو حجم دفعة مختلف. عند المحاولة بحجم مختلف، يتم رفع الخطأ التالي:
+
+```
+`The expanded size of the tensor (3) must match the existing size (7) at non-singleton dimension 2`
+```
+
+نوصي بتتبع النموذج باستخدام حجم مُدخلات وهمية لا يقل عن أكبر مُدخل سيتم تقديمه للنموذج أثناء الاستدلال. يمكن أن تساعد الحشوة(padding) في ملء القيم المفقودة. ومع ذلك، نظرًا لتتبع النموذج بحجم مُدخل أكبر، ستكون أبعاد المصفوفة ستكون كبيرة أيضًا، مما يؤدي عنه المزيد من الحسابات.
+
+انتبه إلى إجمالي عدد العمليات المُنفذة على كل مُدخل وتابع الأداء عن كثب عند تصدير نماذج متغيرة طول التسلسل.
+
+## استخدام TorchScript في Python
+
+يوضح هذا القسم كيفية حفظ النماذج وتحميلها، بالإضافة إلى كيفية استخدام التتبع للاستدلال.
+
+### حفظ نموذج
+
+لتصدير `BertModel` باستخدام TorchScript، قم بتهيئة ـ `BertModel` من فئة `BertConfig` ثم احفظه على القرص تحت اسم الملف `traced_bert.pt`:
+
+```python
+from transformers import BertModel, BertTokenizer, BertConfig
+import torch
+
+enc = BertTokenizer.from_pretrained("google-bert/bert-base-uncased")
+
+# Tokenizing input text
+text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
+tokenized_text = enc.tokenize(text)
+
+# Masking one of the input tokens
+masked_index = 8
+tokenized_text[masked_index] = "[MASK]"
+indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
+segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
+
+# Creating a dummy input
+tokens_tensor = torch.tensor([indexed_tokens])
+segments_tensors = torch.tensor([segments_ids])
+dummy_input = [tokens_tensor, segments_tensors]
+
+# Initializing the model with the torchscript flag
+# Flag set to True even though it is not necessary as this model does not have an LM Head.
+config = BertConfig(
+    vocab_size_or_config_json_file=32000,
+    hidden_size=768,
+    num_hidden_layers=12,
+    num_attention_heads=12,
+    intermediate_size=3072,
+    torchscript=True,
+)
+
+# Instantiating the model
+model = BertModel(config)
+
+# The model needs to be in evaluation mode
+model.eval()
+
+# If you are instantiating the model with *from_pretrained* you can also easily set the TorchScript flag
+model = BertModel.from_pretrained("google-bert/bert-base-uncased", torchscript=True)
+
+# Creating the trace
+traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
+torch.jit.save(traced_model, "traced_bert.pt")
+```
+
+### تحميل نموذج
+
+يمكنك الآن تحميل `BertModel` المُحفظ سابقًا، `traced_bert.pt`، من القرص واستخدامه على `dummy_input` المُهيأ سابقًا:
+
+```python
+loaded_model = torch.jit.load("traced_bert.pt")
+loaded_model.eval()
+
+all_encoder_layers, pooled_output = loaded_model(*dummy_input)
+```
+
+### استخدام نموذج مُتتبع للاستدلال
+
+استخدم النموذج المُتتبع للاستدلال باستخدام أسلوب `__call__` الخاص به:
+
+```python
+traced_model(tokens_tensor, segments_tensors)
+```
+
+## نشر نماذج Hugging Face TorchScript على AWS باستخدام Neuron SDK
+
+قدمت AWS عائلة [Amazon EC2 Inf1](https://aws.amazon.com/ec2/instance-types/inf1/) من اﻷجهزة لخفض التكلفة وأداء التعلم الآلي عالي الأداء في البيئة السحابية. تعمل أجهزة Inf1 بواسطة شريحة Inferentia من AWS، وهي مُسرّع أجهزة مُخصص، متخصص في أعباء عمل الاستدلال للتعلم العميق. [AWS Neuron](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/#) هي SDK لـ Inferentia التي تدعم تتبع نماذج المحولات وتحسينها للنشر على Inf1. توفر Neuron SDK ما يلي:
+
+1. واجهة برمجة تطبيقات سهلة الاستخدام مع تغيير سطر واحد من التعليمات البرمجية لتتبع نموذج TorchScript وتحسينه للاستدلال في البيئة السحابية.
+2. تحسينات الأداء الجاهزة للاستخدام [تحسين التكلفة والأداء](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/benchmark/>).
+3. دعم نماذج Hugging Face المحولات المبنية باستخدام إما [PyTorch](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/bert_tutorial/tutorial_pretrained_bert.html) أو [TensorFlow](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/tensorflow/huggingface_bert/huggingface_bert.html).
+
+### الآثار المترتبة
+
+تعمل نماذج المحولات المستندة إلى بنية [BERT (تمثيلات الترميز ثنائية الاتجاه من المحولات)](https://huggingface.co/docs/transformers/main/model_doc/bert) أو متغيراتها مثل [distilBERT](https://huggingface.co/docs/transformers/main/model_doc/distilbert) و [roBERTa](https://huggingface.co/docs/transformers/main/model_doc/roberta) بشكل أفضل على Inf1 للمهام غير التوليدية مثل الإجابة على الأسئلة الاستخراجية، وتصنيف التسلسلات، وتصنيف الرموز (tokens). ومع ذلك، يمكن تكييف مهام توليد النصوص للعمل على Inf1 وفقًا لهذا [برنامج تعليمي AWS Neuron MarianMT](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/transformers-marianmt.html). يمكن العثور على مزيد من المعلومات حول النماذج التي يمكن تحويلها جاهزة على Inferentia في قسم [ملاءمة بنية النموذج](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/models/models-inferentia.html#models-inferentia) من وثائق Neuron.
+
+### التبعيات (Dependencies)
+
+يتطلب استخدام AWS Neuron لتحويل النماذج [بيئة SDK Neuron](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/neuron-frameworks/pytorch-neuron/index.html#installation-guide) والتي تأتي مسبقًا على [AMI للتعلم العميق من AWS](https://docs.aws.amazon.com/dlami/latest/devguide/tutorial-inferentia-launching.html).
+
+### تحويل نموذج لـ AWS Neuron
+
+قم بتحويل نموذج لـ AWS NEURON باستخدام نفس التعليمات البرمجية من [استخدام TorchScript في Python](torchscript#using-torchscript-in-python) لتتبع `BertModel`. قم باستيراد امتداد إطار عمل `torch.neuron` للوصول إلى مكونات Neuron SDK من خلال واجهة برمجة تطبيقات Python:
+
+```python
+from transformers import BertModel, BertTokenizer, BertConfig
+import torch
+import torch.neuron
+```
+
+كل ما عليك فعله هو تعديل السطر التالي:
+
+```diff
+- torch.jit.trace(model, [tokens_tensor, segments_tensors])
++ torch.neuron.trace(model, [token_tensor, segments_tensors])
+```
+
+يتيح ذلك لـ Neuron SDK تتبع النموذج وتحسينه لمثيلات Inf1.
+
+لمعرفة المزيد حول ميزات AWS Neuron SDK والأدوات ودروس البرامج التعليمية والتحديثات الأخيرة، يرجى الاطلاع على [وثائق AWS NeuronSDK](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/index.html).

docs/source/ar/trainer.md

@@ -611,6 +611,7 @@ accelerate launch \
     --learning_rate 5e-5 \
     --num_train_epochs 3 \
     --output_dir /tmp/$TASK_NAME/ \
+    --overwrite_output_dir
 ```
 
 يمكنك أيضًا تحديد المعلمات من ملف `config_file.yaml` مباشرة في سطر الأوامر:
@@ -633,6 +634,7 @@ accelerate launch --num_processes=2 \
     --learning_rate 5e-5 \
     --num_train_epochs 3 \
     --output_dir /tmp/$TASK_NAME/ \
+    --overwrite_output_dir
 ```
 
 اطلع على برنامج تعليمي [Launching your Accelerate scripts](https://huggingface.co/docs/accelerate/basic_tutorials/launch) لمعرفة المزيد حول `accelerate_launch` والتكوينات المخصصة.

docs/source/de/add_new_model.md

@@ -53,7 +53,7 @@ Lassen Sie uns daher ein wenig tiefer in das allgemeine Design der Bibliothek ei
 ### Überblick über die Modelle
 
 Um ein Modell erfolgreich hinzuzufügen, ist es wichtig, die Interaktion zwischen Ihrem Modell und seiner Konfiguration zu verstehen,
-[`PreTrainedModel`] und [`PreTrainedConfig`]. Als Beispiel werden wir
+[`PreTrainedModel`] und [`PretrainedConfig`]. Als Beispiel werden wir
 das Modell, das zu 🤗 Transformers hinzugefügt werden soll, `BrandNewBert` nennen.
 
 Schauen wir uns das mal an:
@@ -81,10 +81,10 @@ model.config  # model has access to its config
 ```
 
 Ähnlich wie das Modell erbt die Konfiguration grundlegende Serialisierungs- und Deserialisierungsfunktionalitäten von
-[`PreTrainedConfig`]. Beachten Sie, dass die Konfiguration und das Modell immer in zwei verschiedene Formate serialisiert werden
+[`PretrainedConfig`]. Beachten Sie, dass die Konfiguration und das Modell immer in zwei verschiedene Formate serialisiert werden
 unterschiedliche Formate serialisiert werden - das Modell in eine *pytorch_model.bin* Datei und die Konfiguration in eine *config.json* Datei. Aufruf von
 [`~PreTrainedModel.save_pretrained`] wird automatisch
-[`~PreTrainedConfig.save_pretrained`] auf, so dass sowohl das Modell als auch die Konfiguration gespeichert werden.
+[`~PretrainedConfig.save_pretrained`] auf, so dass sowohl das Modell als auch die Konfiguration gespeichert werden.
 
 
 ### Code-Stil

docs/source/de/llm_tutorial.md

@@ -78,10 +78,10 @@ Wenn Sie an der grundlegenden Verwendung von LLMs interessiert sind, ist unsere
 Zunächst müssen Sie das Modell laden.
 
 ```py
->>> from transformers import AutoModelForCausalLM, BitsAndBytesConfig
+>>> from transformers import AutoModelForCausalLM
 
 >>> model = AutoModelForCausalLM.from_pretrained(
-...     "openlm-research/open_llama_7b", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
+...     "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
 ... )
 ```
 
@@ -119,12 +119,12 @@ Und das war's! Mit ein paar Zeilen Code können Sie sich die Macht eines LLM zun
 Es gibt viele [Generierungsstrategien](generation_strategies), und manchmal sind die Standardwerte für Ihren Anwendungsfall vielleicht nicht geeignet. Wenn Ihre Ausgaben nicht mit dem übereinstimmen, was Sie erwarten, haben wir eine Liste der häufigsten Fallstricke erstellt und wie Sie diese vermeiden können.
 
 ```py
->>> from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
 
 >>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b")
 >>> tokenizer.pad_token = tokenizer.eos_token  # Llama has no pad token by default
 >>> model = AutoModelForCausalLM.from_pretrained(
-...     "openlm-research/open_llama_7b", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
+...     "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
 ... )
 ```
 

docs/source/de/run_scripts.md

@@ -98,6 +98,7 @@ python examples/pytorch/summarization/run_summarization.py \
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --predict_with_generate
 ```
 
@@ -121,6 +122,7 @@ torchrun \
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --predict_with_generate
 ```
 
@@ -142,6 +144,7 @@ python xla_spawn.py --num_cores 8 \
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --predict_with_generate
 ```
 
@@ -198,6 +201,7 @@ python examples/pytorch/summarization/run_summarization.py \
     --summary_column summary_column_name \
     --source_prefix "summarize: " \
     --output_dir /tmp/tst-summarization \
+    --overwrite_output_dir \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
     --predict_with_generate
@@ -225,6 +229,7 @@ python examples/pytorch/summarization/run_summarization.py \
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --predict_with_generate
 ```
 
@@ -238,6 +243,8 @@ examples/pytorch/summarization/run_summarization.py -h
 
 Eine weitere hilfreiche Option, die Sie aktivieren können, ist die Wiederaufnahme des Trainings von einem früheren Kontrollpunkt aus. Auf diese Weise können Sie im Falle einer Unterbrechung Ihres Trainings dort weitermachen, wo Sie aufgehört haben, ohne von vorne beginnen zu müssen. Es gibt zwei Methoden, um das Training von einem Kontrollpunkt aus wieder aufzunehmen.
 
+Die erste Methode verwendet das Argument `output_dir previous_output_dir`, um das Training ab dem letzten in `output_dir` gespeicherten Kontrollpunkt wieder aufzunehmen. In diesem Fall sollten Sie `overwrite_output_dir` entfernen:
+
 ```bash
 python examples/pytorch/summarization/run_summarization.py
     --model_name_or_path google-t5/t5-small \
@@ -249,6 +256,24 @@ python examples/pytorch/summarization/run_summarization.py
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --output_dir previous_output_dir \
+    --predict_with_generate
+```
+
+Die zweite Methode verwendet das Argument `Resume_from_checkpoint path_to_specific_checkpoint`, um das Training ab einem bestimmten Checkpoint-Ordner wieder aufzunehmen.
+
+```bash
+python examples/pytorch/summarization/run_summarization.py
+    --model_name_or_path google-t5/t5-small \
+    --do_train \
+    --do_eval \
+    --dataset_name cnn_dailymail \
+    --dataset_config "3.0.0" \
+    --source_prefix "summarize: " \
+    --output_dir /tmp/tst-summarization \
+    --per_device_train_batch_size=4 \
+    --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --resume_from_checkpoint path_to_specific_checkpoint \
     --predict_with_generate
 ```
@@ -280,5 +305,6 @@ python examples/pytorch/summarization/run_summarization.py
     --output_dir /tmp/tst-summarization \
     --per_device_train_batch_size=4 \
     --per_device_eval_batch_size=4 \
+    --overwrite_output_dir \
     --predict_with_generate
 ```

docs/source/en/_toctree.yml

@@ -88,8 +88,6 @@
       title: Tool use
     - local: chat_templating_writing
       title: Writing a chat template
-    - local: chat_response_parsing
-      title: Response parsing
     title: Chat with models
   - sections:
     - local: serving
@@ -218,17 +216,14 @@
   - local: quantization/contribute
     title: Contribute
   title: Quantization
-- isExpanded: false
-  sections:
-  - local: kernel_doc/overview
-    title: Kernels in transformers
-  title: Kernels
 - isExpanded: false
   sections:
   - local: serialization
     title: ONNX
   - local: executorch
     title: ExecuTorch
+  - local: torchscript
+    title: TorchScript
   title: Export to production
 - isExpanded: false
   sections:
@@ -284,8 +279,6 @@
         title: Knowledge Distillation for Computer Vision
       - local: tasks/keypoint_matching
         title: Keypoint matching
-      - local: tasks/training_vision_backbone
-        title: Training vision models using Backbone API
       title: Computer vision
     - sections:
       - local: tasks/image_captioning
@@ -312,8 +305,6 @@
     title: Glossary
   - local: philosophy
     title: Philosophy
-  - local: models_timeline
-    title: Models Timeline
   - local: notebooks
     title: Notebooks with examples
   - local: community
@@ -349,6 +340,8 @@
       title: Models
     - local: main_classes/text_generation
       title: Text Generation
+    - local: main_classes/onnx
+      title: ONNX
     - local: main_classes/optimizer_schedules
       title: Optimization
     - local: main_classes/output
@@ -375,8 +368,6 @@
       title: Image Processor
     - local: main_classes/video_processor
       title: Video Processor
-    - local: main_classes/kernels
-      title: Kernels
     title: Main Classes
   - sections:
     - sections:
@@ -546,6 +537,8 @@
         title: Helium
       - local: model_doc/herbert
         title: HerBERT
+      - local: model_doc/hgnet_v2
+        title: HGNet-V2
       - local: model_doc/hunyuan_v1_dense
         title: HunYuanDenseV1
       - local: model_doc/hunyuan_v1_moe
@@ -562,8 +555,6 @@
         title: LED
       - local: model_doc/lfm2
         title: LFM2
-      - local: model_doc/lfm2_moe
-        title: LFM2Moe
       - local: model_doc/llama
         title: LLaMA
       - local: model_doc/llama2
@@ -1026,8 +1017,6 @@
         title: CLIPSeg
       - local: model_doc/clvp
         title: CLVP
-      - local: model_doc/cwm
-        title: Code World Model (CWM)
       - local: model_doc/cohere2_vision
         title: Cohere2Vision
       - local: model_doc/colpali
@@ -1044,10 +1033,6 @@
         title: DePlot
       - local: model_doc/donut
         title: Donut
-      - local: model_doc/edgetam
-        title: EdgeTAM
-      - local: model_doc/edgetam_video
-        title: EdgeTamVideo
       - local: model_doc/emu3
         title: Emu3
       - local: model_doc/evolla
@@ -1188,8 +1173,6 @@
         title: TVP
       - local: model_doc/udop
         title: UDOP
-      - local: model_doc/video_llama_3
-        title: VideoLlama3
       - local: model_doc/video_llava
         title: VideoLlava
       - local: model_doc/vilt
@@ -1255,8 +1238,6 @@
       title: Importing Utilities
     - local: internal/time_series_utils
       title: Utilities for Time Series
-    - local: internal/rope_utils
-      title: Rotary Embeddings Utilities
     title: Internal helpers
   - sections:
     - local: reference/environment_variables

docs/source/en/accelerator_selection.md

@@ -55,7 +55,6 @@ deepspeed --num_gpus 2 trainer-program.py ...
 </hfoptions>
 
 ## Order of accelerators
-
 To select specific accelerators to use and their order, use the environment variable appropriate for your hardware. This is often set on the command line for each run, but can also be added to your `~/.bashrc` or other startup config file.
 
 For example, if there are 4 accelerators (0, 1, 2, 3) and you only want to run accelerators 0 and 2:

docs/source/en/add_new_model.md

@@ -51,7 +51,7 @@ This section describes how the model and configuration classes interact and the
 
 ### Model and configuration
 
-All Transformers' models inherit from a base [`PreTrainedModel`] and [`PreTrainedConfig`] class. The configuration is the models blueprint.
+All Transformers' models inherit from a base [`PreTrainedModel`] and [`PretrainedConfig`] class. The configuration is the models blueprint.
 
 There is never more than two levels of abstraction for any model to keep the code readable. The example model here, BrandNewLlama, inherits from `BrandNewLlamaPreTrainedModel` and [`PreTrainedModel`]. It is important that a new model only depends on [`PreTrainedModel`] so that it can use the [`~PreTrainedModel.from_pretrained`] and [`~PreTrainedModel.save_pretrained`] methods.
 
@@ -66,9 +66,9 @@ model = BrandNewLlamaModel.from_pretrained("username/brand_new_llama")
 model.config
 ```
 
-[`PreTrainedConfig`] provides the [`~PreTrainedConfig.from_pretrained`] and [`~PreTrainedConfig.save_pretrained`] methods.
+[`PretrainedConfig`] provides the [`~PretrainedConfig.from_pretrained`] and [`~PretrainedConfig.save_pretrained`] methods.
 
-When you use [`PreTrainedModel.save_pretrained`], it automatically calls [`PreTrainedConfig.save_pretrained`] so that both the model and configuration are saved together.
+When you use [`PreTrainedModel.save_pretrained`], it automatically calls [`PretrainedConfig.save_pretrained`] so that both the model and configuration are saved together.
 
 A model is saved to a `model.safetensors` file and a configuration is saved to a `config.json` file.
 

docs/source/en/attention_interface.md

@@ -193,4 +193,4 @@ def custom_attention_mask(
 
 It mostly works thanks to the `mask_function`, which is a `Callable` in the form of [torch's mask_mod functions](https://pytorch.org/blog/flexattention/), taking 4 indices as input and returning a boolean to indicate if this position should take part in the attention computation.
 
-If you cannot use the `mask_function` to create your mask for some reason, you can try to work around it by doing something similar to our [torch export workaround](https://github.com/huggingface/transformers/blob/main/src/transformers/integrations/executorch.py).
+If you cannot use the `mask_function` to create your mask for some reason, you can try to work around it by doing something similar to our [torch export workaround](https://github.com/huggingface/transformers/blob/main/src/transformers/integrations/executorch.py).

docs/source/en/auto_docstring.md

@@ -210,9 +210,9 @@ There are some rules for documenting different types of arguments and they're li
         This can span multiple lines.
     ```
 
-  * Include `type` in backticks.
-  * Add *optional* if the argument is not required or has a default value.
-  * Add "defaults to X" if it has a default value. You don't need to add "defaults to `None`" if the default value is `None`.
+    * Include `type` in backticks.
+    * Add *optional* if the argument is not required or has a default value.
+    * Add "defaults to X" if it has a default value. You don't need to add "defaults to `None`" if the default value is `None`.
 
     These arguments can also be passed to `@auto_docstring` as a `custom_args` argument. It is used to define the docstring block for new arguments once if they are repeated in multiple places in the modeling file.
 
@@ -292,7 +292,7 @@ The `@auto_docstring` decorator automatically generates docstrings by:
 
 8. Unrolling kwargs typed with the unpack operator. For specific methods (defined in `UNROLL_KWARGS_METHODS`) or classes (defined in `UNROLL_KWARGS_CLASSES`), the decorator processes `**kwargs` parameters that are typed with `Unpack[KwargsTypedDict]`. It extracts the documentations from the `TypedDict` and adds each parameter to the function's docstring.
 
-    Currently only supported for [`ImagesKwargs`].
+    Currently only supported for [`FastImageProcessorKwargs`].
 
 ## Best practices
 

docs/source/en/backbones.md

@@ -22,7 +22,7 @@ Higher-level computer visions tasks, such as object detection or image segmentat
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/Backbone.png"/>
 </div>
 
-Load a backbone with [`~PreTrainedConfig.from_pretrained`] and use the `out_indices` parameter to determine which layer, given by the index, to extract a feature map from.
+Load a backbone with [`~PretrainedConfig.from_pretrained`] and use the `out_indices` parameter to determine which layer, given by the index, to extract a feature map from.
 
 ```py
 from transformers import AutoBackbone
@@ -46,7 +46,7 @@ There are two ways to load a Transformers backbone, [`AutoBackbone`] and a model
 <hfoptions id="backbone-classes">
 <hfoption id="AutoBackbone">
 
-The [AutoClass](./model_doc/auto) API automatically loads a pretrained vision model with [`~PreTrainedConfig.from_pretrained`] as a backbone if it's supported.
+The [AutoClass](./model_doc/auto) API automatically loads a pretrained vision model with [`~PretrainedConfig.from_pretrained`] as a backbone if it's supported.
 
 Set the `out_indices` parameter to the layer you'd like to get the feature map from. If you know the name of the layer, you could also use `out_features`. These parameters can be used interchangeably, but if you use both, make sure they refer to the same layer.
 

docs/source/en/cache_explanation.md

@@ -41,13 +41,13 @@ $$
 
 The query (`Q`), key (`K`), and value (`V`) matrices are projections from the input embeddings of shape `(b, h, T, d_head)`.
 
-For causal attention, the mask prevents the model from attending to future tokens. Once a token is processed, its representation never changes with respect to future tokens, which means $ K_{\text{past}} $ and $ V_{\text{past}} $ can be cached and reused to compute the last token's representation.
+For causal attention, the mask prevents the model from attending to future tokens. Once a token is processed, its representation never changes with respect to future tokens, which means \\( K_{\text{past}} \\) and \\( V_{\text{past}} \\) can be cached and reused to compute the last token's representation.
 
 $$
 \text{Attention}(q_t, [\underbrace{k_1, k_2, \dots, k_{t-1}}_{\text{cached}}, k_{t}], [\underbrace{v_1, v_2, \dots, v_{t-1}}_{\text{cached}}, v_{t}])
 $$
 
-At inference time, you only need the last token's query to compute the representation $ x_t $ that predicts the next token $ t+1 $. At each step, the new key and value vectors are **stored** in the cache and **appended** to the past keys and values.
+At inference time, you only need the last token's query to compute the representation \\( x_t \\) that predicts the next token \\( t+1 \\). At each step, the new key and value vectors are **stored** in the cache and **appended** to the past keys and values.
 
 $$
 K_{\text{cache}} \leftarrow \text{concat}(K_{\text{past}}, k_t), \quad V_{\text{cache}} \leftarrow \text{concat}(V_{\text{past}}, v_t)
@@ -59,7 +59,7 @@ Refer to the table below to compare how caching improves efficiency.
 
 | without caching | with caching |
 |---|---|
-| for each step, recompute all previous `K` and `V`  | for each step, only compute current `K` and `V` |
+| for each step, recompute all previous `K` and `V`  | for each step, only compute current `K` and `V`
 | attention cost per step is **quadratic** with sequence length | attention cost per step is **linear** with sequence length (memory grows linearly, but compute/token remains low) |
 
 ## Cache class
@@ -98,10 +98,9 @@ The example below demonstrates how to create a generation loop with [`DynamicCac
 
 ```py
 import torch
-from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache
-from accelerate import Accelerator
+from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache, infer_device
 
-device = Accelerator().device
+device = f"{infer_device()}:0"
 
 model_id = "meta-llama/Llama-2-7b-chat-hf"
 model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map=device)
@@ -144,10 +143,9 @@ The generation loop usually takes care of the cache position, but if you're writ
 
 ```py
 import torch
-from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache
-from accelerate import Accelerator
+from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache, infer_device
 
-device = Accelerator().device
+device = f"{infer_device()}:0"
 
 model_id = "meta-llama/Llama-2-7b-chat-hf"
 model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map=device)
@@ -158,3 +156,30 @@ inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, ret
 generated_ids = model.generate(**inputs, use_cache=True, max_new_tokens=10)
 
 ```
+
+## Legacy cache format
+
+Before the [`Cache`] class, the cache used to be stored as a tuple of tuples of tensors. This format is dynamic because it grows as text is generated, similar to [`DynamicCache`].
+
+The legacy format is essentially the same data structure but organized differently.
+- It's a tuple of tuples, where each inner tuple contains the key and value tensors for a layer.
+- The tensors have the same shape `[batch_size, num_heads, seq_len, head_dim]`.
+- The format is less flexible and doesn't support features like quantization or offloading.
+
+If your project depends on this legacy format, we recommend to convert to [`DynamicCache`] with [`~DynamicCache.from_legacy_cache`]. Note that legacy cache format is deprecated and not used anymore in `Transformers`. You can convert back to tuple format with [`DynamicCache.to_legacy_cache`] functions, which is helpful if you have custom logic for manipulating a cache in a specific format.
+
+```py
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
+model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-chat-hf", dtype=torch.float16, device_map="auto")
+inputs = tokenizer("Hello, my name is", return_tensors="pt").to(model.device)
+
+# `return_dict_in_generate=True` is required to return the cache and `return_legacy_cache` forces the returned cache
+# in the legacy format
+generation_outputs = model.generate(**inputs, return_dict_in_generate=True, return_legacy_cache=True, max_new_tokens=5)
+
+cache = DynamicCache.from_legacy_cache(generation_outputs.past_key_values)
+legacy_format_cache = cache.to_legacy_cache()
+```

docs/source/en/chat_extras.md

@@ -95,12 +95,9 @@ print(tokenizer.decode(outputs[0][len(inputs["input_ids"][0]):]))
 
 The chat model called the `get_current_temperature` tool with the correct parameters from the docstring. It inferred France as the location based on Paris, and that it should use Celsius for the units of temperature.
 
-A model **cannot actually call the tool itself**. It requests a tool call, and it's your job to handle the call and append it and the result to the chat history. For
-models that support [response parsing](./chat_response_parsing), the response parsing will be handled automatically, and you can just use
-[`~PreTrainedTokenizer.parse_response] to extract the tool call. For other models, you'll need to manually translate the output
-string into a tool call dict.
+A model **cannot actually call the tool itself**. It requests a tool call, and it's your job to handle the call and append it and the result to the chat history.
 
-Regardless of the approach you use, the tool call should go in the `tool_calls` key of an `assistant` message. This is the recommended API, and should be supported by the chat template of most tool-using models.
+Hold the call in the `tool_calls` key of an `assistant` message. This is the recommended API, and should be supported by the chat template of most tool-using models.
 
 > [!WARNING]
 > Although `tool_calls` is similar to the OpenAI API, the OpenAI API uses a JSON string as its `tool_calls` format. This may cause errors or strange model behavior if used in Transformers, which expects a dict.
@@ -224,4 +221,4 @@ model_input = tokenizer.apply_chat_template(
     messages,
     tools = [current_time, multiply]
 )
-```
+```

docs/source/en/chat_response_parsing.md

@@ -1,233 +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.
-
--->
-
-# Response Parsing
-
-It is increasingly common for chat models to generate structured outputs, rather than just a single reply string. 
-The most common uses for structured outputs are [tool calling](./chat_extras) and [reasoning models](https://huggingface.co/reasoning-course).
-Tool calling models can output tool calls, containing the name of the tool to call and any arguments to be passed to it,
-while reasoning models often output reasoning steps as a "chain of thought". Some recent models even use both of these,
-and may output reasoning and/or one or more tool calls before their final answer.
-
-Models with structured outputs pose a challenge for chat templating, because the output needs to be parsed before it
-can be appended to the chat. For a concrete example, let's say we ask [GPT-OSS](https://huggingface.co/openai/gpt-oss-120b)
-what the weather is like, and it thinks and decides to call a tool. Here's what the raw model output might look like:
-
-```txt
-<|start|>analysis<|message|>The user asks: "What is the weather like in SF?" We need to get the location of the user? The user explicitly asks about SF (San Francisco).
-So we need to get the current weather in San Francisco, CA. We need to call get_current_weather function. But we need to call function to get weather data.
-So we should call get_current_weather with location "San Francisco, CA". Let's do that.
-We will call function get_current_weather.<|end|><|start|>commentary to=functions.get_current_weather<|channel|>commentary <|constrain|>json<|message|>{"location":"San Francisco, CA"}<|call|>
-}
-```
-
-But if you want to append this to a chat, you'll need to format it as a chat message dict, like this:
-
-```json
-{
-  "role": "assistant",
-  "thinking": "The user asks: \"What is the weather like in SF?\" We need to get the location of the user? The user explicitly asks about SF (San Francisco). So we need to get the current weather in San Francisco, CA. We need to call get_current_weather function. But we need to call function to get weather data. So we should call get_current_weather with location \"San Francisco, CA\". Let's do that.",
-  "tool_calls": [
-    {
-      "name": "get_current_weather",
-      "arguments": {
-        "location": "San Francisco, CA"
-      }
-    }
-  ]
-}
-```
-
-Chat **templates** give us a way to turn messages into formatted input for a model, but we need something else to
-parse model output back into a standard message dict. This is what chat **parsing** is for.
-
-## The [parse_response](~PreTrainedTokenizerBase.parse_response) method
-
-Parsing a chat response on a model that supports it is straightforward. Simply take the raw, decoded output from
-[generate](`~generation.GenerationMixin.generate`), and pass it to the tokenizer's [parse_response](~PreTrainedTokenizerBase.parse_response) method:
-
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-checkpoint = "HuggingFaceTB/SmolLM3-3B"
-
-tokenizer = AutoTokenizer.from_pretrained(checkpoint)
-model = AutoModelForCausalLM.from_pretrained(checkpoint, dtype="auto", device_map="auto")
-
-messages = [
-    {
-        "role": "user",
-        "content": "Hey! Can you summarize the end of the Cold War as briefly as possible? Like, comically briefly. It should really leave out almost most of the relevant information."
-    }
-]
-
-input_ids = tokenizer.apply_chat_template(
-    messages,
-    add_generation_prompt=True,
-    tokenize=True,
-    return_tensors="pt"
-).to(model.device)
-
-outputs = model.generate(input_ids, max_new_tokens=1024)[0, input_ids.shape[1]:]
-out_text = tokenizer.decode(outputs)
-parsed = tokenizer.parse_response(out_text)
-print(parsed.keys())
-```
-
-And you should get:
-
-```text
-dict_keys(['thinking', 'content'])
-```
-
-And that's all you need to start using response parsing! `parse_response` should return a complete message dict that is ready to be appended to the chat history. 
-When the tokenizer does not support response parsing, `parse_response` will throw an error. We hope to add support
-to more tokenizers over time.
-
-## Developers: Understanding a simple response schema
-
-Under the hood, `parse_response` uses a **JSON schema** to parse the model output. A JSON schema represents
-the structure of the output message dict. The schema is augmented with additional fields that indicate how the 
-output message string should be parsed into the expected format. Let's take a look at the schema for a SmolLM response,
-excluding tool calls for now:
-
-```python
-{
-    "x-regex": "(?:<think>\n?(?P<thinking>.+?)\n?</think>)?\s*(?P<content>.+?)?\s*(?:<\|im_end\|>|$)",
-    "type": "object",
-    "properties": {
-        "role": {"const": "assistant"},
-        "content": {"type": "string"},
-        "thinking": {"type": "string"}
-    }
-}
-```
-
-We can see that the schema describes a JSON "object" (a `dict`, in other words) with three keys: `role`, `content`, and `thinking`.
-Because all assistant responses have the role "assistant", the `role` key is a `const`(ant). The other two keys are strings, extracted
-from the named groups in the regex in the `x-regex` field.
-
-Like chat templates, response schemas are set as a property of the tokenizer. To enable response parsing, all you need
-to do is set `tokenizer.response_schema` to a valid schema dict, and `tokenizer.parse_response()` will work! Again, like
-chat templates, this schema will be saved with the processor, so once you set it, you can use `save_pretrained()` or `push_to_hub()` to
-save and share the schema. 
-
-## Developers: Complex schemas
-
-Now, let's look at a more complex schema, which includes tool calls, to gain more of an understanding of the parser
-internals. For this, we'll use the `GPT-OSS` schema. GPT-OSS emits both tool calls and thinking blocks, and it uses
-an unusual format where model responses are tagged with one of three "channels": `commentary` for things like
-tool calls, `analysis` for chain of thought blocks, and `final` for messages intended to be sent to the user. 
-A full message where the model calls a tool named `get_current_weather` might look like this, with some extra linebreaks added for clarity:
-
-```text
-<|channel|>analysis<|message|>
-The user asks: "What is the weather like in SF?" So we need to get the current weather in San Francisco, CA. 
-We need to call get_current_weather function. So we should call get_current_weather with location "San Francisco, CA".
-<|end|>
-<|start|>assistant<|channel|>commentary 
-to=functions.get_current_weather <|constrain|>json<|message|>
-{
-  "location": "San Francisco, CA"
-}
-<|call|>
-```
-
-Parsing proceeds recursively; the output of a regex (or other parser) at one level becomes the input to the nodes below it.
-In other words, don't feel like you have to parse the entire output in one enormous regex! Instead, start with the schema,
-and then add regexes to extract the relevant chunks as you go. Here's a schema that will parse it, with some
-explanatory comments:
-
-```python
-{
-    "type": "object",
-    "properties": {
-        "role": {"const": "assistant"},
-        # "content" and "thinking" are both similar to the previous example, and just extract a single string
-        # However, rather than using a single regex with named groups to extract both, we use a regex in each subkey.
-        # When an object node has no parser/regex, the entire input string is passed to all of its children, so 
-        # parsing can either be done with named groups at the object level, or with separate regexes at the property level.
-        "content": {"type": "string", "x-regex": r"<\|channel\|>final<\|message\|>(.*?)(?:<\|end\|>|$)"},
-        "thinking": {"type": "string", "x-regex": r"<\|channel\|>analysis<\|message\|>(.*?)<\|end\|>"},
-        "tool_calls": {
-            # "x-regex-iterator" uses re.findall to find multiple possible manages, and returns them as an
-            # array/list. You don't need to worry about array handling, though - each item in the array will be
-            # parsed by the `items` schema, so just write the schema for a single item.
-            "x-regex-iterator": r"<\|channel\|>commentary (to=functions\..*?<\|message\|>.*?)(?:<\|call\|>|$)",
-            "type": "array",
-            "items": {
-                "type": "object",
-                "properties": {
-                    # A const property is a fixed value, and the input has no effect on it.
-                    "type": {"const": "function"},
-                    # Here, we wrap the entire tool call dict in a `{"function": ...}` block. The input string is passed through to it unchanged.
-                    "function": {
-                        "type": "object",
-                        "properties": {
-                            "name": {"type": "string", "x-regex": r"^to=functions\.(\w+)"},
-                            "arguments": {
-                                "type": "object",
-                                "x-regex": "<\|message\|>(.*)",
-                                # The "x-parser" field indicates that the extracted string should be parsed as JSON.
-                                # The output is then passed to the schema nodes below and recursive parsing continues.
-                                "x-parser": "json",
-                                "additionalProperties": {"type": "any"},
-                            },
-                        },
-                    },
-                },
-            },
-        },
-    },
-}
-```
-
-## Developers: Understanding the parser logic
-
-The parser follows a few simple rules:
-
-1. Each level of the schema receives input from the level above, applies any regex or parser it has, and then passes the output to its children.
-2. The root level receives the entire decoded model output string as input.
-3. If a node has structured content after parsing (for example, if the regex has named groups and returns a dict, or if the parser returns a dict or list),
-   then that structured content is mapped to the node's children, and each child node receives its corresponding value as input.
-4. If an `object` (dict) node has unstructured (string) output, then the entire string is passed to all of its children. This allows child nodes
-   to handle parsing individually rather than requiring a single parent regex to extract all keys at once.
-5. If an `array` (list) node has unstructured (string) output, then this throws an error.
-
-There is a small set of allowable `x-` keys that indicate how parsing should be done at each node:
-- `x-regex`: A regex string to apply to the input. If the regex has named groups, the output is a dict of group names to values. Named groups should only be used in `object` nodes.
-  Otherwise, the regex must have exactly one unnamed capturing group, and the output is the value of that group as a string.
-- `x-regex-iterator`: A regex string to apply to the input using `re.findall()`. The output is a list of all matches.
-  This should only be used in `array` nodes, and the regex must have exactly one unnamed capturing group. The output is distributed to
-  the node's `items` schema.
-- `x-parser`: Calls a built-in parser to apply to the input. Currently, the only supported parser is `json`, which parses the input string as JSON.
-  The output is passed to the child nodes for further parsing. Note that the `json` parser can return deeply nested output - in this case, the output
-  will be progressively unwrapped as it is passed through child nodes. The child nodes do not need additional `x-parser` or `x-regex` fields in this case, 
-  but their structure must match the structure of the parsed JSON.
-- `x-parser-args`: Only allowed in conjunction with `x-parser`. This is a dict of additional arguments that control parsing. Right now, the only supported
-  argument is `transform`, which specifies a `jmespath` transformation to apply to the output. This is useful when the JSON parser returns a structure
-  that needs to be modified to match the schema.
-- `x-regex-key-value`: This is rarely necessary, but it can be useful when parsing key-value pairs in non-JSON format where the names of the keys are not known
-  in advance, such as when a model emits XML tool calls with arbitrary argument names. The regex must have exactly two named capturing groups, 
-  `key` and `value`, and the output is a dict mapping keys to values. This should only be used in `object` nodes.
-
-In general, multiple regexes/parsers cannot be combined at the same level. The exception is that `x-regex`, returning a single string, can be combined with the other parsers. In this case,
-`x-regex` is applied first, and then the output is passed to the other parser, either `x-regex-iterator`, `x-parser`, or `x-regex-key-value`.
-
-Putting these ideas together, you can see that the input flows through the schema, being parsed at each level and then distributed to child nodes. Each level
-only needs to extract the input content that is relevant for that part of the schema, and can then let its child nodes handle the rest. Internally, this is handled
-with a parser function that receives input, applies any regexes/parsers at the current level, then maps the result to its child nodes before recursively calling itself on each of them.
-Recursion terminates when it reaches leaf nodes, usually primitive types like `string` or `number`, which simply return the input they receive.

docs/source/en/chat_templating.md

@@ -77,9 +77,9 @@ Mistral-7B-Instruct uses `[INST]` and `[/INST]` tokens to indicate the start and
 
 The input to `apply_chat_template` should be structured as a list of dictionaries with `role` and `content` keys. The `role` key specifies the speaker, and the `content` key contains the message. The common roles are:
 
-- `user` for messages from the user
-- `assistant` for messages from the model
-- `system` for directives on how the model should act (usually placed at the beginning of the chat)
+ - `user` for messages from the user
+ - `assistant` for messages from the model
+ - `system` for directives on how the model should act (usually placed at the beginning of the chat)
 
 [`apply_chat_template`] takes this list and returns a formatted sequence. Set `tokenize=True` if you want to tokenize the sequence.
 

docs/source/en/community.md

@@ -6,13 +6,13 @@ rendered properly in your Markdown viewer.
 
 This page regroups resources around 🤗 Transformers developed by the community.
 
-## Community resources
+## Community resources:
 
 | Resource     |      Description      |      Author      |
 |:----------|:-------------|------:|
 | [Hugging Face Transformers Glossary Flashcards](https://www.darigovresearch.com/huggingface-transformers-glossary-flashcards) | A set of flashcards based on the [Transformers Docs Glossary](glossary) that has been put into a form which can be easily learned/revised using [Anki](https://apps.ankiweb.net/) an open source, cross platform app specifically designed for long term knowledge retention. See this [Introductory video on how to use the flashcards](https://www.youtube.com/watch?v=Dji_h7PILrw). | [Darigov Research](https://www.darigovresearch.com/) |
 
-## Community notebooks
+## Community notebooks:
 
 | Notebook     |      Description      |      Author      |      |
 |:----------|:-------------|:-------------|------:|

docs/source/en/continuous_batching.md

@@ -1,98 +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.
-
--->
-
----
-# Brainstorm
-
-## Persona
-
-Model developer that wants to evaluate his model implementation on a dataset or a model "trainer" that wants to run inference for his GRPO policy.
-Pre reqs to understand the docs:
-- knows what KV Cache is
-- familiarity with transformers and infernece
-
-## what we want do include in the doc
-
-- CB usage examples
-- CB API reference
-- light refresher on what is CB + links to blog post
-
-- installation / setup instructions
-
-- open telemetry support
-
-- subsection in Transformers > Inference
-
-- supported & unsupported features
-
-- performance considerations
-  - note on benchmarks (CI + space)
-  - cuda graphs
-  - compile
-  - attn impl
-
-- DO NOT mention vllm
-
-- explicit intended use cases, the why of CB in transformers
-
-- integration with serving
----
-
-
-# Continuous Batching
-
-Continuous Batching (CB) is an advanced technique to optimize the inference of transformer models by dynamically grouping multiple requests into batches. This approach maximizes GPU utilization and throughput, specifically for workloads with many variable-length inputs.
-
-We are particularly interested in having Continuous Batching in transformers for the following use cases:
-- Evaluation of models on large datasets with variable-length inputs
-- Generating outputs for multiple sequences for GRPO policies
-
-CB is what makes inference engines like vLLM or SGLang efficient. That being said, transformers does not aim to be a production-ready inference engine, but a complete framework for model development. For this reason, CB is available in `transformers serve`.
-
-If you are not familiar with some of the core concepts CB is built upon, we invite you to read the associated blog post: [Continuous Batching: Efficient Inference for Large Language Models](https://huggingface.co/blog/continuous-batching). _broken link for now_
-
-## Installation
-
-Nothing to do, it comes built-in with `transformers`! :nice:
-
-## API Reference
-
-## Usage Examples
-
-The main way to use CB in transformers is via the `generate_batch` method.
-
-Unlike `generate`, CB takes already tokenized inputs, known as input IDs. Each sequence of input IDs is represented as a list of integers, in python: `list[int]`. Since 
-
-For a more detailed example, please refer to: [examples/continuous_batching](./path/to/example)
-
-### `generate_batch` example
-
-### `ContinuousBatchingManager` example
-
-
-## Supported & Unsupported Features
-
-### Supported Features
-
-
-### Unsupported Features
-
-## Performance Considerations
-
-
-## Integration with Serving
-
-

docs/source/en/cursor.md

@@ -21,7 +21,6 @@ where `port` is the port used by `transformers serve` (`8000` by default). On th
 </h3>
 
 You're now ready to set things up on the app side! In Cursor, while you can't set a new provider, you can change the endpoint for OpenAI requests in the model selection settings. First, navigate to "Settings" > "Cursor Settings", "Models" tab, and expand the "API Keys" collapsible. To set your `transformers serve` endpoint, follow this order:
-
 1. Unselect ALL models in the list above (e.g. `gpt4`, ...);
 2. Add and select the model you want to use (e.g. `Qwen/Qwen3-4B`)
 3. Add some random text to OpenAI API Key. This field won't be used, but it can't be empty;

docs/source/en/custom_models.md

@@ -25,12 +25,12 @@ This guide will show you how to customize a ResNet model, enable [AutoClass](./m
 
 ## Configuration
 
-A configuration, given by the base [`PreTrainedConfig`] class, contains all the necessary information to build a model. This is where you'll configure the attributes of the custom ResNet model. Different attributes gives different ResNet model types.
+A configuration, given by the base [`PretrainedConfig`] class, contains all the necessary information to build a model. This is where you'll configure the attributes of the custom ResNet model. Different attributes gives different ResNet model types.
 
 The main rules for customizing a configuration are:
 
-1. A custom configuration must subclass [`PreTrainedConfig`]. This ensures a custom model has all the functionality of a Transformers' model such as [`~PreTrainedConfig.from_pretrained`], [`~PreTrainedConfig.save_pretrained`], and [`~PreTrainedConfig.push_to_hub`].
-2. The [`PreTrainedConfig`] `__init__` must accept any `kwargs` and they must be passed to the superclass `__init__`. [`PreTrainedConfig`] has more fields than the ones set in your custom configuration, so when you load a configuration with [`~PreTrainedConfig.from_pretrained`], those fields need to be accepted by your configuration and passed to the superclass.
+1. A custom configuration must subclass [`PretrainedConfig`]. This ensures a custom model has all the functionality of a Transformers' model such as [`~PretrainedConfig.from_pretrained`], [`~PretrainedConfig.save_pretrained`], and [`~PretrainedConfig.push_to_hub`].
+2. The [`PretrainedConfig`] `__init__` must accept any `kwargs` and they must be passed to the superclass `__init__`. [`PretrainedConfig`] has more fields than the ones set in your custom configuration, so when you load a configuration with [`~PretrainedConfig.from_pretrained`], those fields need to be accepted by your configuration and passed to the superclass.
 
 > [!TIP]
 > It is useful to check the validity of some of the parameters. In the example below, a check is implemented to ensure `block_type` and `stem_type` belong to one of the predefined values.
@@ -38,10 +38,10 @@ The main rules for customizing a configuration are:
 > Add `model_type` to the configuration class to enable [AutoClass](./models#autoclass) support.
 
 ```py
-from transformers import PreTrainedConfig
+from transformers import PretrainedConfig
 from typing import List
 
-class ResnetConfig(PreTrainedConfig):
+class ResnetConfig(PretrainedConfig):
     model_type = "resnet"
 
     def __init__(
@@ -74,7 +74,7 @@ class ResnetConfig(PreTrainedConfig):
         super().__init__(**kwargs)
 ```
 
-Save the configuration to a JSON file in your custom model folder, `custom-resnet`, with [`~PreTrainedConfig.save_pretrained`].
+Save the configuration to a JSON file in your custom model folder, `custom-resnet`, with [`~PretrainedConfig.save_pretrained`].
 
 ```py
 resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
@@ -83,7 +83,7 @@ resnet50d_config.save_pretrained("custom-resnet")
 
 ## Model
 
-With the custom ResNet configuration, you can now create and customize the model. The model subclasses the base [`PreTrainedModel`] class. Like [`PreTrainedConfig`], inheriting from [`PreTrainedModel`] and initializing the superclass with the configuration extends Transformers' functionalities such as saving and loading to the custom model.
+With the custom ResNet configuration, you can now create and customize the model. The model subclasses the base [`PreTrainedModel`] class. Like [`PretrainedConfig`], inheriting from [`PreTrainedModel`] and initializing the superclass with the configuration extends Transformers' functionalities such as saving and loading to the custom model.
 
 Transformers' models follow the convention of accepting a `config` object in the `__init__` method. This passes the entire `config` to the model sublayers, instead of breaking the `config` object into multiple arguments that are individually passed to the sublayers.
 
@@ -235,7 +235,7 @@ from resnet_model.configuration_resnet import ResnetConfig
 from resnet_model.modeling_resnet import ResnetModel, ResnetModelForImageClassification
 ```
 
-Copy the code from the model and configuration files. To make sure the AutoClass objects are saved with [`~PreTrainedModel.save_pretrained`], call the [`~PreTrainedConfig.register_for_auto_class`] method. This modifies the configuration JSON file to include the AutoClass objects and mapping.
+Copy the code from the model and configuration files. To make sure the AutoClass objects are saved with [`~PreTrainedModel.save_pretrained`], call the [`~PretrainedConfig.register_for_auto_class`] method. This modifies the configuration JSON file to include the AutoClass objects and mapping.
 
 For a model, pick the appropriate `AutoModelFor` class based on the task.
 

docs/source/en/deepspeed.md

@@ -452,7 +452,7 @@ For Ampere GPUs and PyTorch 1.7+, the more efficient [tf32](https://pytorch.org/
 </hfoption>
 <hfoption id="fp16">
 
-To configure fp16 mixed precision, set up the config as shown below with `"auto"` or your own values. [`Trainer`] automatically enables or disables fp16 based on the value of `fp16` or `fp16_full_eval`, and the rest of the config can be set by you. fp16 is enabled from the command line when the following arguments are passed: `--fp16` or `--fp16_full_eval` also.
+To configure AMP-like fp16 mixed precision, set up the config as shown below with `"auto"` or your own values. [`Trainer`] automatically enables or disables fp16 based on the value of `fp16_backend`, and the rest of the config can be set by you. fp16 is enabled from the command line when the following arguments are passed: `--fp16`, `--fp16_backend amp` or `--fp16_full_eval`.
 
 ```json
 {
@@ -469,6 +469,17 @@ To configure fp16 mixed precision, set up the config as shown below with `"auto"
 
 For additional DeepSpeed fp16 training options, take a look at the [FP16 Training Options](https://www.deepspeed.ai/docs/config-json/#fp16-training-options) reference.
 
+To configure Apex-like fp16 mixed precision, set up the config as shown below with `"auto"` or your own values. [`Trainer`] automatically configures `amp` based on the values of `fp16_backend` and `fp16_opt_level`. It can also be enabled from the command line when the following arguments are passed: `--fp16`, `--fp16_backend apex` or `--fp16_opt_level 01`.
+
+```json
+{
+    "amp": {
+        "enabled": "auto",
+        "opt_level": "auto"
+    }
+}
+```
+
 </hfoption>
 <hfoption id="bf16">
 
@@ -593,7 +604,7 @@ To deploy DeepSpeed on multiple GPUs, add `--num_gpus`. You don't need to add `-
 deepspeed --num_gpus=2 examples/pytorch/translation/run_translation.py \
 --deepspeed tests/deepspeed/ds_config_zero3.json \
 --model_name_or_path google-t5/t5-small --per_device_train_batch_size 1 \
---output_dir output_dir --fp16 \
+--output_dir output_dir --overwrite_output_dir --fp16 \
 --do_train --max_train_samples 500 --num_train_epochs 1 \
 --dataset_name wmt16 --dataset_config "ro-en" \
 --source_lang en --target_lang ro
@@ -616,7 +627,7 @@ To deploy DeepSpeed on a single GPU, add `--num_gpus`. You don't need to add `--
 deepspeed --num_gpus=1 examples/pytorch/translation/run_translation.py \
 --deepspeed tests/deepspeed/ds_config_zero2.json \
 --model_name_or_path google-t5/t5-small --per_device_train_batch_size 1 \
---output_dir output_dir --fp16 \
+--output_dir output_dir --overwrite_output_dir --fp16 \
 --do_train --max_train_samples 500 --num_train_epochs 1 \
 --dataset_name wmt16 --dataset_config "ro-en" \
 --source_lang en --target_lang ro

docs/source/en/executorch.md

@@ -16,18 +16,44 @@ rendered properly in your Markdown viewer.
 
 # ExecuTorch
 
-[ExecuTorch](https://pytorch.org/executorch/stable/index.html) runs PyTorch models on mobile and edge devices. Export your Transformers models to the ExecuTorch format with [Optimum ExecuTorch](https://github.com/huggingface/optimum-executorch) with the command below.
+[ExecuTorch](https://pytorch.org/executorch/stable/index.html) is a platform that enables PyTorch training and inference programs to be run on mobile and edge devices. It is powered by [torch.compile](https://pytorch.org/docs/stable/torch.compiler.html) and [torch.export](https://pytorch.org/docs/main/export.html) for performance and deployment.
 
-```bash
-optimum-cli export executorch \
-    --model "HuggingFaceTB/SmolLM2-135M-Instruct" \
-    --task "text-generation" \
-    --recipe "xnnpack" \
-    --use_custom_sdpa \
-    --use_custom_kv_cache \
-    --qlinear 8da4w \
-    --qembedding 8w \
-    --output_dir="hf_smollm2"
+You can use ExecuTorch with Transformers with [torch.export](https://pytorch.org/docs/main/export.html). The [`~transformers.convert_and_export_with_cache`] method converts a [`PreTrainedModel`] into an exportable module. Under the hood, it uses [torch.export](https://pytorch.org/docs/main/export.html) to export the model, ensuring compatibility with ExecuTorch.
+
+```py
+import torch
+from transformers import LlamaForCausalLM, AutoTokenizer, GenerationConfig
+from transformers.integrations.executorch import(
+    TorchExportableModuleWithStaticCache,
+    convert_and_export_with_cache
+)
+
+generation_config = GenerationConfig(
+    use_cache=True,
+    cache_implementation="static",
+    cache_config={
+        "batch_size": 1,
+        "max_cache_len": 20,
+    }
+)
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B", pad_token="</s>", padding_side="right")
+model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B", device_map="auto", dtype=torch.bfloat16, attn_implementation="sdpa", generation_config=generation_config)
+
+exported_program = convert_and_export_with_cache(model)
 ```
 
-Run `optimum-cli export executorch --help` to see all export options. For detailed export instructions, check the [README](optimum/exporters/executorch/README.md).
+The exported PyTorch model is now ready to be used with ExecuTorch. Wrap the model with [`~transformers.TorchExportableModuleWithStaticCache`] to generate text.
+
+```py
+prompts = ["Simply put, the theory of relativity states that "]
+prompt_tokens = tokenizer(prompts, return_tensors="pt", padding=True).to(model.device)
+prompt_token_ids = prompt_tokens["input_ids"]
+
+generated_ids = TorchExportableModuleWithStaticCache.generate(
+    exported_program=exported_program, prompt_token_ids=prompt_token_ids, max_new_tokens=20,
+)
+generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+print(generated_text)
+['Simply put, the theory of relativity states that 1) the speed of light is the']
+```

docs/source/en/generation_strategies.md

@@ -32,10 +32,9 @@ Greedy search works well for tasks with relatively short outputs where creativit
 
 ```py
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from accelerate import Accelerator
+from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device
 
-device = Accelerator().device
+device = infer_device()
 
 tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
 inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to(device)
@@ -55,10 +54,9 @@ Enable multinomial sampling with `do_sample=True` and `num_beams=1`.
 
 ```py
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from accelerate import Accelerator
+from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device
 
-device = Accelerator().device
+device = infer_device()
 
 tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
 inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to(device)
@@ -81,10 +79,9 @@ Enable beam search with the `num_beams` parameter (should be greater than 1 othe
 
 ```py
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from accelerate import Accelerator
+from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device
 
-device = Accelerator().device
+device = infer_device()
 
 tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
 inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to(device)
@@ -169,10 +166,9 @@ Enable prompt lookup decoding with the `prompt_lookup_num_tokens` parameter.
 
 ```py
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from accelerate import Accelerator
+from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device
 
-device = Accelerator().device
+device = infer_device()
 
 tokenizer = AutoTokenizer.from_pretrained("HuggingFaceTB/SmolLM-1.7B")
 model = AutoModelForCausalLM.from_pretrained("HuggingFaceTB/SmolLM-1.7B", dtype=torch.float16).to(device)
@@ -233,7 +229,6 @@ tokenizer.batch_decode(outputs, skip_special_tokens=True)
 ## Custom generation methods
 
 Custom generation methods enable specialized behavior such as:
-
 - have the model continue thinking if it is uncertain;
 - roll back generation if the model gets stuck;
 - handle special tokens with custom logic;
@@ -306,7 +301,6 @@ Updating your Python requirements accordingly will remove this error message.
 ### Creating a custom generation method
 
 To create a new generation method, you need to create a new [**Model**](https://huggingface.co/new) repository and push a few files into it.
-
 1. The model you've designed your generation method with.
 2. `custom_generate/generate.py`, which contains all the logic for your custom generation method.
 3. `custom_generate/requirements.txt`, used to optionally add new Python requirements and/or lock specific versions to correctly use your method.
@@ -383,7 +377,6 @@ def generate(model, input_ids, generation_config=None, left_padding=None, **kwar
 ```
 
 Follow the recommended practices below to ensure your custom generation method works as expected.
-
 - Feel free to reuse the logic for validation and input preparation in the original [`~GenerationMixin.generate`].
 - Pin the `transformers` version in the requirements if you use any private method/attribute in `model`.
 - Consider adding model validation, input validation, or even a separate test file to help users sanity-check your code in their environment.
@@ -417,7 +410,6 @@ tags:
 ```
 
 Recommended practices:
-
 - Document input and output differences in [`~GenerationMixin.generate`].
 - Add self-contained examples to enable quick experimentation.
 - Describe soft-requirements such as if the method only works well with a certain family of models.
@@ -450,7 +442,6 @@ output = model.generate(
 ### Finding custom generation methods
 
 You can find all custom generation methods by [searching for their custom tag.](https://huggingface.co/models?other=custom_generate), `custom_generate`. In addition to the tag, we curate two collections of `custom_generate` methods:
-
 - [Custom generation methods - Community](https://huggingface.co/collections/transformers-community/custom-generation-methods-community-6888fb1da0efbc592d3a8ab6) -- a collection of powerful methods contributed by the community;
 - [Custom generation methods - Tutorials](https://huggingface.co/collections/transformers-community/custom-generation-methods-tutorials-6823589657a94940ea02cfec) -- a collection of reference implementations for methods that previously were part of `transformers`, as well as tutorials for `custom_generate`.
 

docs/source/en/glossary.md

@@ -185,9 +185,9 @@ See the [Fine-tune a pretrained model](https://huggingface.co/docs/transformers/
 
 The model head refers to the last layer of a neural network that accepts the raw hidden states and projects them onto a different dimension. There is a different model head for each task. For example:
 
-* [`GPT2ForSequenceClassification`] is a sequence classification head - a linear layer - on top of the base [`GPT2Model`].
-* [`ViTForImageClassification`] is an image classification head - a linear layer on top of the final hidden state of the `CLS` token - on top of the base [`ViTModel`].
-* [`Wav2Vec2ForCTC`] is a language modeling head with [CTC](#connectionist-temporal-classification-ctc) on top of the base [`Wav2Vec2Model`].
+  * [`GPT2ForSequenceClassification`] is a sequence classification head - a linear layer - on top of the base [`GPT2Model`].
+  * [`ViTForImageClassification`] is an image classification head - a linear layer on top of the final hidden state of the `CLS` token - on top of the base [`ViTModel`].
+  * [`Wav2Vec2ForCTC`] is a language modeling head with [CTC](#connectionist-temporal-classification-ctc) on top of the base [`Wav2Vec2Model`].
 
 ## I
 

docs/source/en/how_to_hack_models.md

@@ -149,4 +149,4 @@ Call [print_trainable_parameters](https://huggingface.co/docs/peft/package_refer
 ```py
 model.print_trainable_parameters()
 "trainable params: 589,824 || all params: 94,274,096 || trainable%: 0.6256"
-```
+```

docs/source/en/hpo_train.md

@@ -15,12 +15,15 @@ rendered properly in your Markdown viewer.
 
 # Hyperparameter search
 
-Hyperparameter search discovers an optimal set of hyperparameters that produces the best model performance. [`Trainer`] supports several hyperparameter search backends - [Optuna](https://optuna.readthedocs.io/en/stable/index.html), [Weights & Biases](https://docs.wandb.ai/), [Ray Tune](https://docs.ray.io/en/latest/tune/index.html) - through  [`~Trainer.hyperparameter_search`] to optimize an objective or even multiple objectives.
+Hyperparameter search discovers an optimal set of hyperparameters that produces the best model performance. [`Trainer`] supports several hyperparameter search backends - [Optuna](https://optuna.readthedocs.io/en/stable/index.html), [SigOpt](https://docs.sigopt.com/), [Weights & Biases](https://docs.wandb.ai/), [Ray Tune](https://docs.ray.io/en/latest/tune/index.html) - through  [`~Trainer.hyperparameter_search`] to optimize an objective or even multiple objectives.
 
 This guide will go over how to set up a hyperparameter search for each of the backends.
 
+> [!WARNING]
+> [SigOpt](https://github.com/sigopt/sigopt-server) is in public archive mode and is no longer actively maintained. Try using Optuna, Weights & Biases or Ray Tune instead.
+
 ```bash
-pip install optuna/wandb/ray[tune]
+pip install optuna/sigopt/wandb/ray[tune]
 ```
 
 To use [`~Trainer.hyperparameter_search`], you need to create a `model_init` function. This function includes basic model information (arguments and configuration) because it needs to be reinitialized for each search trial in the run.
@@ -37,6 +40,7 @@ def model_init(trial):
         config=config,
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
+        token=True if model_args.use_auth_token else None,
     )
 ```
 
@@ -105,7 +109,31 @@ best_trials = trainer.hyperparameter_search(
     n_trials=20,
     compute_objective=compute_objective,
 )
+```
 
+</hfoption>
+<hfoption id="SigOpt">
+
+[SigOpt](https://docs.sigopt.com/ai-module-api-references/api_reference/objects/object_parameter) optimizes double, integer, and categorical parameters.
+
+```py
+def sigopt_hp_space(trial):
+    return [
+        {"bounds": {"min": 1e-6, "max": 1e-4}, "name": "learning_rate", "type": "double"},
+        {
+            "categorical_values": ["16", "32", "64", "128"],
+            "name": "per_device_train_batch_size",
+            "type": "categorical",
+        },
+    ]
+
+best_trials = trainer.hyperparameter_search(
+    direction=["minimize", "maximize"],
+    backend="sigopt",
+    hp_space=sigopt_hp_space,
+    n_trials=20,
+    compute_objective=compute_objective,
+)
 ```
 
 </hfoption>
@@ -138,4 +166,4 @@ best_trials = trainer.hyperparameter_search(
 
 ## Distributed Data Parallel
 
-[`Trainer`] only supports hyperparameter search for distributed data parallel (DDP) on the Optuna backends. Only the rank-zero process is used to generate the search trial, and the resulting parameters are passed along to the other ranks.
+[`Trainer`] only supports hyperparameter search for distributed data parallel (DDP) on the Optuna and SigOpt backends. Only the rank-zero process is used to generate the search trial, and the resulting parameters are passed along to the other ranks.

docs/source/en/index.md

@@ -34,8 +34,6 @@ There are over 1M+ Transformers [model checkpoints](https://huggingface.co/model
 
 Explore the [Hub](https://huggingface.com/) today to find a model and use Transformers to help you get started right away.
 
-Explore the [Models Timeline](./models_timeline) to discover the latest text, vision, audio and multimodal model architectures in Transformers.
-
 ## Features
 
 Transformers provides everything you need for inference or training with state-of-the-art pretrained models. Some of the main features include:
@@ -62,4 +60,4 @@ Transformers is designed for developers and machine learning engineers and resea
 
 ## Learn
 
-If you're new to Transformers or want to learn more about transformer models, we recommend starting with the [LLM course](https://huggingface.co/learn/llm-course/chapter1/1?fw=pt). This comprehensive course covers everything from the fundamentals of how transformer models work to practical applications across various tasks. You'll learn the complete workflow, from curating high-quality datasets to fine-tuning large language models and implementing reasoning capabilities. The course contains both theoretical and hands-on exercises to build a solid foundational knowledge of transformer models as you learn.
+If you're new to Transformers or want to learn more about transformer models, we recommend starting with the [LLM course](https://huggingface.co/learn/llm-course/chapter1/1?fw=pt). This comprehensive course covers everything from the fundamentals of how transformer models work to practical applications across various tasks. You'll learn the complete workflow, from curating high-quality datasets to fine-tuning large language models and implementing reasoning capabilities. The course contains both theoretical and hands-on exercises to build a solid foundational knowledge of transformer models as you learn.

docs/source/en/internal/file_utils.md

@@ -43,3 +43,4 @@ Most of those are only useful if you are studying the general code in the librar
 ## Other Utilities
 
 [[autodoc]] utils._LazyModule
+[[autodoc]] pytorch_utils.infer_device

docs/source/en/internal/generation_utils.md

@@ -153,9 +153,6 @@ generation.
 [[autodoc]] TemperatureLogitsWarper
     - __call__
 
-[[autodoc]] TopHLogitsWarper
-    - __call__
-
 [[autodoc]] TopKLogitsWarper
     - __call__
 
@@ -196,6 +193,28 @@ A [`StoppingCriteria`] can be used to change when to stop generation (other than
 [[autodoc]] EosTokenCriteria
     - __call__
 
+## Constraints
+
+A [`Constraint`] can be used to force the generation to include specific tokens or sequences in the output. Please note that this is exclusively available to our PyTorch implementations.
+
+[[autodoc]] Constraint
+
+[[autodoc]] PhrasalConstraint
+
+[[autodoc]] DisjunctiveConstraint
+
+[[autodoc]] ConstraintListState
+
+## BeamSearch
+
+[[autodoc]] BeamScorer
+    - process
+    - finalize
+
+[[autodoc]] ConstrainedBeamSearchScorer
+    - process
+    - finalize
+
 ## Streamers
 
 [[autodoc]] TextStreamer
@@ -251,19 +270,19 @@ A [`StoppingCriteria`] can be used to change when to stop generation (other than
     - batch_select_indices
 
 [[autodoc]] DynamicCache
-
-[[autodoc]] StaticCache
+    - to_legacy_cache
+    - from_legacy_cache
 
 [[autodoc]] QuantizedCache
 
-[[autodoc]] EncoderDecoderCache
-
 [[autodoc]] QuantoQuantizedCache
 
 [[autodoc]] HQQQuantizedCache
 
 [[autodoc]] OffloadedCache
 
+[[autodoc]] StaticCache
+
 [[autodoc]] OffloadedStaticCache
 
 [[autodoc]] HybridCache
@@ -272,6 +291,10 @@ A [`StoppingCriteria`] can be used to change when to stop generation (other than
 
 [[autodoc]] SlidingWindowCache
 
+[[autodoc]] EncoderDecoderCache
+    - to_legacy_cache
+    - from_legacy_cache
+
 ## Watermark Utils
 
 [[autodoc]] WatermarkingConfig

docs/source/en/internal/model_debugging_utils.md

@@ -218,9 +218,9 @@ path reference to the associated `.safetensors` file. Each tensor is written to
 the state dictionary. File names are constructed using the `module_path` as a prefix with a few possible postfixes that
 are built recursively.
 
-* Module inputs are denoted with the `_inputs` and outputs by `_outputs`.
-* `list` and `tuple` instances, such as `args` or function return values, will be postfixed with `_{index}`.
-* `dict` instances will be postfixed with `_{key}`.
+*   Module inputs are denoted with the `_inputs` and outputs by `_outputs`.
+*   `list` and `tuple` instances, such as `args` or function return values, will be postfixed with `_{index}`.
+*   `dict` instances will be postfixed with `_{key}`.
 
 ### Comparing between implementations
 
@@ -255,7 +255,6 @@ how many tests are being skipped and for which models.
 When porting models to transformers, tests fail as they should, and sometimes `test_modeling_common` feels irreconcilable with the peculiarities of our brand new model. But how can we be sure we're not breaking everything by adding a seemingly innocent skip?
 
 This utility:
-
 - scans all test_modeling_common methods
 - looks for times where a method is skipped
 - returns a summary json you can load as a DataFrame/inspect
@@ -320,7 +319,7 @@ df.sort_values(by=['skipped_proportion'], ascending=False)
 You can focus on a specific test method using `--test_method_name`:
 
 ```bash
-python utils/scan_skipped_tests.py --test_method_name test_inputs_embeds --output_dir path/to/output
+$ python utils/scan_skipped_tests.py --test_method_name test_inputs_embeds --output_dir path/to/output
 ```
 
 - `--test_method_name`: Name of the test method to scan (e.g., `test_inputs_embeds`).
@@ -345,95 +344,10 @@ Skipped : 124/323 (38.4%)
  - bit: Bit does not use inputs_embeds
  - blip: Blip does not use inputs_embeds
  - blip_2: Inputs_embeds is tested in individual model tests
- - bridgetower:
+ - bridgetower: 
  - canine: CANINE does not have a get_input_embeddings() method.
  - ...
 
 📄 JSON saved to /home/pablo/git/transformers/scan_test_inputs_embeds.json
 
 ```
-
-## Modular model detector
-
-### Code similarity analyzer - for model adders
-
-This utility analyzes code similarities between model implementations to identify opportunities for modularization. It compares a new or existing modeling file against all models in the library using embedding-based and token-based similarity metrics.
-
-### Rationale
-
-When adding a new model to transformers, many components (attention layers, MLPs, outputs, etc.) may already exist in similar form in other models. Instead of implementing everything from scratch, model adders can identify which existing classes are similar and potentially reusable through modularization.
-
-The tool computes two similarity scores:
-
-- **Embedding score**: Uses semantic code embeddings (via `Qwen/Qwen3-Embedding-4B`) to detect functionally similar code even with different naming
-- **Jaccard score**: Measures token set overlap to identify structurally similar code patterns
-
-A score of 1.00 means the code is identical.
-
-### Usage
-
-From the root of the `transformers` repository:
-
-```bash
-python utils/modular_model_detector.py --modeling-file path/to/modeling_file.py
-```
-
-The tool will automatically download the pre-built index from the Hub (requires RAM/VRAM for the embedding model).
-
-**Example output:**
-
-```text
-Loading checkpoint shards: 100%|████████████████████| 2/2 [00:00<00:00, 33.62it/s]
-encoding 21 query definitions with Qwen/Qwen3-Embedding-4B (device=cuda, batch=16, max_length=4096)
-
-stuff.py::Beit3ImageTextMatchingOutput:
-embedding:
-    blip_2::Blip2ImageTextMatchingModelOutput (0.9994)
-    chinese_clip::ChineseCLIPOutput (0.9818)
-    owlvit::OwlViTOutput (0.9818)
-jaccard:
-    owlv2::Owlv2Output (0.9667)
-    metaclip_2::MetaClip2Output (0.9667)
-    altclip::AltCLIPOutput (0.9667)
-intersection:
-    blip::BlipOutput
-    owlvit::OwlViTOutput
-
-stuff.py::Beit3MLP:
-embedding:
-    efficientloftr::EfficientLoFTRMLP (0.9718)
-    seggpt::SegGptMlp (0.9650)
-jaccard:
-    chinese_clip::ChineseCLIPTextSelfOutput (0.5294)
-    bert::BertSelfOutput (0.5294)
-intersection:
-```
-
-The `intersection` field shows classes that appear in both top-5 results, indicating high confidence for modularization candidates.
-
-### Building a custom index
-
-To rebuild the index from your local codebase (useful after adding new models or using a different embedding model):
-
-```bash
-python utils/modular_model_detector.py --build
-```
-
-To push the rebuilt index to a Hub dataset:
-
-```bash
-python utils/modular_model_detector.py --build --push-new-index --hub-dataset your-org/your-dataset
-```
-
-### Options
-
-- `--modeling-file`: Path to the modeling file to analyze
-- `--build`: Build the code similarity index from all modeling files in `src/transformers/models/`
-- `--push-new-index`: After building, push the index to a Hub dataset (requires `--build`)
-- `--hub-dataset`: Hub dataset repository ID to pull/push the index (default: `hf-internal-testing/transformers_code_embeddings`)
-
-### Limitations
-
-This tool requires GPU/CPU resources to run the embedding model (`Qwen/Qwen3-Embedding-4B`). The pre-built index is downloaded from the Hub by default, which requires an internet connection on first use.
-
-Results are suggestions based on code similarity and should be manually reviewed before modularization. High similarity scores don't guarantee perfect compatibility.

docs/source/en/internal/modeling_utils.md

@@ -46,4 +46,10 @@ Most of those are only useful if you are studying the code of the models in the
 
 [[autodoc]] pytorch_utils.apply_chunking_to_forward
 
+[[autodoc]] pytorch_utils.find_pruneable_heads_and_indices
+
+[[autodoc]] pytorch_utils.prune_layer
+
+[[autodoc]] pytorch_utils.prune_conv1d_layer
+
 [[autodoc]] pytorch_utils.prune_linear_layer

docs/source/en/internal/rope_utils.md

@@ -1,83 +0,0 @@
-<!--Copyright 2020 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
-
-⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
-rendered properly in your Markdown viewer.
-
--->
-
-# Utilities for Rotary Embedding
-
-This page explains how the Rotary Embedding is computed and applied in Transformers and what types of RoPE are supported.
-
-## Overview
-
-Rotary Position Embeddings are a technique used to inject positional information into attention mechanisms without relying on explicit position encodings.  
-Instead of adding position vectors to token embeddings, RoPE rotates query and key vectors in the complex plane according to their positions enabling relative positional awareness and better extrapolation to unseen sequence lengths.
-
-The Transformers library provides a flexible and extensible implementation of various RoPE types defined in `[`~modeling_rope_utils.ROPE_VALIDATION_FUNCTIONS`]`, including both the default and scaled variants:
-
-| Rope Type | Description |
-|------------|-------------|
-| `"default"` | Standard rotary embedding as in LLaMA. |
-| `"linear"` | Linear-scaled RoPE which allows longer context windows. |
-| `"dynamic"` | NTK-aware scaling computed by rescaling frequency base (`θ`) for longer context. |
-| `"yarn"` | YaRN scaling variant providing smoother extrapolation and stability. |
-| `"longrope"` | [LongRoPE](https://github.com/microsoft/LongRoPE) scaling as in Phi-2 model series. |
-| `"llama3"` | RoPE scaling as in Llama3.1. |
-
-## Configuration in Model Configs
-
-To enable and customize rotary embeddings, add a `rope_parameters` field to your model’s configuration file (`config.json`). This field controls the RoPE behavior across model layers. Note that each RoPE variant defines its own set of expected keys and missing keys will raise an error. See the example below which creates a llama config with default RoPE parameters:
-
-```python
-from transformers import LlamaConfig
-
-config = LlamaConfig()
-config.rope_parameters = {
-    "rope_type": "default", # type of RoPE to use
-    "rope_theta": 10000.0 # base frequency parameter
-}
-
-# If we want to apply a scaled RoPE type, we need to pass extra parameters
-config.rope_parameters = {
-    "rope_type": "linear",
-    "rope_theta": 10000.0,
-    "factor": 8.0  # scale factor for context extension
-}
-```
-
-## Per-Layer-Type RoPE Configuration
-
-Some models such as Gemma-3 use different layer types with different attention mechanisms, i.e. "full attention" in some blocks and "sliding-window attention" in others. Transformers supports specifying distinct RoPE parameters per layer type for these models. In this case, `rope_parameters` should be a nested dictionary, where top-level keys correspond to `config.layer_types` and values are per-type RoPE parameters. During model initialization, each decoder layer will automatically look up the matching RoPE configuration based on its declared layer type.
-
-```python
-from transformers import Gemma3Config
-
-config = Gemma3Config()
-config.rope_parameters = {
-    "full_attention": {
-        "rope_type": "dynamic",
-        "rope_theta": 1000000.0,
-        "factor": 8.0,
-        "original_max_position_embeddings": 8096,
-    },
-    "sliding_attention": {
-        "rope_type": "default",
-        "rope_theta": 10000.0,
-    }
-}
-```
-
-## Utilities
-
-[[autodoc]] RopeParameters
-    - __call__

docs/source/en/internal/trainer_utils.md

@@ -36,6 +36,10 @@ Most of those are only useful if you are studying the code of the Trainer in the
 
 [[autodoc]] trainer_callback.CallbackHandler
 
+## Distributed Evaluation
+
+[[autodoc]] trainer_pt_utils.DistributedTensorGatherer
+
 ## Trainer Argument Parser
 
 [[autodoc]] HfArgumentParser

docs/source/en/kernel_doc/overview.md

@@ -1,3 +0,0 @@
-# Overview
-
-Kernels in transformers are used to optimize the performance of models with custom layers from the hub and very low effort.

docs/source/en/kv_cache.md

@@ -124,12 +124,11 @@ The example below shows how you can fallback to an offloaded cache if you run ou
 
 ```py
 import torch
-from transformers import AutoTokenizer, AutoModelForCausalLM
-from accelerate import Accelerator
+from transformers import AutoTokenizer, AutoModelForCausalLM, infer_device
 
 def resilient_generate(model, *args, **kwargs):
     oom = False
-    device = Accelerator().device
+    device = infer_device()
     torch_device_module = getattr(torch, device, torch.cuda)
     try:
         return model.generate(*args, **kwargs)
@@ -208,7 +207,7 @@ Some models have a unique way of storing past kv pairs or states that is not com
 
 Mamba models, such as [Mamba](./model_doc/mamba), require a specific cache because the model doesn't have an attention mechanism or kv states. Thus, they are not compatible with the above [`Cache`] classes.
 
-## Iterative generation
+# Iterative generation
 
 A cache can also work in iterative generation settings where there is back-and-forth interaction with a model (chatbots). Like regular generation, iterative generation with a cache allows a model to efficiently handle ongoing conversations without recomputing the entire context at each step.
 

docs/source/en/llm_optims.md

@@ -114,8 +114,7 @@ print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
 Another option for using [`StaticCache`] is to pass it to a models forward pass using the same `past_key_values` argument. This allows you to write your own custom decoding function to decode the next token given the current token, position, and cache position of previously generated tokens.
 
 ```py
-from transformers import LlamaTokenizer, LlamaForCausalLM, StaticCache, logging
-from accelerate import Accelerator
+from transformers import LlamaTokenizer, LlamaForCausalLM, StaticCache, logging, infer_device
 from transformers.testing_utils import CaptureLogger
 import torch
 
@@ -125,7 +124,7 @@ prompts = [
 ]
 
 NUM_TOKENS_TO_GENERATE = 40
-torch_device = Accelerator().device
+torch_device = infer_device()
 
 tokenizer = LlamaTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf", pad_token="</s>", padding_side="right")
 model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", device_map="sequential")
@@ -209,11 +208,10 @@ Enable speculative decoding by loading an assistant model and passing it to [`~G
 <hfoption id="greedy search">
 
 ```py
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from accelerate import Accelerator
+from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device
 import torch
 
-device = Accelerator().device
+device = infer_device()
 
 tokenizer = AutoTokenizer.from_pretrained("facebook/opt-1.3b")
 inputs = tokenizer("Einstein's theory of relativity states", return_tensors="pt").to(device)
@@ -231,11 +229,10 @@ tokenizer.batch_decode(outputs, skip_special_tokens=True)
 For speculative sampling decoding, add the [do_sample](https://hf.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationConfig.do_sample) and [temperature](https://hf.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationConfig.temperature) parameters to [`~GenerationMixin.generate`].
 
 ```py
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from accelerate import Accelerator
+from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device
 import torch
 
-device = Accelerator().device
+device = infer_device()
 
 tokenizer = AutoTokenizer.from_pretrained("facebook/opt-1.3b")
 inputs = tokenizer("Einstein's theory of relativity states", return_tensors="pt").to(device)
@@ -260,11 +257,10 @@ To enable prompt lookup decoding, specify the number of tokens that should be ov
 <hfoption id="greedy decoding">
 
 ```py
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from accelerate import Accelerator
+from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device
 import torch
 
-device = Accelerator().device
+device = infer_device()
 
 tokenizer = AutoTokenizer.from_pretrained("facebook/opt-1.3b")
 inputs = tokenizer("The second law of thermodynamics states", return_tensors="pt").to(device)
@@ -282,11 +278,10 @@ print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
 For prompt lookup decoding with sampling, add the [do_sample](https://hf.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationConfig.do_sample) and [temperature](https://hf.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationConfig.temperature) parameters to [`~GenerationMixin.generate`].
 
 ```py
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from accelerate import Accelerator
+from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device
 import torch
 
-device = Accelerator().device
+device = infer_device()
 
 tokenizer = AutoTokenizer.from_pretrained("facebook/opt-1.3b")
 inputs = tokenizer("The second law of thermodynamics states", return_tensors="pt").to(device)

docs/source/en/llm_tutorial.md

@@ -94,7 +94,6 @@ model.generate(**inputs, num_beams=4, do_sample=True)
 ```
 
 [`~GenerationMixin.generate`] can also be extended with external libraries or custom code:
-
 1. the `logits_processor` parameter accepts custom [`LogitsProcessor`] instances for manipulating the next token probability distribution;
 2. the `stopping_criteria` parameters supports custom [`StoppingCriteria`] to stop text generation;
 3. other custom generation methods can be loaded through the `custom_generate` flag ([docs](generation_strategies.md/#custom-decoding-methods)).
@@ -259,11 +258,11 @@ Some models and tasks expect a certain input prompt format, and if the format is
 For example, a chat model expects the input as a [chat template](./chat_templating). Your prompt should include a `role` and `content` to indicate who is participating in the conversation. If you try to pass your prompt as a single string, the model doesn't always return the expected output.
 
 ```py
-from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
 tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-alpha")
 model = AutoModelForCausalLM.from_pretrained(
-    "HuggingFaceH4/zephyr-7b-alpha", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
+    "HuggingFaceH4/zephyr-7b-alpha", device_map="auto", load_in_4bit=True
 )
 ```
 

docs/source/en/llm_tutorial_optimization.md

@@ -16,18 +16,18 @@ rendered properly in your Markdown viewer.
 Large Language Models (LLMs) such as GPT3/4, [Falcon](https://huggingface.co/tiiuae/falcon-40b), and [Llama](https://huggingface.co/meta-llama/Llama-2-70b-hf) are rapidly advancing in their ability to tackle human-centric tasks, establishing themselves as essential tools in modern knowledge-based industries.
 Deploying these models in real-world tasks remains challenging, however:
 
-- To exhibit near-human text understanding and generation capabilities, LLMs currently require to be composed of billions of parameters (see [Kaplan et al](https://huggingface.co/papers/2001.08361), [Wei et. al](https://huggingface.co/papers/2206.07682)). This consequently amplifies the memory demands for inference.
-- In many real-world tasks, LLMs need to be given extensive contextual information. This necessitates the model's capability to manage very long input sequences during inference.
+-   To exhibit near-human text understanding and generation capabilities, LLMs currently require to be composed of billions of parameters (see [Kaplan et al](https://huggingface.co/papers/2001.08361), [Wei et. al](https://huggingface.co/papers/2206.07682)). This consequently amplifies the memory demands for inference.
+-   In many real-world tasks, LLMs need to be given extensive contextual information. This necessitates the model's capability to manage very long input sequences during inference.
 
 The crux of these challenges lies in augmenting the computational and memory capabilities of LLMs, especially when handling expansive input sequences.
 
 In this guide, we will go over the effective techniques for efficient LLM deployment:
 
-1. **Lower Precision:** Research has shown that operating at reduced numerical precision, namely [8-bit and 4-bit](./main_classes/quantization) can achieve computational advantages without a considerable decline in model performance.
+1.  **Lower Precision:** Research has shown that operating at reduced numerical precision, namely [8-bit and 4-bit](./main_classes/quantization) can achieve computational advantages without a considerable decline in model performance.
 
-2. **Flash Attention:** Flash Attention is a variation of the attention algorithm that not only provides a more memory-efficient approach but also realizes increased efficiency due to optimized GPU memory utilization.
+2.  **Flash Attention:** Flash Attention is a variation of the attention algorithm that not only provides a more memory-efficient approach but also realizes increased efficiency due to optimized GPU memory utilization.
 
-3. **Architectural Innovations:** Considering that LLMs are always deployed in the same way during inference, namely autoregressive text generation with a long input context, specialized model architectures have been proposed that allow for more efficient inference. The most important advancement in model architectures hereby are [Alibi](https://huggingface.co/papers/2108.12409), [Rotary embeddings](https://huggingface.co/papers/2104.09864), [Multi-Query Attention (MQA)](https://huggingface.co/papers/1911.02150) and [Grouped-Query-Attention (GQA)](https://huggingface.co/papers/2305.13245).
+3.  **Architectural Innovations:** Considering that LLMs are always deployed in the same way during inference, namely autoregressive text generation with a long input context, specialized model architectures have been proposed that allow for more efficient inference. The most important advancement in model architectures hereby are [Alibi](https://huggingface.co/papers/2108.12409), [Rotary embeddings](https://huggingface.co/papers/2104.09864), [Multi-Query Attention (MQA)](https://huggingface.co/papers/1911.02150) and [Grouped-Query-Attention (GQA)](https://huggingface.co/papers/2305.13245).
 
 Throughout this guide, we will offer an analysis of auto-regressive generation from a tensor's perspective. We delve into the pros and cons of adopting lower precision, provide a comprehensive exploration of the latest attention algorithms, and discuss improved LLM architectures. While doing so, we run practical examples showcasing each of the feature improvements.
 
@@ -37,22 +37,22 @@ Memory requirements of LLMs can be best understood by seeing the LLM as a set of
 
 At the time of writing this guide, LLMs consist of at least a couple billion parameters. Each parameter thereby is made of a decimal number, e.g. `4.5689` which is usually stored in either [float32](https://en.wikipedia.org/wiki/Single-precision_floating-point_format), [bfloat16](https://en.wikipedia.org/wiki/Bfloat16_floating-point_format), or [float16](https://en.wikipedia.org/wiki/Half-precision_floating-point_format) format. This allows us to easily compute the memory requirement to load the LLM into memory:
 
-> *Loading the weights of a model having X billion parameters requires roughly 4 \* X GB of VRAM in float32 precision*
+> *Loading the weights of a model having X billion parameters requires roughly 4 * X GB of VRAM in float32 precision*
 
 Nowadays, models are however rarely trained in full float32 precision, but usually in bfloat16 precision or less frequently in float16 precision. Therefore the rule of thumb becomes:
 
-> *Loading the weights of a model having X billion parameters requires roughly 2 \* X GB of VRAM in bfloat16/float16 precision*
+> *Loading the weights of a model having X billion parameters requires roughly 2 * X GB of VRAM in bfloat16/float16 precision*
 
 For shorter text inputs (less than 1024 tokens), the memory requirement for inference is very much dominated by the memory requirement to load the weights. Therefore, for now, let's assume that the memory requirement for inference is equal to the memory requirement to load the model into the GPU VRAM.
 
 To give some examples of how much VRAM it roughly takes to load a model in bfloat16:
 
-- **GPT3** requires 2 \* 175 GB = **350 GB** VRAM
-- [**Bloom**](https://huggingface.co/bigscience/bloom) requires 2 \* 176 GB = **352 GB** VRAM
-- [**Llama-2-70b**](https://huggingface.co/meta-llama/Llama-2-70b-hf) requires 2 \* 70 GB = **140 GB** VRAM
-- [**Falcon-40b**](https://huggingface.co/tiiuae/falcon-40b) requires 2 \* 40 GB = **80 GB** VRAM
-- [**MPT-30b**](https://huggingface.co/mosaicml/mpt-30b) requires 2 \* 30 GB = **60 GB** VRAM
-- [**bigcode/starcoder**](https://huggingface.co/bigcode/starcoder) requires 2 \* 15.5 = **31 GB** VRAM
+-   **GPT3** requires 2 \* 175 GB = **350 GB** VRAM
+-   [**Bloom**](https://huggingface.co/bigscience/bloom) requires 2 \* 176 GB = **352 GB** VRAM
+-   [**Llama-2-70b**](https://huggingface.co/meta-llama/Llama-2-70b-hf) requires 2 \* 70 GB = **140 GB** VRAM
+-   [**Falcon-40b**](https://huggingface.co/tiiuae/falcon-40b) requires 2 \* 40 GB = **80 GB** VRAM
+-   [**MPT-30b**](https://huggingface.co/mosaicml/mpt-30b) requires 2 \* 30 GB = **60 GB** VRAM
+-   [**bigcode/starcoder**](https://huggingface.co/bigcode/starcoder) requires 2 \* 15.5 = **31 GB** VRAM
 
 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).
 
@@ -169,11 +169,11 @@ All that matters is that the next token *logit* distribution stays roughly the s
 
 There are various quantization techniques, which we won't discuss in detail here, but in general, all quantization techniques work as follows:
 
-- 1. Quantize all weights to the target precision
-- 2. Load the quantized weights, and pass the input sequence of vectors in bfloat16 precision
-- 3. Dynamically dequantize weights to bfloat16 to perform the computation with their input vectors in bfloat16 precision
+-   1.  Quantize all weights to the target precision
+-   2.  Load the quantized weights, and pass the input sequence of vectors in bfloat16 precision
+-   3.  Dynamically dequantize weights to bfloat16 to perform the computation with their input vectors in bfloat16 precision
 
-In a nutshell, this means that *inputs-weight matrix* multiplications, with $X$ being the *inputs*, $W$ being a weight matrix and $Y$ being the output:
+In a nutshell, this means that *inputs-weight matrix* multiplications, with \\( X \\) being the *inputs*, \\( W \\) being a weight matrix and \\( Y \\) being the output:
 
 $$ Y = X * W $$
 
@@ -194,7 +194,7 @@ the [`bitsandbytes`](https://github.com/bitsandbytes-foundation/bitsandbytes) li
 We can then load models in 8-bit quantization by simply adding a `load_in_8bit=True` flag to `from_pretrained`.
 
 ```python
-model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", quantization_config=BitsAndBytesConfig(load_in_8bit=True), pad_token_id=0)
+model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_8bit=True, pad_token_id=0)
 ```
 
 Now, let's run our example again and measure the memory usage.
@@ -241,7 +241,7 @@ flush()
 Let's see what peak GPU memory consumption 4-bit quantization gives. Quantizing the model to 4-bit can be done with the same API as before - this time by passing `load_in_4bit=True` instead of `load_in_8bit=True`.
 
 ```python
-model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", quantization_config=BitsAndBytesConfig(load_in_4bit=True), pad_token_id=0)
+model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_4bit=True, pad_token_id=0)
 
 pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
 
@@ -271,7 +271,7 @@ Just 9.5GB! That's really not a lot for a >15 billion parameter model.
 
 While we see very little degradation in accuracy for our model here, 4-bit quantization can in practice often lead to different results compared to 8-bit quantization or full `bfloat16` inference. It is up to the user to try it out.
 
-Also note that inference here was again a bit slower compared to 8-bit quantization which is due to the more aggressive quantization method used for 4-bit quantization leading to $\text{quantize}$ and $\text{dequantize}$ taking longer during inference.
+Also note that inference here was again a bit slower compared to 8-bit quantization which is due to the more aggressive quantization method used for 4-bit quantization leading to \\( \text{quantize} \\) and \\( \text{dequantize} \\) taking longer during inference.
 
 ```python
 del model
@@ -300,117 +300,279 @@ Next, let's look into how we can improve computational and memory efficiency by
 Today's top-performing LLMs share more or less the same fundamental architecture that consists of feed-forward layers, activation layers, layer normalization layers, and most crucially, self-attention layers.
 
 Self-attention layers are central to Large Language Models (LLMs) in that they enable the model to understand the contextual relationships between input tokens.
-However, the peak GPU memory consumption for self-attention layers grows *quadratically* both in compute and memory complexity with number of input tokens (also called *sequence length*) that we denote in the following by $N$ .
+However, the peak GPU memory consumption for self-attention layers grows *quadratically* both in compute and memory complexity with number of input tokens (also called *sequence length*) that we denote in the following by \\( N \\) .
 While this is not really noticeable for shorter input sequences (of up to 1000 input tokens), it becomes a serious problem for longer input sequences (at around 16000 input tokens).
 
-Let's take a closer look. The formula to compute the output $\mathbf{O}$ of a self-attention layer for an input $\mathbf{X}$ of length $N$ is:
+Let's take a closer look. The formula to compute the output \\( \mathbf{O} \\) of a self-attention layer for an input \\( \mathbf{X} \\) of length \\( N \\) is:
 
 $$ \textbf{O} = \text{Attn}(\mathbf{X}) = \mathbf{V} \times \text{Softmax}(\mathbf{QK}^T) \text{ with } \mathbf{Q} = \mathbf{W}_q \mathbf{X}, \mathbf{V} = \mathbf{W}_v \mathbf{X}, \mathbf{K} = \mathbf{W}_k \mathbf{X} $$
 
-$\mathbf{X} = (\mathbf{x}_1, ... \mathbf{x}_{N})$ is thereby the input sequence to the attention layer. The projections $\mathbf{Q}$ and $\mathbf{K}$ will each consist of $N$ vectors resulting in the $\mathbf{QK}^T$ being of size $N^2$ .
+\\(  \mathbf{X} = (\mathbf{x}_1, ... \mathbf{x}_{N}) \\) is thereby the input sequence to the attention layer. The projections \\( \mathbf{Q} \\) and \\( \mathbf{K} \\) will each consist of \\( N \\) vectors resulting in the \\( \mathbf{QK}^T \\) being of size \\( N^2 \\) .
 
 LLMs usually have multiple attention heads, thus doing multiple self-attention computations in parallel.
-Assuming, the LLM has 40 attention heads and runs in bfloat16 precision, we can calculate the memory requirement to store the $\mathbf{QK^T}$ matrices to be $40 * 2 * N^2$ bytes. For $N=1000$ only around 50 MB of VRAM are needed, however, for $N=16000$ we would need 19 GB of VRAM, and for $N=100,000$ we would need almost 1TB just to store the $\mathbf{QK}^T$ matrices.
+Assuming, the LLM has 40 attention heads and runs in bfloat16 precision, we can calculate the memory requirement to store the \\( \mathbf{QK^T} \\) matrices to be \\( 40 * 2 * N^2 \\) bytes. For \\( N=1000 \\) only around 50 MB of VRAM are needed, however, for \\( N=16000 \\) we would need 19 GB of VRAM, and for \\( N=100,000 \\) we would need almost 1TB just to store the \\( \mathbf{QK}^T \\) matrices.
 
 Long story short, the default self-attention algorithm quickly becomes prohibitively memory-expensive for large input contexts.
 
 As LLMs improve in text comprehension and generation, they are applied to increasingly complex tasks. While models once handled the translation or summarization of a few sentences, they now manage entire pages, demanding the capability to process extensive input lengths.
 
-How can we get rid of the exorbitant memory requirements for large input lengths? We need a new way to compute the self-attention mechanism that gets rid of the $\mathbf{QK}^T$ matrix. [Tri Dao et al.](https://huggingface.co/papers/2205.14135) developed exactly such a new algorithm and called it **Flash Attention**.
+How can we get rid of the exorbitant memory requirements for large input lengths? We need a new way to compute the self-attention mechanism that gets rid of the \\( QK^T \\) matrix. [Tri Dao et al.](https://huggingface.co/papers/2205.14135) developed exactly such a new algorithm and called it **Flash Attention**.
 
-In a nutshell, Flash Attention breaks the $\mathbf{V} \times \text{Softmax}(\mathbf{QK}^T)$ computation apart and instead computes smaller chunks of the output by iterating over multiple softmax computation steps:
+In a nutshell, Flash Attention breaks the  \\(\mathbf{V} \times \text{Softmax}(\mathbf{QK}^T\\)) computation apart and instead computes smaller chunks of the output by iterating over multiple softmax computation steps:
 
 $$ \textbf{O}_i \leftarrow s^a_{ij} * \textbf{O}_i + s^b_{ij} * \mathbf{V}_{j} \times \text{Softmax}(\mathbf{QK}^T_{i,j}) \text{ for multiple } i, j \text{ iterations} $$
 
-with $s^a_{ij}$ and $s^b_{ij}$ being some softmax normalization statistics that need to be recomputed for every $i$ and $j$ .
+with \\( s^a_{ij} \\) and \\( s^b_{ij} \\) being some softmax normalization statistics that need to be recomputed for every \\( i \\) and \\( j \\) .
 
 Please note that the whole Flash Attention is a bit more complex and is greatly simplified here as going in too much depth is out of scope for this guide. The reader is invited to take a look at the well-written [Flash Attention paper](https://huggingface.co/papers/2205.14135) for more details.
 
 The main takeaway here is:
 
-> By keeping track of softmax normalization statistics and by using some smart mathematics, Flash Attention gives **numerical identical** outputs compared to the default self-attention layer at a memory cost that only increases linearly with $N$ .
+> By keeping track of softmax normalization statistics and by using some smart mathematics, Flash Attention gives **numerical identical** outputs compared to the default self-attention layer at a memory cost that only increases linearly with \\( N \\) .
 
 Looking at the formula, one would intuitively say that Flash Attention must be much slower compared to the default self-attention formula as more computation needs to be done. Indeed Flash Attention requires more FLOPs compared to normal attention as the softmax normalization statistics have to constantly be recomputed (see [paper](https://huggingface.co/papers/2205.14135) for more details if interested)
 
 > However, Flash Attention is much faster in inference compared to default attention which comes from its ability to significantly reduce the demands on the slower, high-bandwidth memory of the GPU (VRAM), focusing instead on the faster on-chip memory (SRAM).
 
-Essentially, Flash Attention makes sure that all intermediate write and read operations can be done using the fast *on-chip* SRAM memory instead of having to access the slower VRAM memory to compute the output vector $\mathbf{O}$ .
+Essentially, Flash Attention makes sure that all intermediate write and read operations can be done using the fast *on-chip* SRAM memory instead of having to access the slower VRAM memory to compute the output vector \\( \mathbf{O} \\) .
 
 In practice, there is currently absolutely no reason to **not** use Flash Attention if available. The algorithm gives mathematically the same outputs, and is both faster and more memory-efficient.
 
+Let's look at a practical example.
+
+Our OctoCoder model now gets a significantly longer input prompt which includes a so-called *system prompt*. System prompts are used to steer the LLM into a better assistant that is tailored to the users' task.
+In the following, we use a system prompt that will make OctoCoder a better coding assistant.
+
+```python
+system_prompt = """Below are a series of dialogues between various people and an AI technical assistant.
+The assistant tries to be helpful, polite, honest, sophisticated, emotionally aware, and humble but knowledgeable.
+The assistant is happy to help with code questions and will do their best to understand exactly what is needed.
+It also tries to avoid giving false or misleading information, and it caveats when it isn't entirely sure about the right answer.
+That said, the assistant is practical really does its best, and doesn't let caution get too much in the way of being useful.
+
+The Starcoder models are a series of 15.5B parameter models trained on 80+ programming languages from The Stack (v1.2) (excluding opt-out requests).
+The model uses Multi Query Attention, was trained using the Fill-in-the-Middle objective, and with 8,192 tokens context window for a trillion tokens of heavily deduplicated data.
+
+-----
+
+Question: Write a function that takes two lists and returns a list that has alternating elements from each input list.
+
+Answer: Sure. Here is a function that does that.
+
+def alternating(list1, list2):
+   results = []
+   for i in range(len(list1)):
+       results.append(list1[i])
+       results.append(list2[i])
+   return results
+
+Question: Can you write some test cases for this function?
+
+Answer: Sure, here are some tests.
+
+assert alternating([10, 20, 30], [1, 2, 3]) == [10, 1, 20, 2, 30, 3]
+assert alternating([True, False], [4, 5]) == [True, 4, False, 5]
+assert alternating([], []) == []
+
+Question: Modify the function so that it returns all input elements when the lists have uneven length. The elements from the longer list should be at the end.
+
+Answer: Here is the modified function.
+
+def alternating(list1, list2):
+   results = []
+   for i in range(min(len(list1), len(list2))):
+       results.append(list1[i])
+       results.append(list2[i])
+   if len(list1) > len(list2):
+       results.extend(list1[i+1:])
+   else:
+       results.extend(list2[i+1:])
+   return results
+
+-----
+"""
+```
+
+For demonstration purposes, we duplicate the system prompt by ten so that the input length is long enough to observe Flash Attention's memory savings.
+We append the original text prompt `"Question: Please write a function in Python that transforms bytes to Giga bytes.\n\nAnswer: Here"`
+
+```python
+long_prompt = 10 * system_prompt + prompt
+```
+
+We instantiate our model again in bfloat16 precision.
+
+```python
+model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", dtype=torch.bfloat16, device_map="auto")
+tokenizer = AutoTokenizer.from_pretrained("bigcode/octocoder")
+
+pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
+```
+
+Let's now run the model just like before *without Flash Attention* and measure the peak GPU memory requirement and inference time.
+
+```python
+import time
+
+start_time = time.time()
+result = pipe(long_prompt, max_new_tokens=60)[0]["generated_text"][len(long_prompt):]
+
+print(f"Generated in {time.time() - start_time} seconds.")
+result
+```
+
+**Output**:
+
+```text
+Generated in 10.96854019165039 seconds.
+Sure. Here is a function that does that.\n\ndef bytes_to_giga(bytes):\n   return bytes / 1024 / 1024 / 1024\n\nAnswer: Sure. Here is a function that does that.\n\ndef
+````
+
+We're getting the same output as before, however this time, the model repeats the answer multiple times until it's 60 tokens cut-off. This is not surprising as we've repeated the system prompt ten times for demonstration purposes and thus cued the model to repeat itself.
+
+**Note** that the system prompt should not be repeated ten times in real-world applications - one time is enough!
+
+Let's measure the peak GPU memory requirement.
+
+```python
+bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
+```
+
+**Output**:
+
+```text
+37.668193340301514
+```
+
+As we can see the peak GPU memory requirement is now significantly higher than in the beginning, which is largely due to the longer input sequence. Also the generation takes a little over a minute now.
+
+We call `flush()` to free GPU memory for our next experiment.
+
+```python
+flush()
+```
+
+For comparison, let's run the same function, but enable Flash Attention instead.
+To do so, we convert the model to [BetterTransformer](https://huggingface.co/docs/optimum/bettertransformer/overview) and by doing so enabling PyTorch's [SDPA self-attention](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention) which in turn is able to use Flash Attention.
+
+```python
+model.to_bettertransformer()
+```
+
+Now we run the exact same code snippet as before and under the hood Transformers will make use of Flash Attention.
+
+```py
+start_time = time.time()
+with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
+    result = pipe(long_prompt, max_new_tokens=60)[0]["generated_text"][len(long_prompt):]
+
+print(f"Generated in {time.time() - start_time} seconds.")
+result
+```
+
+**Output**:
+
+```text
+Generated in 3.0211617946624756 seconds.
+ Sure. Here is a function that does that.\n\ndef bytes_to_giga(bytes):\n   return bytes / 1024 / 1024 / 1024\n\nAnswer: Sure. Here is a function that does that.\n\ndef
+```
+
+We're getting the exact same result as before, but can observe a very significant speed-up thanks to Flash Attention.
+
+Let's measure the memory consumption one last time.
+
+```python
+bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
+```
+
+**Output**:
+
+```text
+32.617331981658936
+```
+
+And we're almost back to our original 29GB peak GPU memory from the beginning.
+
+We can observe that we only use roughly 100MB more GPU memory when passing a very long input sequence with Flash Attention compared to passing a short input sequence as done in the beginning.
+
+```py
+flush()
+```
+
+For more information on how to use Flash Attention, please have a look at [this doc page](https://huggingface.co/docs/transformers/en/perf_infer_gpu_one#flashattention-2).
+
 ## 3. Architectural Innovations
 
 So far we have looked into improving computational and memory efficiency by:
 
-- Casting the weights to a lower precision format
-- Replacing the self-attention algorithm with a more memory- and compute efficient version
+-   Casting the weights to a lower precision format
+-   Replacing the self-attention algorithm with a more memory- and compute efficient version
 
-Let's now look into how we can change the architecture of an LLM so that it is most effective and efficient for tasks that require long text inputs, *e.g.*:
-
-- Retrieval augmented Questions Answering,
-- Summarization,
-- Chat
+Let's now look into how we can change the architecture of an LLM so that it is most effective and efficient for task that require long text inputs, *e.g.*:
+-   Retrieval augmented Questions Answering,
+-   Summarization,
+-   Chat
 
 Note that *chat* not only requires the LLM to handle long text inputs, but it also necessitates that the LLM is able to efficiently handle the back-and-forth dialogue between user and assistant (such as ChatGPT).
 
 Once trained, the fundamental LLM architecture is difficult to change, so it is important to make considerations about the LLM's tasks beforehand and accordingly optimize the model's architecture.
 There are two important components of the model architecture that quickly become memory and/or performance bottlenecks for large input sequences.
 
-- The positional embeddings
-- The key-value cache
+-   The positional embeddings
+-   The key-value cache
 
 Let's go over each component in more detail
 
 ### 3.1 Improving positional embeddings of LLMs
 
 Self-attention puts each token in relation to each other's tokens.
-As an example, the $\text{Softmax}(\mathbf{QK}^T)$ matrix of the text input sequence *"Hello", "I", "love", "you"* could look as follows:
+As an example, the \\( \text{Softmax}(\mathbf{QK}^T) \\) matrix of the text input sequence *"Hello", "I", "love", "you"* could look as follows:
 
 ![](/blog/assets/163_optimize_llm/self_attn_tokens.png)
 
 Each word token is given a probability mass at which it attends all other word tokens and, therefore is put into relation with all other word tokens. E.g. the word *"love"* attends to the word *"Hello"* with 5%, to *"I"* with 30%, and to itself with 65%.
 
 A LLM based on self-attention, but without position embeddings would have great difficulties in understanding the positions of the text inputs to each other.
-This is because the probability score computed by $\mathbf{QK}^T$ relates each word token to each other word token in $O(1)$ computations regardless of their relative positional distance to each other.
+This is because the probability score computed by \\( \mathbf{QK}^T \\) relates each word token to each other word token in \\( O(1) \\) computations regardless of their relative positional distance to each other.
 Therefore, for the LLM without position embeddings each token appears to have the same distance to all other tokens, *e.g.* differentiating between *"Hello I love you"* and *"You love I hello"* would be very challenging.
 
 For the LLM to understand sentence order, an additional *cue* is needed and is usually applied in the form of *positional encodings* (or also called *positional embeddings*).
 Positional encodings, encode the position of each token into a numerical presentation that the LLM can leverage to better understand sentence order.
 
-The authors of the [*Attention Is All You Need*](https://huggingface.co/papers/1706.03762) paper introduced sinusoidal positional embeddings $\mathbf{P} = \mathbf{p}_1, \ldots, \mathbf{p}_N$ .
-where each vector $\mathbf{p}_i$ is computed as a sinusoidal function of its position $i$ .
-The positional encodings are then simply added to the input sequence vectors $\mathbf{\hat{X}} = \mathbf{\hat{x}}_1, \ldots, \mathbf{\hat{x}}_N$ = $\mathbf{x}_1 + \mathbf{p}_1, \ldots, \mathbf{x}_N + \mathbf{p}_N$ thereby cueing the model to better learn sentence order.
+The authors of the [*Attention Is All You Need*](https://huggingface.co/papers/1706.03762) paper introduced sinusoidal positional embeddings \\( \mathbf{P} = \mathbf{p}_1, \ldots, \mathbf{p}_N \\) .
+where each vector \\( \mathbf{p}_i \\) is computed as a sinusoidal function of its position \\( i \\) .
+The positional encodings are then simply added to the input sequence vectors \\( \mathbf{\hat{X}} = \mathbf{\hat{x}}_1, \ldots, \mathbf{\hat{x}}_N \\) = \\( \mathbf{x}_1 + \mathbf{p}_1, \ldots, \mathbf{x}_N + \mathbf{p}_N \\) thereby cueing the model to better learn sentence order.
 
 Instead of using fixed position embeddings, others (such as [Devlin et al.](https://huggingface.co/papers/1810.04805)) used learned positional encodings for which the positional embeddings
-$\mathbf{P}$ are learned during training.
+\\( \mathbf{P} \\) are learned during training.
 
 Sinusoidal and learned position embeddings used to be the predominant methods to encode sentence order into LLMs, but a couple of problems related to these positional encodings were found:
 
-  1. Sinusoidal and learned position embeddings are both absolute positional embeddings, *i.e.* encoding a unique embedding for each position id: $0, \ldots, N$ . As shown by [Huang et al.](https://huggingface.co/papers/2009.13658) and [Su et al.](https://huggingface.co/papers/2104.09864), absolute positional embeddings lead to poor LLM performance for long text inputs. For long text inputs, it is advantageous if the model learns the relative positional distance input tokens have to each other instead of their absolute position.
-  2. When using learned position embeddings, the LLM has to be trained on a fixed input length $N$, which makes it difficult to extrapolate to an input length longer than what it was trained on.
+  1. Sinusoidal and learned position embeddings are both absolute positional embeddings, *i.e.* encoding a unique embedding for each position id: \\( 0, \ldots, N \\) . As shown by [Huang et al.](https://huggingface.co/papers/2009.13658) and [Su et al.](https://huggingface.co/papers/2104.09864), absolute positional embeddings lead to poor LLM performance for long text inputs. For long text inputs, it is advantageous if the model learns the relative positional distance input tokens have to each other instead of their absolute position.
+  2. When using learned position embeddings, the LLM has to be trained on a fixed input length \\( N \\), which makes it difficult to extrapolate to an input length longer than what it was trained on.
 
 Recently, relative positional embeddings that can tackle the above mentioned problems have become more popular, most notably:
 
-- [Rotary Position Embedding (RoPE)](https://huggingface.co/papers/2104.09864)
-- [ALiBi](https://huggingface.co/papers/2108.12409)
+-   [Rotary Position Embedding (RoPE)](https://huggingface.co/papers/2104.09864)
+-   [ALiBi](https://huggingface.co/papers/2108.12409)
 
-Both *RoPE* and *ALiBi* argue that it's best to cue the LLM about sentence order directly in the self-attention algorithm as it's there that word tokens are put into relation with each other. More specifically, sentence order should be cued by modifying the $\mathbf{QK}^T$ computation.
+Both *RoPE* and *ALiBi* argue that it's best to cue the LLM about sentence order directly in the self-attention algorithm as it's there that word tokens are put into relation with each other. More specifically, sentence order should be cued by modifying the \\( \mathbf{QK}^T \\) computation.
 
-Without going into too many details, *RoPE* notes that positional information can be encoded into query-key pairs, *e.g.* $\mathbf{q}_i$ and $\mathbf{x}_j$ by rotating each vector by an angle $\theta * i$ and $\theta * j$ respectively with $i, j$ describing each vectors sentence position:
+Without going into too many details, *RoPE* notes that positional information can be encoded into query-key pairs, *e.g.* \\( \mathbf{q}_i \\) and \\( \mathbf{x}_j \\) by rotating each vector by an angle \\( \theta * i \\) and \\( \theta * j \\) respectively with \\( i, j \\) describing each vectors sentence position:
 
 $$ \mathbf{\hat{q}}_i^T \mathbf{\hat{x}}_j = \mathbf{{q}}_i^T \mathbf{R}_{\theta, i -j} \mathbf{{x}}_j. $$
 
-$\mathbf{R}_{\theta, i - j}$ thereby represents a rotational matrix. $\theta$ is *not* learned during training, but instead set to a pre-defined value that depends on the maximum input sequence length during training.
+\\( \mathbf{R}_{\theta, i - j} \\) thereby represents a rotational matrix. \\( \theta \\) is *not* learned during training, but instead set to a pre-defined value that depends on the maximum input sequence length during training.
 
-> By doing so, the probability score between $\mathbf{q}_i$ and $\mathbf{q}_j$ is only affected if $i \ne j$ and solely depends on the relative distance $i - j$ regardless of each vector's specific positions $i$ and $j$ .
+> By doing so, the probability score between \\( \mathbf{q}_i \\) and \\( \mathbf{q}_j \\) is only affected if \\( i \ne j \\) and solely depends on the relative distance \\( i - j \\) regardless of each vector's specific positions \\( i \\) and \\( j \\) .
 
 *RoPE* is used in multiple of today's most important LLMs, such as:
 
-- [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
-- [**Llama**](https://huggingface.co/papers/2302.13971)
-- [**PaLM**](https://huggingface.co/papers/2204.02311)
+-   [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
+-   [**Llama**](https://huggingface.co/papers/2302.13971)
+-   [**PaLM**](https://huggingface.co/papers/2204.02311)
 
-As an alternative, *ALiBi* proposes a much simpler relative position encoding scheme. The relative distance that input tokens have to each other is added as a negative integer scaled by a pre-defined value `m` to each query-key entry of the $\mathbf{QK}^T$ matrix right before the softmax computation.
+As an alternative, *ALiBi* proposes a much simpler relative position encoding scheme. The relative distance that input tokens have to each other is added as a negative integer scaled by a pre-defined value `m` to each query-key entry of the \\( \mathbf{QK}^T \\) matrix right before the softmax computation.
 
 ![](/blog/assets/163_optimize_llm/alibi.png)
 
@@ -418,20 +580,19 @@ As shown in the [ALiBi](https://huggingface.co/papers/2108.12409) paper, this si
 
 *ALiBi* is used in multiple of today's most important LLMs, such as:
 
-- [**MPT**](https://huggingface.co/mosaicml/mpt-30b)
-- [**BLOOM**](https://huggingface.co/bigscience/bloom)
+-   [**MPT**](https://huggingface.co/mosaicml/mpt-30b)
+-   [**BLOOM**](https://huggingface.co/bigscience/bloom)
 
 Both *RoPE* and *ALiBi* position encodings can extrapolate to input lengths not seen during training whereas it has been shown that extrapolation works much better out-of-the-box for *ALiBi* as compared to *RoPE*.
 For ALiBi, one simply increases the values of the lower triangular position matrix to match the length of the input sequence.
-For *RoPE*, keeping the same $\theta$ that was used during training leads to poor results when passing text inputs much longer than those seen during training, *c.f* [Press et al.](https://huggingface.co/papers/2108.12409). However, the community has found a couple of effective tricks that adapt $\theta$, thereby allowing *RoPE* position embeddings to work well for extrapolated text input sequences (see [here](https://github.com/huggingface/transformers/pull/24653)).
+For *RoPE*, keeping the same \\( \theta \\) that was used during training leads to poor results when passing text inputs much longer than those seen during training, *c.f* [Press et al.](https://huggingface.co/papers/2108.12409). However, the community has found a couple of effective tricks that adapt \\( \theta \\), thereby allowing *RoPE* position embeddings to work well for extrapolated text input sequences (see [here](https://github.com/huggingface/transformers/pull/24653)).
 
 > Both RoPE and ALiBi are relative positional embeddings that are *not* learned during training, but instead are based on the following intuitions:
+ -   Positional cues about the text inputs should be given directly to the \\( QK^T \\) matrix of the self-attention layer
+ -   The LLM should be incentivized to learn a constant *relative* distance positional encodings have to each other
+ -   The further text input tokens are from each other, the lower the probability of their query-value probability. Both RoPE and ALiBi lower the query-key probability of tokens far away from each other. RoPE by decreasing their vector product by increasing the angle between the query-key vectors. ALiBi by adding large negative numbers to the vector product
 
-- Positional cues about the text inputs should be given directly to the $\mathbf{QK}^T$ matrix of the self-attention layer.
-- The LLM should be incentivized to learn a constant *relative* distance positional encoding.
-- The further text input tokens are from each other, the lower the probability of their query-value probability. Both RoPE and ALiBi lower the query-key probability of tokens far away from each other. RoPE lowers by decreasing their vector product by increasing the angle between the query-key vectors. ALiBi lowers by adding large negative numbers to the vector product.
-
-In conclusion, LLMs that are intended to be deployed in tasks that require handling large text inputs are better trained with relative positional embeddings, such as RoPE and ALiBi. Also note that even if an LLM with RoPE and ALiBi has been trained only on a fixed length of say $N_1 = 2048$ it can still be used in practice with text inputs much larger than $N_1$, like $N_2 = 8192 > N_1$ by extrapolating the positional embeddings.
+In conclusion, LLMs that are intended to be deployed in tasks that require handling large text inputs are better trained with relative positional embeddings, such as RoPE and ALiBi. Also note that even if an LLM with RoPE and ALiBi has been trained only on a fixed length of say \\( N_1 = 2048 \\) it can still be used in practice with text inputs much larger than \\( N_1 \\), like \\( N_2 = 8192 > N_1 \\) by extrapolating the positional embeddings.
 
 ### 3.2 The key-value cache
 
@@ -470,7 +631,7 @@ As we can see every time we increase the text input tokens by the just sampled t
 
 With very few exceptions, LLMs are trained using the [causal language modeling objective](https://huggingface.co/docs/transformers/tasks/language_modeling#causal-language-modeling) and therefore mask the upper triangle matrix of the attention score - this is why in the two diagrams above the attention scores are left blank (*a.k.a* have 0 probability). For a quick recap on causal language modeling you can refer to the [*Illustrated Self Attention blog*](https://jalammar.github.io/illustrated-gpt2/#part-2-illustrated-self-attention).
 
-As a consequence, tokens *never* depend on later tokens, more specifically the $\mathbf{q}_i$ vector is never put in relation with any key, values vectors $\mathbf{k}_j, \mathbf{v}_j$ if $j > i$ . Instead $\mathbf{q}_i$ only attends to previous key-value vectors $\mathbf{k}_{m < i}, \mathbf{v}_{m < i} \text{ , for } m \in \{0, \ldots i - 1\}$. In order to reduce unnecessary computation, one can therefore cache each layer's key-value vectors for all previous timesteps.
+As a consequence, tokens *never* depend on previous tokens, more specifically the \\( \mathbf{q}_i \\) vector is never put in relation with any key, values vectors \\( \mathbf{k}_j, \mathbf{v}_j \\) if \\( j > i \\) . Instead \\( \mathbf{q}_i \\) only attends to previous key-value vectors \\( \mathbf{k}_{m < i}, \mathbf{v}_{m < i} \text{ , for } m \in \{0, \ldots i - 1\} \\). In order to reduce unnecessary computation, one can therefore cache each layer's key-value vectors for all previous timesteps.
 
 In the following, we will tell the LLM to make use of the key-value cache by retrieving and forwarding it for each forward pass.
 In Transformers, we can retrieve the key-value cache by passing the `use_cache` flag to the `forward` call and can then pass it with the current token.
@@ -486,7 +647,7 @@ for _ in range(5):
   next_token_id = torch.argmax(next_logits, dim=-1)
 
   print("shape of input_ids", next_token_id.shape)
-  print("length of key-value cache", past_key_values.get_seq_length())  # past_key_values are of shape [num_layers, 0 for k, 1 for v, batch_size, length, hidden_dim]
+  print("length of key-value cache", len(past_key_values[0][0]))  # past_key_values are of shape [num_layers, 0 for k, 1 for v, batch_size, length, hidden_dim]
   generated_tokens.append(next_token_id.item())
 
 generated_text = tokenizer.batch_decode(generated_tokens)
@@ -511,12 +672,11 @@ length of key-value cache 24
 
 As one can see, when using the key-value cache the text input tokens are *not* increased in length, but remain a single input vector. The length of the key-value cache on the other hand is increased by one at every decoding step.
 
-> Making use of the key-value cache means that the $\mathbf{QK}^T$ is essentially reduced to $\mathbf{q}_c\mathbf{K}^T$ with $\mathbf{q}_c$ being the query projection of the currently passed input token which is *always* just a single vector.
+> Making use of the key-value cache means that the \\( \mathbf{QK}^T \\) is essentially reduced to \\( \mathbf{q}_c\mathbf{K}^T \\) with \\( \mathbf{q}_c \\) being the query projection of the currently passed input token which is *always* just a single vector.
 
 Using the key-value cache has two advantages:
-
-- Significant increase in computational efficiency as less computations are performed compared to computing the full $\mathbf{QK}^T$ matrix. This leads to an increase in inference speed
-- The maximum required memory is not increased quadratically with the number of generated tokens, but only increases linearly.
+-   Significant increase in computational efficiency as less computations are performed compared to computing the full \\( \mathbf{QK}^T \\) matrix. This leads to an increase in inference speed
+-   The maximum required memory is not increased quadratically with the number of generated tokens, but only increases linearly.
 
 > One should *always* make use of the key-value cache as it leads to identical results and a significant speed-up for longer input sequences. Transformers has the key-value cache enabled by default when making use of the text pipeline or the [`generate` method](https://huggingface.co/docs/transformers/main_classes/text_generation). We have an entire guide dedicated to caches [here](./kv_cache).
 
@@ -538,12 +698,10 @@ Assistant: Germany has ca. 81 million inhabitants
 ```
 
 In this chat, the LLM runs auto-regressive decoding twice:
-
   1. The first time, the key-value cache is empty and the input prompt is `"User: How many people live in France?"` and the model auto-regressively generates the text `"Roughly 75 million people live in France"` while increasing the key-value cache at every decoding step.
   2. The second time the input prompt is `"User: How many people live in France? \n Assistant: Roughly 75 million people live in France \n User: And how many in Germany?"`. Thanks to the cache, all key-value vectors for the first two sentences are already computed. Therefore the input prompt only consists of `"User: And how many in Germany?"`. While processing the shortened input prompt, its computed key-value vectors are concatenated to the key-value cache of the first decoding. The second Assistant's answer `"Germany has ca. 81 million inhabitants"` is then auto-regressively generated with the key-value cache consisting of encoded key-value vectors of `"User: How many people live in France? \n Assistant: Roughly 75 million people live in France \n User: And how many are in Germany?"`.
 
 Two things should be noted here:
-
   1. Keeping all the context is crucial for LLMs deployed in chat so that the LLM understands all the previous context of the conversation. E.g. for the example above the LLM needs to understand that the user refers to the population when asking `"And how many are in Germany"`.
   2. The key-value cache is extremely useful for chat as it allows us to continuously grow the encoded chat history instead of having to re-encode the chat history again from scratch (as e.g. would be the case when using an encoder-decoder architecture).
 
@@ -579,7 +737,7 @@ def bytes_to_megabytes(bytes):
 Answer: The function takes a number of bytes as input and returns the number of
 ```
 
-Great, no additional time is spent recomputing the same key and values for the attention layer! There is however one catch. While the required peak memory for the $\mathbf{QK}^T$ matrix is significantly reduced, holding the key-value cache in memory can become very memory expensive for long input sequences or multi-turn chat. Remember that the key-value cache needs to store the key-value vectors for all previous input vectors $\mathbf{x}_i \text{, for } i \in \{1, \ldots, c - 1\}$ for all self-attention layers and for all attention heads.
+Great, no additional time is spent recomputing the same key and values for the attention layer! There is however one catch. While the required peak memory for the \\( \mathbf{QK}^T \\) matrix is significantly reduced, holding the key-value cache in memory can become very memory expensive for long input sequences or multi-turn chat. Remember that the key-value cache needs to store the key-value vectors for all previous input vectors \\( \mathbf{x}_i \text{, for } i \in \{1, \ldots, c - 1\} \\) for all self-attention layers and for all attention heads.
 
 Let's compute the number of float values that need to be stored in the key-value cache for the LLM `bigcode/octocoder` that we used before.
 The number of float values amounts to two times the sequence length times the number of attention heads times the attention head dimension and times the number of layers.
@@ -603,21 +761,21 @@ Researchers have proposed two methods that allow to significantly reduce the mem
 
 [Multi-Query-Attention](https://huggingface.co/papers/1911.02150) was proposed in Noam Shazeer's *Fast Transformer Decoding: One Write-Head is All You Need* paper. As the title says, Noam found out that instead of using `n_head` key-value projections weights, one can use a single head-value projection weight pair that is shared across all attention heads without that the model's performance significantly degrades.
 
-> By using a single head-value projection weight pair, the key value vectors $\mathbf{k}_i, \mathbf{v}_i$ have to be identical across all attention heads which in turn means that we only need to store 1 key-value projection pair in the cache instead of `n_head` ones.
+> By using a single head-value projection weight pair, the key value vectors \\( \mathbf{k}_i, \mathbf{v}_i \\) have to be identical across all attention heads which in turn means that we only need to store 1 key-value projection pair in the cache instead of `n_head` ones.
 
 As most LLMs use between 20 and 100 attention heads, MQA significantly reduces the memory consumption of the key-value cache. For the LLM used in this notebook we could therefore reduce the required memory consumption from 15 GB to less than 400 MB at an input sequence length of 16000.
 
 In addition to memory savings, MQA also leads to improved computational efficiency as explained in the following.
-In auto-regressive decoding, large key-value vectors need to be reloaded, concatenated with the current key-value vector pair to be then fed into the $\mathbf{q}_c\mathbf{K}^T$ computation at every step. For auto-regressive decoding, the required memory bandwidth for the constant reloading can become a serious time bottleneck. By reducing the size of the key-value vectors less memory needs to be accessed, thus reducing the memory bandwidth bottleneck. For more detail, please have a look at [Noam's paper](https://huggingface.co/papers/1911.02150).
+In auto-regressive decoding, large key-value vectors need to be reloaded, concatenated with the current key-value vector pair to be then fed into the \\( \mathbf{q}_c\mathbf{K}^T \\) computation at every step. For auto-regressive decoding, the required memory bandwidth for the constant reloading can become a serious time bottleneck. By reducing the size of the key-value vectors less memory needs to be accessed, thus reducing the memory bandwidth bottleneck. For more detail, please have a look at [Noam's paper](https://huggingface.co/papers/1911.02150).
 
-The important part to understand here is that reducing the number of key-value attention heads to 1 only makes sense if a key-value cache is used. The peak memory consumption of the model for a single forward pass without key-value cache stays unchanged as every attention head still has a unique query vector so that each attention head still has a different $\mathbf{QK}^T$ matrix.
+The important part to understand here is that reducing the number of key-value attention heads to 1 only makes sense if a key-value cache is used. The peak memory consumption of the model for a single forward pass without key-value cache stays unchanged as every attention head still has a unique query vector so that each attention head still has a different \\( \mathbf{QK}^T \\) matrix.
 
 MQA has seen wide adoption by the community and is now used by many of the most popular LLMs:
 
-- [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
-- [**PaLM**](https://huggingface.co/papers/2204.02311)
-- [**MPT**](https://huggingface.co/mosaicml/mpt-30b)
-- [**BLOOM**](https://huggingface.co/bigscience/bloom)
+-   [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
+-   [**PaLM**](https://huggingface.co/papers/2204.02311)
+-   [**MPT**](https://huggingface.co/mosaicml/mpt-30b)
+-   [**BLOOM**](https://huggingface.co/bigscience/bloom)
 
 Also, the checkpoint used in this notebook - `bigcode/octocoder` - makes use of MQA.
 

docs/source/en/main_classes/configuration.md

@@ -16,7 +16,7 @@ rendered properly in your Markdown viewer.
 
 # Configuration
 
-The base class [`PreTrainedConfig`] implements the common methods for loading/saving a configuration
+The base class [`PretrainedConfig`] implements the common methods for loading/saving a configuration
 either from a local file or directory, or from a pretrained model configuration provided by the library (downloaded
 from HuggingFace's AWS S3 repository).
 
@@ -24,8 +24,8 @@ Each derived config class implements model specific attributes. Common attribute
 `hidden_size`, `num_attention_heads`, and `num_hidden_layers`. Text models further implement:
 `vocab_size`.
 
-## PreTrainedConfig
+## PretrainedConfig
 
-[[autodoc]] PreTrainedConfig
+[[autodoc]] PretrainedConfig
     - push_to_hub
     - all

docs/source/en/main_classes/data_collator.md

@@ -67,6 +67,6 @@ Examples of use can be found in the [example scripts](../examples) or [example n
 
 [[autodoc]] data.data_collator.DataCollatorWithFlattening
 
-## DataCollatorForMultipleChoice
+# DataCollatorForMultipleChoice
 
 [[autodoc]] data.data_collator.DataCollatorForMultipleChoice

docs/source/en/main_classes/feature_extractor.md

@@ -16,7 +16,7 @@ rendered properly in your Markdown viewer.
 
 # Feature Extractor
 
-A feature extractor is in charge of preparing input features for audio models. This includes feature extraction from sequences, e.g., pre-processing audio files to generate Log-Mel Spectrogram features, and conversion to NumPy and PyTorch tensors.
+A feature extractor is in charge of preparing input features for audio or vision models. This includes feature extraction from sequences, e.g., pre-processing audio files to generate Log-Mel Spectrogram features, feature extraction from images, e.g., cropping image files, but also padding, normalization, and conversion to NumPy and PyTorch tensors.
 
 ## FeatureExtractionMixin
 

docs/source/en/main_classes/kernels.md

@@ -1,7 +0,0 @@
-## Kernels
-
-This page documents the kernels configuration utilities.
-
-### KernelConfig
-
-[[autodoc]] KernelConfig

docs/source/en/main_classes/logging.md

@@ -80,7 +80,6 @@ We use both in the `transformers` library. We leverage and adapt `logging`'s `ca
 management of these warning messages by the verbosity setters above.
 
 What does that mean for developers of the library? We should respect the following heuristics:
-
 - `warnings` should be favored for developers of the library and libraries dependent on `transformers`
 - `logging` should be used for end-users of the library using it in every-day projects
 

docs/source/en/main_classes/model.md

@@ -22,6 +22,7 @@ file or directory, or from a pretrained model configuration provided by the libr
 [`PreTrainedModel`] also implements a few methods which are common among all the models to:
 
 - resize the input token embeddings when new tokens are added to the vocabulary
+- prune the attention heads of the model.
 
 The other methods that are common to each model are defined in [`~modeling_utils.ModuleUtilsMixin`] and [`~generation.GenerationMixin`].
 
@@ -42,3 +43,7 @@ set this to `False`.
 ## Pushing to the Hub
 
 [[autodoc]] utils.PushToHubMixin
+
+## Sharded checkpoints
+
+[[autodoc]] modeling_utils.load_sharded_checkpoint

docs/source/en/main_classes/onnx.md

@@ -0,0 +1,53 @@
+<!--Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Exporting 🤗 Transformers models to ONNX
+
+🤗 Transformers provides a `transformers.onnx` package that enables you to
+convert model checkpoints to an ONNX graph by leveraging configuration objects.
+
+See the [guide](../serialization) on exporting 🤗 Transformers models for more
+details.
+
+## ONNX Configurations
+
+We provide three abstract classes that you should inherit from, depending on the
+type of model architecture you wish to export:
+
+* Encoder-based models inherit from [`~onnx.config.OnnxConfig`]
+* Decoder-based models inherit from [`~onnx.config.OnnxConfigWithPast`]
+* Encoder-decoder models inherit from [`~onnx.config.OnnxSeq2SeqConfigWithPast`]
+
+### OnnxConfig
+
+[[autodoc]] onnx.config.OnnxConfig
+
+### OnnxConfigWithPast
+
+[[autodoc]] onnx.config.OnnxConfigWithPast
+
+### OnnxSeq2SeqConfigWithPast
+
+[[autodoc]] onnx.config.OnnxSeq2SeqConfigWithPast
+
+## ONNX Features
+
+Each ONNX configuration is associated with a set of _features_ that enable you
+to export models for different types of topologies or tasks.
+
+### FeaturesManager
+
+[[autodoc]] onnx.features.FeaturesManager

docs/source/en/main_classes/pipelines.md

@@ -267,7 +267,6 @@ about how many forward passes you inputs are actually going to trigger, you can
 independently of the inputs. The caveats from the previous section still apply.
 
 ## Pipeline FP16 inference
-
 Models can be run in FP16 which can be significantly faster on GPU while saving memory. Most models will not suffer noticeable performance loss from this. The larger the model, the less likely that it will.
 
 To enable FP16 inference, you can simply pass `dtype=torch.float16` or `dtype='float16'` to the pipeline constructor. Note that this only works for models with a PyTorch backend. Your inputs will be converted to FP16 internally.
@@ -335,7 +334,6 @@ Pipelines available for audio tasks include the following.
 Pipelines available for computer vision tasks include the following.
 
 ### DepthEstimationPipeline
-
 [[autodoc]] DepthEstimationPipeline
     - __call__
     - all

docs/source/en/main_classes/processors.md

@@ -17,7 +17,6 @@ rendered properly in your Markdown viewer.
 # Processors
 
 Processors can mean two different things in the Transformers library:
-
 - the objects that pre-process inputs for multi-modal models such as [Wav2Vec2](../model_doc/wav2vec2) (speech and text)
   or [CLIP](../model_doc/clip) (text and vision)
 - deprecated objects that were used in older versions of the library to preprocess data for GLUE or SQUAD.

docs/source/en/main_classes/quantization.md

@@ -43,7 +43,6 @@ Learn how to quantize models in the [Quantization](../quantization) guide.
 [[autodoc]] AwqConfig
 
 ## EetqConfig
-
 [[autodoc]] EetqConfig
 
 ## GPTQConfig

docs/source/en/main_classes/text_generation.md

@@ -30,15 +30,15 @@ like token streaming.
 ## GenerationConfig
 
 [[autodoc]] generation.GenerationConfig
-    - from_pretrained
-    - from_model_config
-    - save_pretrained
-    - update
-    - validate
-    - get_generation_mode
+	- from_pretrained
+	- from_model_config
+	- save_pretrained
+	- update
+	- validate
+	- get_generation_mode
 
 ## GenerationMixin
 
 [[autodoc]] GenerationMixin
-    - generate
-    - compute_transition_scores
+	- generate
+	- compute_transition_scores

docs/source/en/main_classes/tokenizer.md

@@ -50,14 +50,14 @@ several advanced alignment methods which can be used to map between the original
 token space (e.g., getting the index of the token comprising a given character or the span of characters corresponding
 to a given token).
 
-## Multimodal Tokenizer
+# Multimodal Tokenizer
 
 Apart from that each tokenizer can be a "multimodal" tokenizer which means that the tokenizer will hold all relevant special tokens
 as part of tokenizer attributes for easier access. For example, if the tokenizer is loaded from a vision-language model like LLaVA, you will
 be able to access `tokenizer.image_token_id` to obtain the special image token used as a placeholder.
 
 To enable extra special tokens for any type of tokenizer, you have to add the following lines and save the tokenizer. Extra special tokens do not
-have to be modality related and can be anything that the model often needs access to. In the below code, tokenizer at `output_dir` will have direct access
+have to be modality related and can ne anything that the model often needs access to. In the below code, tokenizer at `output_dir` will have direct access
 to three more special tokens.  
 
 ```python