Merge branch 'huggingface:main' into videollama3

Author: steven-ccq

.github/workflows/model_jobs.yml

@@ -12,9 +12,6 @@ on:
       slice_id:
         required: true
         type: number
-      runner_map:
-        required: false
-        type: string
       docker:
         required: true
         type: string
@@ -54,10 +51,12 @@ jobs:
       matrix:
         folders: ${{ fromJson(inputs.folder_slices)[inputs.slice_id] }}
     runs-on:
-      group: ${{ fromJson(inputs.runner_map)[matrix.folders][inputs.machine_type] }}
+      group: '${{ inputs.machine_type }}'
     container:
       image: ${{ inputs.docker }}
       options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
+    outputs:
+      machine_type: ${{ steps.set_machine_type.outputs.machine_type }}
     steps:
       - name: Echo input and matrix info
         shell: bash
@@ -111,6 +110,7 @@ jobs:
         run: pip freeze
 
       - name: Set `machine_type` for report and artifact names
+        id: set_machine_type
         working-directory: /transformers
         shell: bash
         run: |
@@ -126,6 +126,7 @@ jobs:
 
           echo "$machine_type"
           echo "machine_type=$machine_type" >> $GITHUB_ENV
+          echo "machine_type=$machine_type" >> $GITHUB_OUTPUT
 
       - name: Run all tests on GPU
         working-directory: /transformers
@@ -159,5 +160,5 @@ jobs:
       job: run_models_gpu
       report_repo_id: ${{ inputs.report_repo_id }}
       gpu_name: ${{ inputs.runner_type }}
-      machine_type: ${{ inputs.machine_type }}
+      machine_type: ${{ needs.run_models_gpu.outputs.machine_type }}
     secrets: inherit

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

@@ -88,6 +88,7 @@ jobs:
       job: run_trainer_and_fsdp_gpu
       slack_report_channel: "#transformers-ci-daily-training"
       docker: huggingface/transformers-all-latest-gpu
+      runner_type: "a10"
       ci_event: Daily CI
       report_repo_id: hf-internal-testing/transformers_daily_ci
       commit_sha: ${{ github.sha }}

.github/workflows/self-scheduled.yml

@@ -68,7 +68,6 @@ jobs:
     outputs:
       folder_slices: ${{ steps.set-matrix.outputs.folder_slices }}
       slice_ids: ${{ steps.set-matrix.outputs.slice_ids }}
-      runner_map: ${{ steps.set-matrix.outputs.runner_map }}
       quantization_matrix: ${{ steps.set-matrix-quantization.outputs.quantization_matrix }}
     steps:
       - name: Update clone
@@ -95,7 +94,6 @@ jobs:
           if [ "${{ inputs.job }}" = "run_models_gpu" ]; then
             echo "folder_slices=$(python3 ../utils/split_model_tests.py --models '${{ inputs.models }}' --num_splits ${{ env.NUM_SLICES }})" >> $GITHUB_OUTPUT
             echo "slice_ids=$(python3 -c 'd = list(range(${{ env.NUM_SLICES }})); print(d)')" >> $GITHUB_OUTPUT
-            echo "runner_map=$(python3 ../utils/get_runner_map.py)" >> $GITHUB_OUTPUT
           elif [ "${{ inputs.job }}" = "run_trainer_and_fsdp_gpu" ]; then
             echo "folder_slices=[['trainer'], ['fsdp']]" >> $GITHUB_OUTPUT
             echo "slice_ids=[0, 1]" >> $GITHUB_OUTPUT
@@ -119,14 +117,13 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        machine_type: [single-gpu, multi-gpu]
+        machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
         slice_id: ${{ fromJSON(needs.setup.outputs.slice_ids) }}
     uses: ./.github/workflows/model_jobs.yml
     with:
       folder_slices: ${{ needs.setup.outputs.folder_slices }}
       machine_type: ${{ matrix.machine_type }}
       slice_id: ${{ matrix.slice_id }}
-      runner_map: ${{ needs.setup.outputs.runner_map }}
       docker: ${{ inputs.docker }}
       commit_sha: ${{ inputs.commit_sha || github.sha }}
       runner_type: ${{ inputs.runner_type }}
@@ -147,9 +144,10 @@ jobs:
       folder_slices: ${{ needs.setup.outputs.folder_slices }}
       machine_type: ${{ matrix.machine_type }}
       slice_id: ${{ matrix.slice_id }}
-      runner_map: ${{ needs.setup.outputs.runner_map }}
       docker: ${{ inputs.docker }}
       commit_sha: ${{ inputs.commit_sha || github.sha }}
+      runner_type: ${{ inputs.runner_type }}
+      report_repo_id: ${{ inputs.report_repo_id }}
       report_name_prefix: run_trainer_and_fsdp_gpu
     secrets: inherit
 

conftest.py

@@ -16,6 +16,7 @@
 # by pytest before any tests are run
 
 import doctest
+import os
 import sys
 import warnings
 from os.path import abspath, dirname, join
@@ -27,6 +28,7 @@
     HfDoctestModule,
     HfDocTestParser,
     is_torch_available,
+    patch_testing_methods_to_collect_info,
     patch_torch_compile_force_graph,
 )
 
@@ -145,3 +147,7 @@ def check_output(self, want, got, optionflags):
     # patch `torch.compile`: if `TORCH_COMPILE_FORCE_FULLGRAPH=1` (or values considered as true, e.g. yes, y, etc.),
     # the patched version will always run with `fullgraph=True`.
     patch_torch_compile_force_graph()
+
+
+if os.environ.get("PATCH_TESTING_METHODS_TO_COLLECT_OUTPUTS", "").lower() in ("yes", "true", "on", "y", "1"):
+    patch_testing_methods_to_collect_info()

docker/transformers-intel-cpu/Dockerfile

@@ -0,0 +1,71 @@
+FROM intel/deep-learning-essentials:2025.1.3-0-devel-ubuntu24.04 AS base
+LABEL maintainer="Hugging Face"
+SHELL ["/bin/bash", "-c"]
+
+ARG PYTHON_VERSION=3.12
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt-get update && \
+    apt-get install -y software-properties-common && \
+    add-apt-repository -y ppa:deadsnakes/ppa && \
+    apt-get update
+
+RUN apt-get update && \
+    apt-get -y install \
+    apt-utils \
+    build-essential \
+    ca-certificates \
+    clinfo \
+    curl \
+    git \
+    git-lfs \
+    vim \
+    numactl \
+    gnupg2 \
+    gpg-agent \
+    python3-dev \
+    python3-opencv \
+    unzip \
+    ffmpeg \
+    tesseract-ocr \
+    espeak-ng \
+    wget \
+    ncurses-term \
+    google-perftools \
+    libjemalloc-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_VERSION} --seed ${VIRTUAL_ENV}
+ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+
+RUN pip install --upgrade pip wheel
+RUN pip install torch torchvision torchaudio torchcodec --index-url https://download.pytorch.org/whl/cpu --no-cache-dir
+RUN pip install av pyctcdecode pytesseract decord galore-torch fire scipy scikit-learn sentencepiece sentence_transformers sacremoses nltk rouge_score librosa soundfile mpi4py pytorch_msssim
+RUN pip install onnx optimum onnxruntime
+RUN pip install autoawq
+RUN pip install gptqmodel --no-build-isolation
+RUN pip install -U datasets timm transformers accelerate peft diffusers opencv-python kenlm evaluate
+RUN pip install -U intel-openmp
+
+# install bitsandbytes
+RUN git clone https://github.com/bitsandbytes-foundation/bitsandbytes.git && cd bitsandbytes/ && \
+    cmake -DCOMPUTE_BACKEND=cpu -S . && make && pip install . && cd ../
+
+# CPU don't need triton
+RUN pip uninstall triton -y
+
+ENV LD_PRELOAD=${LD_PRELOAD}:/opt/venv/lib/libiomp5.so:/usr/lib/x86_64-linux-gnu/libtcmalloc.so.4
+ENV KMP_AFFINITY=granularity=fine,compact,1,0
+
+RUN touch /entrypoint.sh
+RUN chmod +x /entrypoint.sh
+RUN echo "#!/bin/bash" >> /entrypoint.sh
+RUN echo "/bin/bash" >> /entrypoint.sh
+
+ENTRYPOINT ["/entrypoint.sh"]

docs/source/en/_toctree.yml

@@ -199,6 +199,8 @@
     title: HIGGS
   - local: quantization/hqq
     title: HQQ
+  - local: quantization/mxfp4
+    title: MXFP4
   - local: quantization/optimum
     title: Optimum
   - local: quantization/quanto
@@ -557,6 +559,8 @@
         title: Llama2
       - local: model_doc/llama3
         title: Llama3
+      - local: model_doc/longcat_flash
+        title: LongCatFlash
       - local: model_doc/longformer
         title: Longformer
       - local: model_doc/longt5
@@ -625,6 +629,8 @@
         title: OLMo
       - local: model_doc/olmo2
         title: OLMo2
+      - local: model_doc/olmo3
+        title: Olmo3
       - local: model_doc/olmoe
         title: OLMoE
       - local: model_doc/open-llama

docs/source/en/cache_explanation.md

@@ -85,7 +85,7 @@ When you use Transformers' [`Cache`] class, the self-attention module performs s
 
 Caches are structured as a list of layers, where each layer contains a key and value cache. The key and value caches are tensors with the shape `[batch_size, num_heads, seq_len, head_dim]`.
 
-Layers can be of different types (e.g. `DynamicLayer`, `StaticLayer`, `SlidingWindowLayer`), which mostly changes how sequence length is handled and how the cache is updated.
+Layers can be of different types (e.g. `DynamicLayer`, `StaticLayer`, `StaticSlidingWindowLayer`), which mostly changes how sequence length is handled and how the cache is updated.
 
 The simplest is a `DynamicLayer` that grows as more tokens are processed. The sequence length dimension (`seq_len`) increases with each new token:
 
@@ -94,7 +94,7 @@ cache.layers[idx].keys = torch.cat([cache.layers[idx].keys, key_states], dim=-2)
 cache.layers[idx].values = torch.cat([cache.layers[idx].values, value_states], dim=-2)
 ```
 
-Other layer types like `StaticLayer` and `SlidingWindowLayer` have a fixed sequence length that is set when the cache is created. This makes them compatible with `torch.compile`. In the case of `SlidingWindowLayer`, existing tokens are shifted out of the cache when a new token is added.
+Other layer types like `StaticLayer` and `StaticSlidingWindowLayer` have a fixed sequence length that is set when the cache is created. This makes them compatible with `torch.compile`. In the case of `StaticSlidingWindowLayer`, existing tokens are shifted out of the cache when a new token is added.
 
 The example below demonstrates how to create a generation loop with [`DynamicCache`]. As discussed, the attention mask is a concatenation of past and current token values and `1` is added to the cache position for the next token.
 

docs/source/en/internal/generation_utils.md

@@ -250,7 +250,7 @@ A [`Constraint`] can be used to force the generation to include specific tokens
     - update
     - lazy_initialization
 
-[[autodoc]] SlidingWindowLayer
+[[autodoc]] StaticSlidingWindowLayer
     - update
     - lazy_initialization
 

docs/source/en/model_doc/apertus.md

@@ -13,6 +13,9 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 
 -->
+*This model was released on 2025-09-02 and added to Hugging Face Transformers on 2025-08-28.*
+
+# Apertus
 
 <div style="float: right;">
     <div class="flex flex-wrap space-x-1">
@@ -23,7 +26,7 @@ rendered properly in your Markdown viewer.
     </div>
 </div>
 
-# Apertus
+## Overview
 
 [Apertus](https://www.swiss-ai.org) is a family of large language models from the Swiss AI Initiative.
 

docs/source/en/model_doc/florence2.md

@@ -13,6 +13,9 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 
 -->
+*This model was released on 2024-06-16 and added to Hugging Face Transformers on 2025-08-20.*
+
+# Florence-2
 
 <div style="float: right;">
     <div class="flex flex-wrap space-x-1">
@@ -21,7 +24,7 @@ rendered properly in your Markdown viewer.
     </div>
 </div>
 
-# Florence-2
+## Overview
 
 [Florence-2](https://huggingface.co/papers/2311.06242) is an advanced vision foundation model that uses a prompt-based approach to handle a wide range of vision and vision-language tasks. Florence-2 can interpret simple text prompts to perform tasks like captioning, object detection, and segmentation. It leverages the FLD-5B dataset, containing 5.4 billion annotations across 126 million images, to master multi-task learning. The model's sequence-to-sequence architecture enables it to excel in both zero-shot and fine-tuned settings, proving to be a competitive vision foundation model.
 
@@ -44,7 +47,7 @@ from transformers import pipeline
 
 pipeline = pipeline(
     "image-text-to-text",
-    model="ducviet00/Florence-2-base-hf",
+    model="florence-community/Florence-2-base",
     device=0,
     dtype=torch.bfloat16
 )