undo

Fix docs

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -56,6 +56,12 @@ body:
           - ray/raytune: @richardliaw, @amogkam
           - Big Model Inference: @SunMarc
           - quantization (bitsandbytes, autogpt): @SunMarc @MekkCyber
+        
+        Devices/Backends:
+        
+          - AMD ROCm: @ivarflakstad
+          - Intel XPU: @IlyasMoutawwakil
+          - Ascend NPU: @ivarflakstad 
 
         Documentation: @stevhliu
 

.github/workflows/build_pr_documentation.yml

@@ -14,4 +14,4 @@ jobs:
       commit_sha: ${{ github.event.pull_request.head.sha }}
       pr_number: ${{ github.event.number }}
       package: transformers
-      languages: ar de en es fr hi it ko pt tr zh ja te
+      languages: en

ISSUES.md

@@ -26,7 +26,7 @@ There are two main venues to receive support: [the forums](https://discuss.huggi
 
 [The user forums](https://discuss.huggingface.co/) are supported by the wide community of the library users and backed up by developers when needed.
 
-If you have a difficulty with deploying this library or some questions, or you'd like to discuss a new feature, please first consider discussing those things at the forums. Only when you feel your subject matter has been crystalized and you still need support from the library developers do proceed to file an [issue](https://github.com/huggingface/transformers/issues).
+If you have a difficulty with deploying this library or some questions, or you'd like to discuss a new feature, please first consider discussing those things at the forums. Only when you feel your subject matter has been crystallized and you still need support from the library developers do proceed to file an [issue](https://github.com/huggingface/transformers/issues).
 
 In particular all "Please explain" questions or objectively very user-specific feature requests belong to the forums. Here are some example of such questions:
 

benchmark/benchmark.py

@@ -90,7 +90,7 @@ def summarize(run_dir, metrics, expand_metrics=False):
 
         model = benchmark.config.backend["model"]
 
-        # Ths looks like `benchmark.input_shapes.batch_size=1,benchmark.input_shapes.sequence_length=5`.
+        # This looks like `benchmark.input_shapes.batch_size=1,benchmark.input_shapes.sequence_length=5`.
         # (we rely on the usage of hydra's `${hydra.job.override_dirname}`.)
         benchmark_name = re.sub(f"backend.model={model},*", "", report_dir)
         benchmark_name = str(Path(benchmark_name).parts[-1])

benchmark/llama.py

@@ -293,7 +293,7 @@ def run_benchmark(logger: Logger, branch: str, commit_id: str, commit_msg: str,
                 max_cache_len=seq_length + 128,
             )
 
-            # 3nd call
+            # 3rd call
             start = perf_counter()
             output = model.generate(**inputs, past_key_values=past_key_values)
             end = perf_counter()

docker/consistency.dockerfile

@@ -5,7 +5,7 @@ ARG REF=main
 RUN apt-get update && apt-get install -y time git g++ pkg-config make git-lfs
 ENV UV_PYTHON=/usr/local/bin/python
 RUN pip install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools GitPython
-RUN uv pip install --no-cache-dir --upgrade 'torch' 'torchaudio' 'torchvision' --index-url https://download.pytorch.org/whl/cpu
+RUN uv pip install --no-cache-dir --upgrade 'torch==2.6.0' 'torchaudio==2.6.0' 'torchvision==0.21.0' --index-url https://download.pytorch.org/whl/cpu
 # tensorflow pin matching setup.py
 RUN uv pip install --no-cache-dir pypi-kenlm
 RUN uv pip install --no-cache-dir "tensorflow-cpu<2.16" "tf-keras<2.16"

docker/custom-tokenizers.dockerfile

@@ -16,7 +16,7 @@ RUN cmake .. -DCMAKE_INSTALL_PREFIX=/usr/local
 RUN make install -j 10
 
 
-RUN uv pip install --no-cache --upgrade 'torch' --index-url https://download.pytorch.org/whl/cpu
+RUN uv pip install --no-cache --upgrade 'torch==2.6.0' --index-url https://download.pytorch.org/whl/cpu
 RUN uv pip install --no-cache-dir  --no-deps accelerate --extra-index-url https://download.pytorch.org/whl/cpu
 RUN uv pip install  --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[ja,testing,sentencepiece,jieba,spacy,ftfy,rjieba]" unidic unidic-lite
 # spacy is not used so not tested. Causes to failures. TODO fix later

docker/examples-torch.dockerfile

@@ -5,7 +5,7 @@ USER root
 RUN apt-get update &&  apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git g++ cmake pkg-config openssh-client git
 ENV UV_PYTHON=/usr/local/bin/python
 RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
-RUN uv pip install --no-cache-dir 'torch' 'torchvision' 'torchaudio' --index-url https://download.pytorch.org/whl/cpu
+RUN uv pip install --no-cache-dir 'torch==2.6.0' 'torchaudio==2.6.0' 'torchvision==0.21.0' --index-url https://download.pytorch.org/whl/cpu
 RUN uv pip install --no-deps timm accelerate --extra-index-url https://download.pytorch.org/whl/cpu
 RUN uv pip install --no-cache-dir librosa "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[sklearn,sentencepiece,vision,testing]" seqeval albumentations jiwer
 RUN uv pip uninstall transformers

docker/exotic-models.dockerfile

@@ -5,7 +5,7 @@ USER root
 RUN apt-get update && apt-get install -y libsndfile1-dev espeak-ng time git libgl1-mesa-glx libgl1 g++ tesseract-ocr
 ENV UV_PYTHON=/usr/local/bin/python
 RUN pip --no-cache-dir install uv &&  uv venv && uv pip install --no-cache-dir -U pip setuptools
-RUN uv pip install --no-cache-dir 'torch' 'torchvision' 'torchaudio' --index-url https://download.pytorch.org/whl/cpu
+RUN uv pip install --no-cache-dir 'torch==2.6.0' 'torchaudio==2.6.0' 'torchvision==0.21.0' --index-url https://download.pytorch.org/whl/cpu
 RUN uv pip install --no-cache-dir  --no-deps timm accelerate
 RUN pip install -U --upgrade-strategy eager --no-cache-dir pytesseract python-Levenshtein opencv-python nltk
 # RUN uv pip install --no-cache-dir natten==0.15.1+torch210cpu -f https://shi-labs.com/natten/wheels

docker/pipeline-torch.dockerfile

@@ -5,7 +5,7 @@ USER root
 RUN apt-get update &&  apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git pkg-config openssh-client git
 ENV UV_PYTHON=/usr/local/bin/python
 RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
-RUN uv pip install --no-cache-dir 'torch' 'torchvision' 'torchaudio' --index-url https://download.pytorch.org/whl/cpu
-RUN uv pip install --no-deps timm accelerate --extra-index-url https://download.pytorch.org/whl/cpu 
+RUN uv pip install --no-cache-dir --upgrade 'torch==2.6.0' 'torchaudio==2.6.0' 'torchvision==0.21.0' --index-url https://download.pytorch.org/whl/cpu
+RUN uv pip install --no-deps timm accelerate --extra-index-url https://download.pytorch.org/whl/cpu
 RUN uv pip install --no-cache-dir librosa "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[sklearn,sentencepiece,vision,testing]"
 RUN uv pip uninstall transformers

docker/torch-light.dockerfile

@@ -5,7 +5,7 @@ USER root
 RUN apt-get update &&  apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git g++ cmake pkg-config openssh-client git git-lfs
 ENV UV_PYTHON=/usr/local/bin/python
 RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
-RUN uv pip install --no-cache-dir 'torch' 'torchvision' 'torchaudio' --index-url https://download.pytorch.org/whl/cpu
+RUN uv pip install --no-cache-dir --upgrade 'torch==2.6.0' 'torchaudio==2.6.0' 'torchvision==0.21.0' --index-url https://download.pytorch.org/whl/cpu
 RUN uv pip install --no-deps timm accelerate --extra-index-url https://download.pytorch.org/whl/cpu
 RUN uv pip install --no-cache-dir librosa "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[sklearn,sentencepiece,vision,testing,tiktoken,num2words,video]"
 RUN uv pip uninstall transformers

docker/torch-tf-light.dockerfile

@@ -7,7 +7,7 @@ RUN apt-get update &&  apt-get install -y --no-install-recommends libsndfile1-de
 ENV UV_PYTHON=/usr/local/bin/python
 RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
 RUN uv pip install --no-cache-dir  --no-deps accelerate --extra-index-url https://download.pytorch.org/whl/cpu 
-RUN uv pip install --no-cache-dir 'torch' 'torchvision' 'torchaudio' --index-url https://download.pytorch.org/whl/cpu
+RUN uv pip install --no-cache-dir 'torch==2.6.0' 'torchaudio==2.6.0' 'torchvision==0.21.0' --index-url https://download.pytorch.org/whl/cpu
 RUN git lfs install
 
 RUN uv pip install --no-cache-dir pypi-kenlm

docker/transformers-quantization-latest-gpu/Dockerfile

@@ -84,6 +84,9 @@ RUN python3 -m pip install --no-cache-dir compressed-tensors
 # Add AMD Quark for quantization testing
 RUN python3 -m pip install --no-cache-dir amd-quark
 
+# Add AutoRound for quantization testing
+RUN python3 -m pip install --no-cache-dir "auto-round>=0.5.0"
+
 # Add transformers in editable mode
 RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-torch]
 

docs/source/en/_toctree.yml

@@ -163,8 +163,12 @@
     title: Overview
   - local: quantization/selecting
     title: Selecting a quantization method
+  - local: quantization/concept_guide
+    title: Quantization concepts
   - local: quantization/aqlm
     title: AQLM
+  - local: quantization/auto_round
+    title: AutoRound
   - local: quantization/awq
     title: AWQ
   - local: quantization/bitnet
@@ -281,6 +285,8 @@
         title: Image-text-to-text
       - local: tasks/video_text_to_text
         title: Video-text-to-text
+      - local: tasks/visual_document_retrieval
+        title: Visual Document Retrieval
       title: Multimodal
     title: Task recipes
   - local: run_scripts
@@ -487,8 +493,6 @@
         title: GraniteMoe
       - local: model_doc/granitemoeshared
         title: GraniteMoeShared
-      - local: model_doc/granitevision
-        title: GraniteVision
       - local: model_doc/helium
         title: Helium
       - local: model_doc/herbert
@@ -509,8 +513,6 @@
         title: Llama2
       - local: model_doc/llama3
         title: Llama3
-      - local: model_doc/llama4
-        title: Llama4
       - local: model_doc/longformer
         title: Longformer
       - local: model_doc/longt5
@@ -539,8 +541,6 @@
         title: MegatronGPT2
       - local: model_doc/mistral
         title: Mistral
-      - local: model_doc/mistral3
-        title: Mistral3
       - local: model_doc/mixtral
         title: Mixtral
       - local: model_doc/mluke
@@ -591,8 +591,6 @@
         title: Phi
       - local: model_doc/phi3
         title: Phi-3
-      - local: model_doc/phi4_multimodal
-        title: Phi4 Multimodal
       - local: model_doc/phimoe
         title: PhiMoE
       - local: model_doc/phobert
@@ -737,6 +735,8 @@
         title: Mask2Former
       - local: model_doc/maskformer
         title: MaskFormer
+      - local: model_doc/mlcd
+        title: MLCD
       - local: model_doc/mobilenet_v1
         title: MobileNetV1
       - local: model_doc/mobilenet_v2
@@ -933,6 +933,8 @@
         title: GIT
       - local: model_doc/got_ocr2
         title: GOT-OCR2
+      - local: model_doc/granitevision
+        title: GraniteVision
       - local: model_doc/grounding-dino
         title: Grounding DINO
       - local: model_doc/groupvit
@@ -947,6 +949,10 @@
         title: InstructBLIP
       - local: model_doc/instructblipvideo
         title: InstructBlipVideo
+      - local: model_doc/internvl
+        title: InternVL
+      - local: model_doc/janus
+        title: Janus
       - local: model_doc/kosmos-2
         title: KOSMOS-2
       - local: model_doc/layoutlm
@@ -959,6 +965,8 @@
         title: LayoutXLM
       - local: model_doc/lilt
         title: LiLT
+      - local: model_doc/llama4
+        title: Llama4
       - local: model_doc/llava
         title: Llava
       - local: model_doc/llava_next
@@ -973,6 +981,8 @@
         title: MatCha
       - local: model_doc/mgp-str
         title: MGP-STR
+      - local: model_doc/mistral3
+        title: Mistral3
       - local: model_doc/mllama
         title: mllama
       - local: model_doc/nougat
@@ -989,6 +999,8 @@
         title: PaliGemma
       - local: model_doc/perceiver
         title: Perceiver
+      - local: model_doc/phi4_multimodal
+        title: Phi4 Multimodal
       - local: model_doc/pix2struct
         title: Pix2Struct
       - local: model_doc/pixtral
@@ -1055,6 +1067,8 @@
         title: PatchTST
       - local: model_doc/time_series_transformer
         title: Time Series Transformer
+      - local: model_doc/timesfm
+        title: TimesFM
       title: Time series models
     - sections:
       - local: model_doc/graphormer

docs/source/en/generation_strategies.md

@@ -31,7 +31,7 @@ import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
-inputs = tokenizer("I look forward to", return_tensors="pt").to("cuda")
+inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to("cuda")
 
 model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.float16).to("cuda")
 # explicitly set to default length because Llama2 generation length is 4096

docs/source/en/internal/model_debugging_utils.md

@@ -28,7 +28,7 @@ Most of those are only useful if you are adding new models in the library.
 
 This context manager is a power user tool intended for model adders.
 It tracks all forward calls within a model forward and logs a slice of each input and output on a nested Json.
-To note, this context manager enforces `torch.inference_mode()`.
+To note, this context manager enforces `torch.no_grad()`.
 
 ### Rationale
 
@@ -43,6 +43,7 @@ import torch
 from PIL import Image
 import requests
 from transformers import LlavaProcessor, LlavaForConditionalGeneration
+from transformers.model_debugging_utils import model_addition_debugger_context
 torch.random.manual_seed(673)
 
 # load pretrained model and processor
@@ -60,12 +61,153 @@ prompt = "<image>Describe this image."
 inputs = processor(text=prompt, images=random_image, return_tensors="pt")
 
 # call forward method (not .generate!)
-with model_addition_debugger_context(model, "optional_path_to_your_output_file.json"):
+with model_addition_debugger_context(
+  model,
+  debug_path="optional_path_to_your_directory",
+  do_prune_layers=False # This will output ALL the layers of a model.
+  ):
     output = model.forward(**inputs)
 
 ```
 
 
-[[autodoc]] model_addition_debugger
+### Reading results
+
+The debugger generates two files from the forward call, both with the same base name, 
+but ending either with `_SUMMARY.json` or with `_FULL_TENSORS.json`. 
+
+The first one will contain a summary of each module's _input_ and _output_ tensor values and shapes.
+
+```json
+{
+  "module_path": "MolmoForConditionalGeneration",
+  "inputs": {
+    "args": [],
+    "kwargs": {
+      "input_ids": {
+        "shape": "torch.Size([1, 589])",
+        "dtype": "torch.int64"
+      },
+      "attention_mask": {
+        "shape": "torch.Size([1, 589])",
+        "dtype": "torch.int64"
+      },
+      "pixel_values": {
+        "shape": "torch.Size([1, 5, 576, 588])",
+        "dtype": "torch.float32",
+        "mean": "tensor(-8.9514e-01, device='cuda:0')",
+        "std": "tensor(9.2586e-01, device='cuda:0')",
+        "min": "tensor(-1.7923e+00, device='cuda:0')",
+        "max": "tensor(1.8899e+00, device='cuda:0')"
+    }
+  },
+  "children": [
+    {
+      "module_path": "MolmoForConditionalGeneration.language_model.model.embed_tokens",
+      "inputs": {
+        "args": [
+          {
+            "shape": "torch.Size([1, 589])",
+            "dtype": "torch.int64"
+          }
+        ]
+      },
+      "outputs": {
+        "shape": "torch.Size([1, 589, 3584])",
+        "dtype": "torch.float32",
+        "mean": "tensor(6.5460e-06, device='cuda:0')",
+        "std": "tensor(2.3807e-02, device='cuda:0')",
+        "min": "tensor(-3.3398e-01, device='cuda:0')",
+        "max": "tensor(3.9453e-01, device='cuda:0')"
+      }
+    },
+    {
+      "module_path": "MolmoForConditionalGeneration.vision_tower",
+      "inputs": {
+        "args": [
+          {
+            "shape": "torch.Size([5, 1, 576, 588])",
+            "dtype": "torch.float32",
+            "mean": "tensor(-8.9514e-01, device='cuda:0')",
+            "std": "tensor(9.2586e-01, device='cuda:0')",
+            "min": "tensor(-1.7923e+00, device='cuda:0')",
+            "max": "tensor(1.8899e+00, device='cuda:0')"
+          }
+        ],
+        "kwargs": {
+          "output_hidden_states": "True"
+        }
+      },
+      "children": [
+        { ... and so on
+```
+
+The `_FULL_TENSORS.json` file will display a full view of all tensors, which is useful
+for comparing two files. 
+```json
+      "pixel_values": {
+        "shape": "torch.Size([1, 5, 576, 588])",
+        "dtype": "torch.float32",
+        "value": [
+          "tensor([[[[-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          ...,",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00]],",
+          "",
+          "         [[-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          ...,",
+          "          [-1.4857e+00, -1.4820e+00, -1.2100e+00,  ..., -6.0979e-01, -5.9650e-01, -3.8527e-01],",
+          "          [-1.6755e+00, -1.7221e+00, -1.4518e+00,  ..., -7.5577e-01, -7.4658e-01, -5.5592e-01],",
+          "          [-7.9957e-01, -8.2162e-01, -5.7014e-01,  ..., -1.3689e+00, -1.3169e+00, -1.0678e+00]],",
+          "",
+          "         [[-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          ...,",
+          "          [-3.0322e-01, -5.0645e-01, -5.8436e-01,  ..., -6.2439e-01, -7.9160e-01, -8.1188e-01],",
+          "          [-4.4921e-01, -6.5653e-01, -7.2656e-01,  ..., -3.4702e-01, -5.2146e-01, -5.1326e-01],",
+          "          [-3.4702e-01, -5.3647e-01, -5.4170e-01,  ..., -1.0915e+00, -1.1968e+00, -1.0252e+00]],",
+          "",
+          "         [[-1.1207e+00, -1.2718e+00, -1.0678e+00,  ..., 1.2013e-01, -1.3126e-01, -1.7197e-01],",
+          "          [-6.9738e-01, -9.1166e-01, -8.5454e-01,  ..., -5.5050e-02, -2.8134e-01, -4.2793e-01],",
+          "          [-3.4702e-01, -5.5148e-01, -5.8436e-01,  ..., 1.9312e-01, -8.6235e-02, -2.1463e-01],",
+          "          ...,",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00]],",
+          "",
+          "         [[-1.0039e+00, -9.5669e-01, -6.5546e-01,  ..., -1.4711e+00, -1.4219e+00, -1.1389e+00],",
+          "          [-1.0039e+00, -9.5669e-01, -6.5546e-01,  ..., -1.7193e+00, -1.6771e+00, -1.4091e+00],",
+          "          [-1.6317e+00, -1.6020e+00, -1.2669e+00,  ..., -1.2667e+00, -1.2268e+00, -8.9720e-01],",
+          "          ...,",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00],",
+          "          [-1.7923e+00, -1.7521e+00, -1.4802e+00,  ..., -1.7923e+00, -1.7521e+00, -1.4802e+00]]]], device='cuda:0')"
+        ],
+        "mean": "tensor(-8.9514e-01, device='cuda:0')",
+        "std": "tensor(9.2586e-01, device='cuda:0')",
+        "min": "tensor(-1.7923e+00, device='cuda:0')",
+        "max": "tensor(1.8899e+00, device='cuda:0')"
+      },
+```
+
+### Comparing between implementations
+
+Once the forward passes of two models have been traced by the debugger, one can compare the `json` output files. See below: we can see slight differences between these two implementations' key projection layer. Inputs are mostly identical, but not quite. Looking through the file differences makes it easier to pinpoint which layer is wrong. 
+
+
+![download-icon](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/files_difference_debugging.png)
+
+
+### Limitations and scope
+
+This feature will only work for torch-based models, and would require more work and case-by-case approach for say `jax`-based models that are usually compiled. Models relying heavily on external kernel calls may work, but trace will probably miss some things. Regardless, any python implementation that aims at mimicking another implementation can be traced once instead of reran N times with breakpoints.
+
+If you pass `do_prune_layers=False` to your model debugger, ALL the layers will be outputted to `json`. Else, only the first and last layer will be shown. This is useful when some layers (typically cross-attention) appear only after N layers. 
 
 [[autodoc]] model_addition_debugger_context

docs/source/en/internal/modeling_utils.md

@@ -20,6 +20,10 @@ This page lists all the custom layers used by the library, as well as the utilit
 
 Most of those are only useful if you are studying the code of the models in the library.
 
+## Layers
+
+[[autodoc]] GradientCheckpointingLayer
+
 ## Attention Functions
 
 [[autodoc]] AttentionInterface
@@ -33,23 +37,6 @@ Most of those are only useful if you are studying the code of the models in the
 
 [[autodoc]] pytorch_utils.Conv1D
 
-[[autodoc]] modeling_utils.PoolerStartLogits
-    - forward
-
-[[autodoc]] modeling_utils.PoolerEndLogits
-    - forward
-
-[[autodoc]] modeling_utils.PoolerAnswerClass
-    - forward
-
-[[autodoc]] modeling_utils.SquadHeadOutput
-
-[[autodoc]] modeling_utils.SQuADHead
-    - forward
-
-[[autodoc]] modeling_utils.SequenceSummary
-    - forward
-
 ## PyTorch Helper Functions
 
 [[autodoc]] pytorch_utils.apply_chunking_to_forward

docs/source/en/kv_cache.md

@@ -18,7 +18,7 @@ rendered properly in your Markdown viewer.
 
 The key-value (KV) vectors are used to calculate attention scores. For autoregressive models, KV scores are calculated *every* time because the model predicts one token at a time. Each prediction depends on the previous tokens, which means the model performs the same computations each time.
 
-A KV *cache* stores these calculations so they can be reused without recomputing them. Efficient caching is crucial for optimizing model performance because it reduces computation time and improves response rates. Refer to the [Caching](./cache_explanation.md) doc for a more detailed explanation about how a cache works.
+A KV *cache* stores these calculations so they can be reused without recomputing them. Efficient caching is crucial for optimizing model performance because it reduces computation time and improves response rates. Refer to the [Caching](./cache_explanation) doc for a more detailed explanation about how a cache works.
 
 Transformers offers several [`Cache`] classes that implement different caching mechanisms. Some of these [`Cache`] classes are optimized to save memory while others are designed to maximize generation speed. Refer to the table below to compare cache types and use it to help you select the best cache for your use case.
 

docs/source/en/main_classes/quantization.md

@@ -92,3 +92,7 @@ Learn how to quantize models in the [Quantization](../quantization) guide.
 ## QuarkConfig
 
 [[autodoc]] QuarkConfig
+
+## AutoRoundConfig
+
+[[autodoc]] AutoRoundConfig

docs/source/en/model_doc/bridgetower.md

@@ -147,6 +147,11 @@ Tips:
 [[autodoc]] BridgeTowerImageProcessor
     - preprocess
 
+## BridgeTowerImageProcessorFast
+
+[[autodoc]] BridgeTowerImageProcessorFast
+    - preprocess
+
 ## BridgeTowerProcessor
 
 [[autodoc]] BridgeTowerProcessor

docs/source/en/model_doc/chinese_clip.md

@@ -90,6 +90,11 @@ Currently, following scales of pretrained Chinese-CLIP models are available on
 [[autodoc]] ChineseCLIPImageProcessor
     - preprocess
 
+## ChineseCLIPImageProcessorFast
+
+[[autodoc]] ChineseCLIPImageProcessorFast
+    - preprocess
+
 ## ChineseCLIPFeatureExtractor
 
 [[autodoc]] ChineseCLIPFeatureExtractor

docs/source/en/model_doc/colpali.md

@@ -1,5 +1,4 @@
 <!--Copyright 2024 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
 
@@ -9,76 +8,134 @@ Unless required by applicable law or agreed to in writing, software distributed
 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
+⚠️ Note that this file is in Markdown but contains specific syntax for our doc-builder (similar to MDX) that may not be
 rendered properly in your Markdown viewer.
-
 -->
 
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
+</div>
+
 # ColPali
 
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-</div>
+[ColPali](https://huggingface.co/papers/2407.01449) is a model designed to retrieve documents by analyzing their visual features. Unlike traditional systems that rely heavily on text extraction and OCR, ColPali treats each page as an image. It uses [Paligemma-3B](./paligemma) to capture not only text, but also the layout, tables, charts, and other visual elements to create detailed embeddings. This offers a more comprehensive understanding of documents and enables more efficient and accurate retrieval.
 
-## Overview
+You can find all the original ColPali checkpoints under the [ColPali](https://huggingface.co/collections/vidore/hf-native-colvision-models-6755d68fc60a8553acaa96f7) collection.
 
-The *ColPali* model was proposed in [ColPali: Efficient Document Retrieval with Vision Language Models](https://doi.org/10.48550/arXiv.2407.01449) by **Manuel Faysse***, **Hugues Sibille***, **Tony Wu***, Bilel Omrani, Gautier Viaud, Céline Hudelot, Pierre Colombo (* denotes equal contribution). Work lead by ILLUIN Technology.
+> [!TIP]
+> Click on the ColPali models in the right sidebar for more examples of how to use ColPali for image retrieval.
 
-In our proposed *ColPali* approach, we leverage VLMs to construct efficient multi-vector embeddings directly from document images (“screenshots”) for document retrieval. We train the model to maximize the similarity between these document embeddings and the corresponding query embeddings, using the late interaction method introduced in ColBERT.
+<hfoptions id="usage">
+<hfoption id="image retrieval">
 
-Using *ColPali* removes the need for potentially complex and brittle layout recognition and OCR pipelines with a single model that can take into account both the textual and visual content (layout, charts, etc.) of a document.
-
-## Resources
-
-- The *ColPali* arXiv paper can be found [here](https://doi.org/10.48550/arXiv.2407.01449). 📄
-- The official blog post detailing ColPali can be found [here](https://huggingface.co/blog/manu/colpali). 📝
-- The original model implementation code for the ColPali model and for the `colpali-engine` package can be found [here](https://github.com/illuin-tech/colpali). 🌎
-- Cookbooks for learning to use the transformers-native version of *ColPali*, fine-tuning, and similarity maps generation can be found [here](https://github.com/tonywu71/colpali-cookbooks). 📚
-
-This model was contributed by [@tonywu71](https://huggingface.co/tonywu71) and [@yonigozlan](https://huggingface.co/yonigozlan).
-
-## Usage
-
-This example demonstrates how to use *ColPali* to embed both queries and images, calculate their similarity scores, and identify the most relevant matches. For a specific query, you can retrieve the top-k most similar images by selecting the ones with the highest similarity scores.
-
-```python
+```py
+import requests
 import torch
 from PIL import Image
-
 from transformers import ColPaliForRetrieval, ColPaliProcessor
 
-model_name = "vidore/colpali-v1.2-hf"
-
+# Load model (bfloat16 support is limited; fallback to float32 if needed)
 model = ColPaliForRetrieval.from_pretrained(
-    model_name,
-    torch_dtype=torch.bfloat16,
-    device_map="cuda:0",  # or "mps" if on Apple Silicon
+    "vidore/colpali-v1.2-hf",
+    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
+    device_map="auto",  # "cpu", "cuda", or "mps" for Apple Silicon
 ).eval()
 
 processor = ColPaliProcessor.from_pretrained(model_name)
 
-# Your inputs (replace dummy images with screenshots of your documents)
+url1 = "https://upload.wikimedia.org/wikipedia/commons/8/89/US-original-Declaration-1776.jpg"
+url2 = "https://upload.wikimedia.org/wikipedia/commons/thumb/4/4c/Romeoandjuliet1597.jpg/500px-Romeoandjuliet1597.jpg"
+
 images = [
-    Image.new("RGB", (32, 32), color="white"),
-    Image.new("RGB", (16, 16), color="black"),
+    Image.open(requests.get(url1, stream=True).raw),
+    Image.open(requests.get(url2, stream=True).raw),
 ]
+
 queries = [
-    "What is the organizational structure for our R&D department?",
-    "Can you provide a breakdown of last year’s financial performance?",
+    "Who printed the edition of Romeo and Juliet?",
+    "When was the United States Declaration of Independence proclaimed?",
 ]
 
 # Process the inputs
-batch_images = processor(images=images).to(model.device)
-batch_queries = processor(text=queries).to(model.device)
+inputs_images = processor(images=images, return_tensors="pt").to(model.device)
+inputs_text = processor(text=queries, return_tensors="pt").to(model.device)
 
 # Forward pass
 with torch.no_grad():
-    image_embeddings = model(**batch_images).embeddings
-    query_embeddings = model(**batch_queries).embeddings
+    image_embeddings = model(**inputs_images).embeddings
+    query_embeddings = model(**inputs_text).embeddings
 
-# Score the queries against the images
 scores = processor.score_retrieval(query_embeddings, image_embeddings)
+
+print("Retrieval scores (query x image):")
+print(scores)
 ```
+</hfoption>
+</hfoptions>
+
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
+
+The example below uses [bitsandbytes](../quantization/bitsandbytes.md) to quantize the weights to int4.
+
+```py
+import requests
+import torch
+from PIL import Image
+from transformers import ColPaliForRetrieval, ColPaliProcessor
+from transformers import BitsAndBytesConfig
+
+# 4-bit quantization configuration
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_use_double_quant=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype=torch.float16,
+)
+
+model_name = "vidore/colpali-v1.2-hf"
+
+# Load model 
+model = ColPaliForRetrieval.from_pretrained(
+    model_name,
+    quantization_config=bnb_config,
+    device_map="cuda"
+).eval()
+
+processor = ColPaliProcessor.from_pretrained(model_name)
+
+url1 = "https://upload.wikimedia.org/wikipedia/commons/8/89/US-original-Declaration-1776.jpg"
+url2 = "https://upload.wikimedia.org/wikipedia/commons/thumb/4/4c/Romeoandjuliet1597.jpg/500px-Romeoandjuliet1597.jpg"
+
+images = [
+    Image.open(requests.get(url1, stream=True).raw),
+    Image.open(requests.get(url2, stream=True).raw),
+]
+
+queries = [
+    "Who printed the edition of Romeo and Juliet?",
+    "When was the United States Declaration of Independence proclaimed?",
+]
+
+# Process the inputs
+inputs_images = processor(images=images, return_tensors="pt").to(model.device)
+inputs_text = processor(text=queries, return_tensors="pt").to(model.device)
+
+# Forward pass
+with torch.no_grad():
+    image_embeddings = model(**inputs_images).embeddings
+    query_embeddings = model(**inputs_text).embeddings
+
+scores = processor.score_retrieval(query_embeddings, image_embeddings)
+
+print("Retrieval scores (query x image):")
+print(scores)
+```
+
+## Notes
+
+- [`~ColPaliProcessor.score_retrieval`] returns a 2D tensor where the first dimension is the number of queries and the second dimension is the number of images. A higher score indicates more similarity between the query and image.
 
 ## ColPaliConfig
 

docs/source/en/model_doc/conditional_detr.md

@@ -48,6 +48,11 @@ This model was contributed by [DepuMeng](https://huggingface.co/DepuMeng). The o
 
 [[autodoc]] ConditionalDetrImageProcessor
     - preprocess
+
+## ConditionalDetrImageProcessorFast
+
+[[autodoc]] ConditionalDetrImageProcessorFast
+    - preprocess
     - post_process_object_detection
     - post_process_instance_segmentation
     - post_process_semantic_segmentation

docs/source/en/model_doc/efficientnet.md

@@ -43,6 +43,11 @@ The original code can be found [here](https://github.com/tensorflow/tpu/tree/mas
 [[autodoc]] EfficientNetImageProcessor
     - preprocess
 
+## EfficientNetImageProcessorFast
+
+[[autodoc]] EfficientNetImageProcessorFast
+    - preprocess
+
 ## EfficientNetModel
 
 [[autodoc]] EfficientNetModel

docs/source/en/model_doc/gemma.md

@@ -1,4 +1,5 @@
-<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+<!--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
@@ -14,31 +15,146 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# Gemma
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="Flax" src="https://img.shields.io/badge/Flax-29a79b.svg?style=flat&logo=data:image/png;base64,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
-">
-<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
-<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
+        <img alt="Flax" src="https://img.shields.io/badge/Flax-29a79b.svg?style=flat&logo=data:image/png;base64,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
+        ">
+        <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
 </div>
 
-## Overview
+# Gemma
 
-The Gemma model was proposed in [Gemma: Open Models Based on Gemini Technology and Research](https://blog.google/technology/developers/gemma-open-models/) by Gemma Team, Google.
-Gemma models are trained on 6T tokens, and released with 2 versions, 2b and 7b.
+[Gemma](https://huggingface.co/papers/2403.08295) is a family of lightweight language models with pretrained and instruction-tuned variants, available in 2B and 7B parameters. The architecture is based on a transformer decoder-only design. It features Multi-Query Attention, rotary positional embeddings (RoPE), GeGLU activation functions, and RMSNorm layer normalization.
 
-The abstract from the paper is the following:
+The instruction-tuned variant was fine-tuned with supervised learning on instruction-following data, followed by reinforcement learning from human feedback (RLHF) to align the model outputs with human preferences.
 
-*This work introduces Gemma, a new family of open language models demonstrating strong performance across academic benchmarks for language understanding, reasoning, and safety. We release two sizes of models (2 billion and 7 billion parameters), and provide both pretrained and fine-tuned checkpoints. Gemma outperforms similarly sized open models on 11 out of 18 text-based tasks, and we present comprehensive evaluations of safety and responsibility aspects of the models, alongside a detailed description of our model development. We believe the responsible release of LLMs is critical for improving the safety of frontier models, and for enabling the next wave of LLM innovations*
+You can find all the original Gemma checkpoints under the [Gemma](https://huggingface.co/collections/google/gemma-release-65d5efbccdbb8c4202ec078b) release.
 
-Tips:
 
-- The original checkpoints can be converted using the conversion script `src/transformers/models/gemma/convert_gemma_weights_to_hf.py` 
+> [!TIP]
+> Click on the Gemma models in the right sidebar for more examples of how to apply Gemma to different language tasks.
 
-This model was contributed by [Arthur Zucker](https://huggingface.co/ArthurZ), [Younes Belkada](https://huggingface.co/ybelkada), [Sanchit Gandhi](https://huggingface.co/sanchit-gandhi), [Pedro Cuenca](https://huggingface.co/pcuenq).
+The example below demonstrates how to generate text with [`Pipeline`] or the [`AutoModel`] class, and from the command line.
 
+<hfoptions id="usage">
+<hfoption id="Pipeline">
+
+```py
+import torch
+from transformers import pipeline
+
+pipeline = pipeline(
+    task="text-generation",
+    model="google/gemma-2b",
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+)
+
+pipeline("LLMs generate text through a process known as", max_new_tokens=50)
+```
+
+</hfoption>
+<hfoption id="AutoModel">
+
+```py
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
+model = AutoModelForCausalLM.from_pretrained(
+    "google/gemma-2b",
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+    attn_implementation="sdpa"
+)
+
+input_text = "LLMs generate text through a process known as"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+outputs = model.generate(**input_ids, max_new_tokens=50, cache_implementation="static")
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+</hfoption>
+<hfoption id="transformers-cli">
+
+```bash
+echo -e "LLMs generate text through a process known as" | transformers-cli run --task text-generation --model google/gemma-2b --device 0
+```
+
+</hfoption>
+</hfoptions>
+
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
+
+The example below uses [bitsandbytes](../quantization/bitsandbytes) to only quantize the weights to int4.
+
+```py
+#!pip install bitsandbytes
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
+
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.bfloat16,
+    bnb_4bit_quant_type="nf4"
+)
+tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
+model = AutoModelForCausalLM.from_pretrained(
+    "google/gemma-7b",
+    quantization_config=quantization_config,
+    device_map="auto",
+    attn_implementation="sdpa"
+)
+
+input_text = "LLMs generate text through a process known as."
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+outputs = model.generate(
+    **input_ids, 
+    max_new_tokens=50, 
+    cache_implementation="static"
+)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+Use the [AttentionMaskVisualizer](https://github.com/huggingface/transformers/blob/beb9b5b02246b9b7ee81ddf938f93f44cfeaad19/src/transformers/utils/attention_visualizer.py#L139) to better understand what tokens the model can and cannot attend to.
+
+```py
+from transformers.utils.attention_visualizer import AttentionMaskVisualizer
+
+visualizer = AttentionMaskVisualizer("google/gemma-2b")
+visualizer("LLMs generate text through a process known as") 
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/gemma-attn-mask.png"/>
+</div>
+
+## Notes
+
+- The original Gemma models support standard kv-caching used in many transformer-based language models. You can use use the default [`DynamicCache`] instance or a tuple of tensors for past key values during generation. This makes it compatible with typical autoregressive generation workflows.
+
+   ```py
+   import torch
+   from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache
+
+   tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
+   model = AutoModelForCausalLM.from_pretrained(
+       "google/gemma-2b",
+       torch_dtype=torch.bfloat16,
+       device_map="auto",
+       attn_implementation="sdpa"
+   )
+   input_text = "LLMs generate text through a process known as"
+   input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+   past_key_values = DynamicCache()
+   outputs = model.generate(**input_ids, max_new_tokens=50, past_key_values=past_key_values)
+   print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+   ```
 
 ## GemmaConfig
 

docs/source/en/model_doc/grounding-dino.md

@@ -102,6 +102,11 @@ A list of official Hugging Face and community (indicated by 🌎) resources to h
 
 [[autodoc]] GroundingDinoImageProcessor
     - preprocess
+
+## GroundingDinoImageProcessorFast
+
+[[autodoc]] GroundingDinoImageProcessorFast
+    - preprocess
     - post_process_object_detection
 
 ## GroundingDinoProcessor

docs/source/en/model_doc/internvl.md

@@ -0,0 +1,350 @@
+<!--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.
+
+-->
+
+
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+    </div>
+</div>
+
+# InternVL
+
+The InternVL3 family of Visual Language Models was introduced in [InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models](https://huggingface.co/papers/2504.10479).
+
+The abstract from the paper is the following:
+
+*We introduce InternVL3, a significant advancement in the InternVL series featuring a native multimodal pre-training paradigm. Rather than adapting a text-only large language model (LLM) into a multimodal large language model (MLLM) that supports visual inputs, InternVL3 jointly acquires multimodal and linguistic capabilities from both diverse multimodal data and pure-text corpora during a single pre-training stage. This unified training paradigm effectively addresses the complexities and alignment challenges commonly encountered in conventional post-hoc training pipelines for MLLMs. To further improve performance and scalability, InternVL3 incorporates variable visual position encoding (V2PE) to support extended multimodal contexts, employs advanced post-training techniques such as supervised fine-tuning (SFT) and mixed preference optimization (MPO), and adopts test-time scaling strategies alongside an optimized training infrastructure. Extensive empirical evaluations demonstrate that InternVL3 delivers superior performance across a wide range of multi-modal tasks. In particular, InternVL3-78B achieves a score of 72.2 on the MMMU benchmark, setting a new state-of-the-art among open-source MLLMs. Its capabilities remain highly competitive with leading proprietary models, including ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 2.5 Pro, while also maintaining strong pure-language proficiency. In pursuit of open-science principles, we will publicly release both the training data and model weights to foster further research and development in next-generation MLLMs.*
+
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/internvl_architecture.png" alt="drawing" width="600"/>
+
+<small> Overview of InternVL3 models architecture, which is the same as InternVL2.5. Taken from the <a href="https://huggingface.co/OpenGVLab/InternVL3-1B">original checkpoint.</a> </small>
+
+
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/internvl_overview_performance.png" alt="drawing" width="600"/>
+
+<small> Comparison of InternVL3 performance on OpenCompass against other SOTA VLLMs. Taken from the <a href="https://huggingface.co/OpenGVLab/InternVL3-1B">original checkpoint.</a> </small>
+
+
+
+This model was contributed by [yonigozlan](https://huggingface.co/yonigozlan).
+The original code can be found [here](https://github.com/OpenGVLab/InternVL).
+
+## Usage example
+
+### Inference with Pipeline
+
+Here is how you can use the `image-text-to-text` pipeline to perform inference with the `InternVL3` models in just a few lines of code:
+
+```python
+>>> from transformers import pipeline
+
+>>> messages = [
+...     {
+...         "role": "user",
+...         "content": [
+...             {
+...                 "type": "image",
+...                 "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg",
+...             },
+...             {"type": "text", "text": "Describe this image."},
+...         ],
+...     },
+... ]
+
+>>> pipe = pipeline("image-text-to-text", model="OpenGVLab/InternVL3-1B-hf")
+>>> outputs = pipe(text=messages, max_new_tokens=50, return_full_text=False)
+>>> outputs[0]["generated_text"]
+'The image showcases a vibrant scene of nature, featuring several flowers and a bee. \n\n1. **Foreground Flowers**: \n   - The primary focus is on a large, pink cosmos flower with a prominent yellow center. The petals are soft and slightly r'
+```
+### Inference on a single image
+
+This example demonstrates how to perform inference on a single image with the InternVL models using chat templates.
+
+> [!NOTE]
+> Note that the model has been trained with a specific prompt format for chatting. Use `processor.apply_chat_template(my_conversation_dict)` to correctly format your prompts.
+
+```python
+>>> from transformers import AutoProcessor, AutoModelForImageTextToText
+>>> import torch
+
+>>> torch_device = "cuda"
+>>> model_checkpoint = "OpenGVLab/InternVL3-1B-hf"
+>>> processor = AutoProcessor.from_pretrained(model_checkpoint)
+>>> model = AutoModelForImageTextToText.from_pretrained(model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16)
+
+>>> messages = [
+...     {
+...         "role": "user",
+...         "content": [
+...             {"type": "image", "url": "http://images.cocodataset.org/val2017/000000039769.jpg"},
+...             {"type": "text", "text": "Please describe the image explicitly."},
+...         ],
+...     }
+... ]
+
+>>> inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device, dtype=torch.bfloat16)
+
+>>> generate_ids = model.generate(**inputs, max_new_tokens=50)
+>>> decoded_output = processor.decode(generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
+
+>>> decoded_output
+'The image shows two cats lying on a pink blanket. The cat on the left is a tabby with a mix of brown, black, and white fur, and it appears to be sleeping with its head resting on the blanket. The cat on the'
+```
+
+### Text-only generation
+This example shows how to generate text using the InternVL model without providing any image input.
+
+
+```python
+>>> from transformers import AutoProcessor, AutoModelForImageTextToText
+>>> import torch
+
+>>> torch_device = "cuda"
+>>> model_checkpoint = "OpenGVLab/InternVL3-1B-hf"
+>>> processor = AutoProcessor.from_pretrained(model_checkpoint)
+>>> model = AutoModelForImageTextToText.from_pretrained(model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16)
+
+>>> messages = [
+...     {
+...         "role": "user",
+...         "content": [
+...             {"type": "text", "text": "Write a haiku"},
+...         ],
+...     }
+... ]
+
+>>> inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(torch_device, dtype=torch.bfloat16)
+
+>>> generate_ids = model.generate(**inputs, max_new_tokens=50)
+>>> decoded_output = processor.decode(generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
+
+>>> print(decoded_output)
+"Whispers of dawn,\nSilent whispers of the night,\nNew day's light begins."
+```
+
+### Batched image and text inputs
+InternVL models also support batched image and text inputs.
+
+```python
+>>> from transformers import AutoProcessor, AutoModelForImageTextToText
+>>> import torch
+
+>>> torch_device = "cuda"
+>>> model_checkpoint = "OpenGVLab/InternVL3-1B-hf"
+>>> processor = AutoProcessor.from_pretrained(model_checkpoint)
+>>> model = AutoModelForImageTextToText.from_pretrained(model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16)
+
+>>> messages = [
+...     [
+...         {
+...             "role": "user",
+...             "content": [
+...                 {"type": "image", "url": "https://llava-vl.github.io/static/images/view.jpg"},
+...                 {"type": "text", "text": "Write a haiku for this image"},
+...             ],
+...         },
+...     ],
+...     [
+...         {
+...             "role": "user",
+...             "content": [
+...                 {"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"},
+...                 {"type": "text", "text": "Describe this image"},
+...             ],
+...         },
+...     ],
+... ]
+
+
+>>> inputs = processor.apply_chat_template(messages, padding=True, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device, dtype=torch.bfloat16)
+
+>>> output = model.generate(**inputs, max_new_tokens=25)
+
+>>> decoded_outputs = processor.batch_decode(output, skip_special_tokens=True)
+>>> decoded_outputs
+["user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature's peace.",
+ 'user\n\nDescribe this image\nassistant\nThe image shows a street scene with a traditional Chinese archway, known as a "Chinese Gate" or "Chinese Gate of']
+```
+
+### Batched multi-image input
+This implementation of the InternVL models supports batched text-images inputs with different number of images for each text.
+
+```python
+>>> from transformers import AutoProcessor, AutoModelForImageTextToText
+>>> import torch
+
+>>> torch_device = "cuda"
+>>> model_checkpoint = "OpenGVLab/InternVL3-1B-hf"
+>>> processor = AutoProcessor.from_pretrained(model_checkpoint)
+>>> model = AutoModelForImageTextToText.from_pretrained(model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16)
+
+>>> messages = [
+...     [
+...         {
+...             "role": "user",
+...             "content": [
+...                 {"type": "image", "url": "https://llava-vl.github.io/static/images/view.jpg"},
+...                 {"type": "text", "text": "Write a haiku for this image"},
+...             ],
+...         },
+...     ],
+...     [
+...         {
+...             "role": "user",
+...             "content": [
+...                 {"type": "image", "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"},
+...                 {"type": "image", "url": "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg"},
+...                 {"type": "text", "text": "These images depict two different landmarks. Can you identify them?"},
+...             ],
+...         },
+...     ],
+>>> ]
+
+>>> inputs = processor.apply_chat_template(messages, padding=True, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device, dtype=torch.bfloat16)
+
+>>> output = model.generate(**inputs, max_new_tokens=25)
+
+>>> decoded_outputs = processor.batch_decode(output, skip_special_tokens=True)
+>>> decoded_outputs
+["user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature's peace.",
+ 'user\n\n\nThese images depict two different landmarks. Can you identify them?\nassistant\nYes, these images depict the Statue of Liberty and the Golden Gate Bridge.']
+```
+
+### Video input
+InternVL models can also handle video inputs. Here is an example of how to perform inference on a video input using chat templates.
+
+```python
+>>> from transformers import AutoProcessor, AutoModelForImageTextToText, BitsAndBytesConfig
+
+>>> model_checkpoint = "OpenGVLab/InternVL3-8B-hf"
+>>> quantization_config = BitsAndBytesConfig(load_in_4bit=True)
+>>> processor = AutoProcessor.from_pretrained(model_checkpoint)
+>>> model = AutoModelForImageTextToText.from_pretrained(model_checkpoint, quantization_config=quantization_config)
+
+>>> messages = [
+...     {
+...         "role": "user",
+...         "content": [
+...             {
+...                 "type": "video",
+...                 "url": "https://huggingface.co/datasets/hf-internal-testing/fixtures_videos/resolve/main/tennis.mp4",
+...             },
+...             {"type": "text", "text": "What type of shot is the man performing?"},
+...         ],
+...     }
+>>> ]
+>>> inputs = processor.apply_chat_template(
+...     messages,
+...     return_tensors="pt",
+...     add_generation_prompt=True,
+...     tokenize=True,
+...     return_dict=True,
+...     num_frames=8,
+>>> ).to(model.device, dtype=torch.float16)
+
+>>> output = model.generate(**inputs, max_new_tokens=25)
+
+>>> decoded_output = processor.decode(output[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
+>>> decoded_output
+'The man is performing a forehand shot.'
+```
+
+### Interleaved image and video inputs
+This example showcases how to handle a batch of chat conversations with interleaved image and video inputs using chat template.
+
+```python
+>>> from transformers import AutoProcessor, AutoModelForImageTextToText, BitsAndBytesConfig
+>>> import torch
+
+>>> torch_device = "cuda"
+>>> model_checkpoint = "OpenGVLab/InternVL3-1B-hf"
+>>> processor = AutoProcessor.from_pretrained(model_checkpoint)
+>>> model = AutoModelForImageTextToText.from_pretrained(model_checkpoint, device_map=torch_device, torch_dtype=torch.bfloat16)
+
+>>> messages = [
+...     [
+...         {
+...             "role": "user",
+...             "content": [
+...                 {"type": "image", "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"},
+...                 {"type": "image", "url": "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg"},
+...                 {"type": "text", "text": "These images depict two different landmarks. Can you identify them?"},
+...             ],
+...         },
+...     ],
+...     [
+...         {
+...             "role": "user",
+...             "content": [
+...                 {"type": "video", "url": "https://huggingface.co/datasets/hf-internal-testing/fixtures_videos/resolve/main/tennis.mp4"},
+...                 {"type": "text", "text": "What type of shot is the man performing?"},
+...             ],
+...         },
+...     ],
+...     [
+...         {
+...             "role": "user",
+...             "content": [
+...                 {"type": "image", "url": "https://llava-vl.github.io/static/images/view.jpg"},
+...                 {"type": "text", "text": "Write a haiku for this image"},
+...             ],
+...         },
+...     ],
+>>> ]
+>>> inputs = processor.apply_chat_template(
+...     messages,
+...     padding=True,
+...     add_generation_prompt=True,
+...     tokenize=True,
+...     return_dict=True,
+...     return_tensors="pt",
+>>> ).to(model.device, dtype=torch.bfloat16)
+
+>>> outputs = model.generate(**inputs, max_new_tokens=25)
+
+>>> decoded_outputs = processor.batch_decode(outputs, skip_special_tokens=True)
+>>> decoded_outputs
+['user\n\n\nThese images depict two different landmarks. Can you identify them?\nassistant\nThe images depict the Statue of Liberty and the Golden Gate Bridge.',
+ 'user\nFrame1: \nFrame2: \nFrame3: \nFrame4: \nFrame5: \nFrame6: \nFrame7: \nFrame8: \nWhat type of shot is the man performing?\nassistant\nA forehand shot',
+ "user\n\nWrite a haiku for this image\nassistant\nSilky lake,  \nWooden pier,  \nNature's peace."]
+```
+
+## InternVLVisionConfig
+
+[[autodoc]] InternVLVisionConfig
+
+## InternVLConfig
+
+[[autodoc]] InternVLConfig
+
+## InternVLVisionModel
+
+[[autodoc]] InternVLVisionModel
+    - forward
+
+## InternVLForConditionalGeneration
+
+[[autodoc]] InternVLForConditionalGeneration
+    - forward
+
+## InternVLProcessor
+
+[[autodoc]] InternVLProcessor

docs/source/en/model_doc/janus.md

@@ -0,0 +1,230 @@
+<!--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.
+
+-->
+
+# Janus
+
+## Overview
+
+The Janus Model was originally proposed in [Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation](https://arxiv.org/abs/2410.13848) by DeepSeek AI team and later refined in [Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling](https://arxiv.org/abs/2501.17811). Janus is a vision-language model that can generate both image and text output, it can also take both images and text as input.
+
+> [!NOTE]
+> The model doesn't generate both images and text in an interleaved format. The user has to pass a parameter indicating whether to generate text or image.
+
+The abstract from the original paper is the following:
+
+*In this paper, we introduce Janus, an autoregressive framework that unifies multimodal understanding and generation. Prior research often relies on a single visual encoder for both tasks, such as Chameleon. However, due to the differing levels of information granularity required by multimodal understanding and generation, this approach can lead to suboptimal performance, particularly in multimodal understanding. To address this issue, we decouple visual encoding into separate pathways, while still leveraging a single, unified transformer architecture for processing. The decoupling not only alleviates the conflict between the visual encoder's roles in understanding and generation, but also enhances the framework's flexibility. For instance, both the multimodal understanding and generation components can independently select their most suitable encoding methods. Experiments show that Janus surpasses previous unified model and matches or exceeds the performance of task-specific models. The simplicity, high flexibility, and effectiveness of Janus make it a strong candidate for next-generation unified multimodal models.*
+
+The abstract from the aforementioned `Janus-Pro` paper, released afterwards, is the following:
+
+*In this work, we introduce Janus-Pro, an advanced version of the previous work Janus. Specifically, Janus-Pro incorporates (1) an optimized training strate (2) expanded training data, and (3) scaling to larger model size. With these improvements, Janus-Pro achieves significant advancements in both multimodal understanding and text-to-image instruction-following capabilities, while also enhancing the stability of text-to-image generation. We hope this work will inspire further exploration in the field. Code and models are publicly available.*
+
+This model was contributed by [Yaswanth Gali](https://huggingface.co/yaswanthgali) and [Hugo Silva](https://huggingface.co/hugosilva664).
+The original code can be found [here](https://github.com/deepseek-ai/Janus).
+
+## Usage Example
+
+### Single image inference
+
+Here is the example of visual understanding with a single image.
+
+> [!NOTE]
+> Note that the model has been trained with a specific prompt format for chatting. Use `processor.apply_chat_template(my_conversation_dict)` to correctly format your prompts.
+
+```python
+import torch  
+from PIL import Image  
+import requests  
+
+from transformers import JanusForConditionalGeneration, JanusProcessor  
+
+model_id = "deepseek-community/Janus-Pro-1B"
+# Prepare Input for generation.
+messages = [
+    {
+        "role": "user",
+        "content": [
+            {'type':'image', 'url': 'http://images.cocodataset.org/val2017/000000039769.jpg'},
+            {'type':"text", "text":"What do you see in this image?."}
+        ]
+    },
+]
+
+# Set generation mode to `text` to perform text generation.
+processor = JanusProcessor.from_pretrained(model_id)
+model = JanusForConditionalGeneration.from_pretrained(model_id,     
+        torch_dtype=torch.bfloat16,
+        device_map="auto")
+
+inputs = processor.apply_chat_template(
+    messages,
+    add_generation_prompt=True,
+    generation_mode="text",
+    tokenize=True,
+    return_dict=True,
+    return_tensors="pt",
+).to(model.device, dtype=torch.bfloat16)
+
+output = model.generate(**inputs, max_new_tokens=40,generation_mode='text',do_sample=True)
+text = processor.decode(output[0], skip_special_tokens=True)
+print(text)
+```
+
+### Multi image inference
+
+Janus can perform inference with multiple images as input, where images can belong to the same prompt or different prompts in batched inference, where the model processes many conversations in parallel. Here is how you can do it:
+
+```python
+import torch
+from PIL import Image
+import requests
+
+from transformers import JanusForConditionalGeneration, JanusProcessor
+
+model_id = "deepseek-community/Janus-Pro-1B"
+
+image_urls = [
+    "http://images.cocodataset.org/val2017/000000039769.jpg",
+    "https://www.ilankelman.org/stopsigns/australia.jpg",
+    "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/snowman.jpg"
+]
+
+messages = [
+    [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What’s the difference between"},
+                {"type": "image", "url": image_urls[0]},
+                {"type": "text", "text": " and "},
+                {"type": "image", "url": image_urls[1]}
+            ]
+        }
+    ],
+    [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image", "url": image_urls[2]},
+                {"type": "text", "text": "What do you see in this image?"}
+            ]
+        }
+    ]
+]
+
+# Load model and processor
+processor = JanusProcessor.from_pretrained(model_id)
+model = JanusForConditionalGeneration.from_pretrained(
+    model_id, torch_dtype=torch.bfloat16, device_map="auto"
+)
+
+inputs = processor.apply_chat_template(
+    messages,
+    add_generation_prompt=True,
+    generation_mode="text",
+    tokenize=True,
+    padding=True,
+    return_dict=True,
+    return_tensors="pt"
+).to(model.device, dtype=torch.bfloat16)
+
+# Generate response
+output = model.generate(**inputs, max_new_tokens=40, generation_mode='text', do_sample=False)
+text = processor.batch_decode(output, skip_special_tokens=True)
+print(text)
+```
+
+## Text to Image generation
+
+Janus can also generate images given a prompt.
+
+```python
+import torch
+from transformers import JanusForConditionalGeneration, JanusProcessor
+
+# Set generation mode to `image` to prepare inputs for image generation..
+
+model_id = "deepseek-community/Janus-Pro-1B"
+processor = JanusProcessor.from_pretrained(model_id)
+model = JanusForConditionalGeneration.from_pretrained(model_id,
+        torch_dtype=torch.bfloat16,
+        device_map="auto")
+
+messages = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "A dog running under the rain."},
+        ],
+     }
+]
+
+prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
+inputs = processor(text=prompt,generation_mode="image",return_tensors="pt").to(model.device, dtype=torch.bfloat16)
+
+# Set num_return_sequence parameter to generate multiple images per prompt.
+model.generation_config.num_return_sequences = 2
+outputs = model.generate(**inputs,
+                         generation_mode="image",
+                         do_sample=True,
+                         use_cache=True,
+                         )
+# Perform post-processing on the generated token ids.
+decoded_image = model.decode_image_tokens(outputs)
+images = processor.postprocess(list(decoded_image.float()),return_tensors="PIL.Image.Image")
+# Save the image
+for i, image in enumerate(images['pixel_values']):
+    image.save(f"result{i}.png")
+```
+
+## JanusConfig
+
+[[autodoc]] JanusConfig
+
+## JanusVisionConfig
+
+[[autodoc]] JanusVisionConfig
+
+## JanusVQVAEConfig
+
+[[autodoc]] JanusVQVAEConfig
+
+## JanusProcessor
+
+[[autodoc]] JanusProcessor
+
+## JanusImageProcessor
+
+[[autodoc]] JanusImageProcessor
+
+## JanusVisionModel
+
+[[autodoc]] JanusVisionModel
+    - forward
+
+## JanusVQVAE
+
+[[autodoc]] JanusVQVAE
+    - forward
+
+## JanusModel
+
+[[autodoc]] JanusModel
+    - forward
+
+## JanusForConditionalGeneration
+
+[[autodoc]] JanusForConditionalGeneration
+    - forward

docs/source/en/model_doc/longformer.md

@@ -1,5 +1,4 @@
-<!--Copyright 2020 The HuggingFace Team. All rights reserved.
-
+<!--Copyright 2024 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
 
@@ -9,93 +8,95 @@ Unless required by applicable law or agreed to in writing, software distributed
 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
+⚠️ Note that this file is in Markdown but contains specific syntax for our doc-builder (similar to MDX) that may not be
 rendered properly in your Markdown viewer.
-
 -->
 
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
+    </div>
+</div>
+
 # Longformer
 
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
-</div>
+[Longformer](https://huggingface.co/papers/2004.05150) is a transformer model designed for processing long documents. The self-attention operation usually scales quadratically with sequence length, preventing transformers from processing longer sequences. The Longformer attention mechanism overcomes this by scaling linearly with sequence length. It combines local windowed attention with task-specific global attention, enabling efficient processing of documents with thousands of tokens.
 
-## Overview
+You can find all the original Longformer checkpoints under the [Ai2](https://huggingface.co/allenai?search_models=longformer) organization.
 
-The Longformer model was presented in [Longformer: The Long-Document Transformer](https://arxiv.org/pdf/2004.05150.pdf) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
+> [!TIP]
+> Click on the Longformer models in the right sidebar for more examples of how to apply Longformer to different language tasks.
 
-The abstract from the paper is the following:
+The example below demonstrates how to fill the `<mask>` token with [`Pipeline`], [`AutoModel`] and from the command line.
 
-*Transformer-based models are unable to process long sequences due to their self-attention operation, which scales
-quadratically with the sequence length. To address this limitation, we introduce the Longformer with an attention
-mechanism that scales linearly with sequence length, making it easy to process documents of thousands of tokens or
-longer. Longformer's attention mechanism is a drop-in replacement for the standard self-attention and combines a local
-windowed attention with a task motivated global attention. Following prior work on long-sequence transformers, we
-evaluate Longformer on character-level language modeling and achieve state-of-the-art results on text8 and enwik8. In
-contrast to most prior work, we also pretrain Longformer and finetune it on a variety of downstream tasks. Our
-pretrained Longformer consistently outperforms RoBERTa on long document tasks and sets new state-of-the-art results on
-WikiHop and TriviaQA.*
-
-This model was contributed by [beltagy](https://huggingface.co/beltagy). The Authors' code can be found [here](https://github.com/allenai/longformer).
-
-## Usage tips
-
-- Since the Longformer is based on RoBERTa, it doesn't have `token_type_ids`. You don't need to indicate which
-  token belongs to which segment. Just separate your segments with the separation token `tokenizer.sep_token` (or
-  `</s>`).
-- A transformer model replacing the attention matrices by sparse matrices to go faster. Often, the local context (e.g., what are the two tokens left and right?) is enough to take action for a given token. Some preselected input tokens are still given global attention, but the attention matrix has way less parameters, resulting in a speed-up. See the local attention section for more information.
-
-## Longformer Self Attention
-
-Longformer self attention employs self attention on both a "local" context and a "global" context. Most tokens only
-attend "locally" to each other meaning that each token attends to its \\(\frac{1}{2} w\\) previous tokens and
-\\(\frac{1}{2} w\\) succeeding tokens with \\(w\\) being the window length as defined in
-`config.attention_window`. Note that `config.attention_window` can be of type `List` to define a
-different \\(w\\) for each layer. A selected few tokens attend "globally" to all other tokens, as it is
-conventionally done for all tokens in `BertSelfAttention`.
-
-Note that "locally" and "globally" attending tokens are projected by different query, key and value matrices. Also note
-that every "locally" attending token not only attends to tokens within its window \\(w\\), but also to all "globally"
-attending tokens so that global attention is *symmetric*.
-
-The user can define which tokens attend "locally" and which tokens attend "globally" by setting the tensor
-`global_attention_mask` at run-time appropriately. All Longformer models employ the following logic for
-`global_attention_mask`:
-
-- 0: the token attends "locally",
-- 1: the token attends "globally".
-
-For more information please also refer to [`~LongformerModel.forward`] method.
-
-Using Longformer self attention, the memory and time complexity of the query-key matmul operation, which usually
-represents the memory and time bottleneck, can be reduced from \\(\mathcal{O}(n_s \times n_s)\\) to
-\\(\mathcal{O}(n_s \times w)\\), with \\(n_s\\) being the sequence length and \\(w\\) being the average window
-size. It is assumed that the number of "globally" attending tokens is insignificant as compared to the number of
-"locally" attending tokens.
-
-For more information, please refer to the official [paper](https://arxiv.org/pdf/2004.05150.pdf).
-
-
-## Training
-
-[`LongformerForMaskedLM`] is trained the exact same way [`RobertaForMaskedLM`] is
-trained and should be used as follows:
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
 ```python
-input_ids = tokenizer.encode("This is a sentence from [MASK] training data", return_tensors="pt")
-mlm_labels = tokenizer.encode("This is a sentence from the training data", return_tensors="pt")
+import torch
+from transformers import pipeline
 
-loss = model(input_ids, labels=input_ids, masked_lm_labels=mlm_labels)[0]
+pipeline = pipeline(
+    task="fill-mask",
+    model="allenai/longformer-base-4096",
+    torch_dtype=torch.float16,
+    device=0
+)
+pipeline("""San Francisco 49ers cornerback Shawntae Spencer will miss the rest of the <mask> with a torn ligament in his left knee.
+Spencer, a fifth-year pro, will be placed on injured reserve soon after undergoing surgery Wednesday to repair the ligament. He injured his knee late in the 49ers’ road victory at Seattle on Sept. 14, and missed last week’s victory over Detroit.
+Tarell Brown and Donald Strickland will compete to replace Spencer with the 49ers, who kept 12 defensive backs on their 53-man roster to start the season. Brown, a second-year pro, got his first career interception last weekend while filling in for Strickland, who also sat out with a knee injury.""")
 ```
 
-## Resources
+</hfoption>
+<hfoption id="AutoModel">
 
-- [Text classification task guide](../tasks/sequence_classification)
-- [Token classification task guide](../tasks/token_classification)
-- [Question answering task guide](../tasks/question_answering)
-- [Masked language modeling task guide](../tasks/masked_language_modeling)
-- [Multiple choice task guide](../tasks/multiple_choice)
+```python
+import torch
+from transformers import AutoModelForMaskedLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("allenai/longformer-base-4096")
+model = AutoModelForMaskedLM.from_pretrained("allenai/longformer-base-4096")
+
+text = (
+"""
+San Francisco 49ers cornerback Shawntae Spencer will miss the rest of the <mask> with a torn ligament in his left knee.
+Spencer, a fifth-year pro, will be placed on injured reserve soon after undergoing surgery Wednesday to repair the ligament. He injured his knee late in the 49ers’ road victory at Seattle on Sept. 14, and missed last week’s victory over Detroit.
+Tarell Brown and Donald Strickland will compete to replace Spencer with the 49ers, who kept 12 defensive backs on their 53-man roster to start the season. Brown, a second-year pro, got his first career interception last weekend while filling in for Strickland, who also sat out with a knee injury.
+"""
+)
+
+input_ids = tokenizer([text], return_tensors="pt")["input_ids"]
+logits = model(input_ids).logits
+
+masked_index = (input_ids[0] == tokenizer.mask_token_id).nonzero().item()
+probs = logits[0, masked_index].softmax(dim=0)
+values, predictions = probs.topk(5)
+tokenizer.decode(predictions).split()
+```
+
+</hfoption>
+<hfoption id="transformers-cli">
+
+```bash
+echo -e "San Francisco 49ers cornerback Shawntae Spencer will miss the rest of the <mask> with a torn ligament in his left knee." | transformers-cli run --task fill-mask --model allenai/longformer-base-4096 --device 0
+```
+
+</hfoption>
+</hfoptions
+
+
+## Notes
+
+- Longformer is based on [RoBERTa](https://huggingface.co/docs/transformers/en/model_doc/roberta) and doesn't have `token_type_ids`. You don't need to indicate which token belongs to which segment. You only need to separate the segments with the separation token `</s>` or `tokenizer.sep_token`.
+- You can set which tokens can attend locally and which tokens attend globally with the `global_attention_mask` at inference (see this [example](https://huggingface.co/docs/transformers/en/model_doc/longformer#transformers.LongformerModel.forward.example) for more details). A value of `0` means a token attends locally and a value of `1` means a token attends globally.
+- [`LongformerForMaskedLM`] is trained like [`RobertaForMaskedLM`] and should be used as shown below.
+
+  ```py
+    input_ids = tokenizer.encode("This is a sentence from [MASK] training data", return_tensors="pt")
+    mlm_labels = tokenizer.encode("This is a sentence from the training data", return_tensors="pt")
+    loss = model(input_ids, labels=input_ids, masked_lm_labels=mlm_labels)[0]
+    ```
 
 ## LongformerConfig
 
@@ -139,9 +140,6 @@ loss = model(input_ids, labels=input_ids, masked_lm_labels=mlm_labels)[0]
 
 [[autodoc]] models.longformer.modeling_tf_longformer.TFLongformerTokenClassifierOutput
 
-<frameworkcontent>
-<pt>
-
 ## LongformerModel
 
 [[autodoc]] LongformerModel
@@ -149,45 +147,42 @@ loss = model(input_ids, labels=input_ids, masked_lm_labels=mlm_labels)[0]
 
 ## LongformerForMaskedLM
 
-[[autodoc]] LongformerForMaskedLM
+[[autodoc]] LongformerForMaskedLM 
     - forward
 
 ## LongformerForSequenceClassification
 
-[[autodoc]] LongformerForSequenceClassification
+[[autodoc]] LongformerForSequenceClassification 
     - forward
 
 ## LongformerForMultipleChoice
 
-[[autodoc]] LongformerForMultipleChoice
+[[autodoc]] LongformerForMultipleChoice 
     - forward
 
 ## LongformerForTokenClassification
 
-[[autodoc]] LongformerForTokenClassification
+[[autodoc]] LongformerForTokenClassification 
     - forward
 
 ## LongformerForQuestionAnswering
 
-[[autodoc]] LongformerForQuestionAnswering
+[[autodoc]] LongformerForQuestionAnswering 
     - forward
 
-</pt>
-<tf>
-
 ## TFLongformerModel
 
-[[autodoc]] TFLongformerModel
+[[autodoc]] TFLongformerModel    
     - call
 
 ## TFLongformerForMaskedLM
 
-[[autodoc]] TFLongformerForMaskedLM
+[[autodoc]] TFLongformerForMaskedLM 
     - call
 
 ## TFLongformerForQuestionAnswering
 
-[[autodoc]] TFLongformerForQuestionAnswering
+[[autodoc]] TFLongformerForQuestionAnswering 
     - call
 
 ## TFLongformerForSequenceClassification
@@ -197,13 +192,10 @@ loss = model(input_ids, labels=input_ids, masked_lm_labels=mlm_labels)[0]
 
 ## TFLongformerForTokenClassification
 
-[[autodoc]] TFLongformerForTokenClassification
+[[autodoc]] TFLongformerForTokenClassification 
     - call
 
 ## TFLongformerForMultipleChoice
 
-[[autodoc]] TFLongformerForMultipleChoice
+[[autodoc]] TFLongformerForMultipleChoice 
     - call
-
-</tf>
-</frameworkcontent>

docs/source/en/model_doc/mbart.md

@@ -14,154 +14,105 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# MBart and MBart-50
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
-<img alt="Flax" src="https://img.shields.io/badge/Flax-29a79b.svg?style=flat&logo=data:image/png;base64,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
-">
-<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
-<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+<div style="float: right;">
+  <div class="flex flex-wrap space-x-1">
+    <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+    <img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
+    <img alt="Flax" src="https://img.shields.io/badge/Flax-29a79b.svg?style=flat">
+    <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+    <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+  </div>
 </div>
 
+# mBART
 
-## Overview of MBart
+[mBART](https://huggingface.co/papers/2001.08210) is a multilingual machine translation model that pretrains the entire translation model (encoder-decoder) unlike previous methods that only focused on parts of the model. The model is trained on a denoising objective which reconstructs the corrupted text. This allows mBART to handle the source language and the target text to translate to.
 
-The MBart model was presented in [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov Marjan
-Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
+[mBART-50](https://huggingface.co/paper/2008.00401) is pretrained on an additional 25 languages.
 
-According to the abstract, MBART is a sequence-to-sequence denoising auto-encoder pretrained on large-scale monolingual
-corpora in many languages using the BART objective. mBART is one of the first methods for pretraining a complete
-sequence-to-sequence model by denoising full texts in multiple languages, while previous approaches have focused only
-on the encoder, decoder, or reconstructing parts of the text.
+You can find all the original mBART checkpoints under the [AI at Meta](https://huggingface.co/facebook?search_models=mbart) organization.
 
-This model was contributed by [valhalla](https://huggingface.co/valhalla). The Authors' code can be found [here](https://github.com/pytorch/fairseq/tree/master/examples/mbart)
+> [!TIP]
+> Click on the mBART models in the right sidebar for more examples of applying mBART to different language tasks.
 
-### Training of MBart
+The example below demonstrates how to translate text with [`Pipeline`] or the [`AutoModel`] class.
 
-MBart is a multilingual encoder-decoder (sequence-to-sequence) model primarily intended for translation task. As the
-model is multilingual it expects the sequences in a different format. A special language id token is added in both the
-source and target text. The source text format is `X [eos, src_lang_code]` where `X` is the source text. The
-target text format is `[tgt_lang_code] X [eos]`. `bos` is never used.
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
-The regular [`~MBartTokenizer.__call__`] will encode source text format passed as first argument or with the `text`
-keyword, and target text format passed with the `text_label` keyword argument.
+```py
+import torch
+from transformers import pipeline
 
-- Supervised training
-
-```python
->>> from transformers import MBartForConditionalGeneration, MBartTokenizer
-
->>> tokenizer = MBartTokenizer.from_pretrained("facebook/mbart-large-en-ro", src_lang="en_XX", tgt_lang="ro_RO")
->>> example_english_phrase = "UN Chief Says There Is No Military Solution in Syria"
->>> expected_translation_romanian = "Şeful ONU declară că nu există o soluţie militară în Siria"
-
->>> inputs = tokenizer(example_english_phrase, text_target=expected_translation_romanian, return_tensors="pt")
-
->>> model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-en-ro")
->>> # forward pass
->>> model(**inputs)
+pipeline = pipeline(
+    task="translation",
+    model="facebook/mbart-large-50-many-to-many-mmt",
+    device=0,
+    torch_dtype=torch.float16,
+    src_lang="en_XX",
+    tgt_lang="fr_XX",
+)
+print(pipeline("UN Chief Says There Is No Military Solution in Syria"))
 ```
 
-- Generation
+</hfoption>
+<hfoption id="AutoModel">
 
-  While generating the target text set the `decoder_start_token_id` to the target language id. The following
-  example shows how to translate English to Romanian using the *facebook/mbart-large-en-ro* model.
+```py
+import torch
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
 
-```python
->>> from transformers import MBartForConditionalGeneration, MBartTokenizer
+article_en = "UN Chief Says There Is No Military Solution in Syria"
 
->>> tokenizer = MBartTokenizer.from_pretrained("facebook/mbart-large-en-ro", src_lang="en_XX")
->>> article = "UN Chief Says There Is No Military Solution in Syria"
->>> inputs = tokenizer(article, return_tensors="pt")
->>> translated_tokens = model.generate(**inputs, decoder_start_token_id=tokenizer.lang_code_to_id["ro_RO"])
->>> tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
-"Şeful ONU declară că nu există o soluţie militară în Siria"
+model = AutoModelForSeq2SeqLM.from_pretrained("facebook/mbart-large-50-many-to-many-mmt", torch_dtype=torch.bfloat16, attn_implementation="sdpa", device_map="auto")
+tokenizer = AutoTokenizer.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
+
+tokenizer.src_lang = "en_XX"
+encoded_hi = tokenizer(article_en, return_tensors="pt").to("cuda")
+generated_tokens = model.generate(**encoded_hi, forced_bos_token_id=tokenizer.lang_code_to_id["fr_XX"], cache_implementation="static")
+print(tokenizer.batch_decode(generated_tokens, skip_special_tokens=True))
 ```
 
-## Overview of MBart-50
+</hfoption>
+</hfoptions>
 
-MBart-50 was introduced in the [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) paper by Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav
-Chaudhary, Jiatao Gu, Angela Fan. MBart-50 is created using the original *mbart-large-cc25* checkpoint by extending
-its embedding layers with randomly initialized vectors for an extra set of 25 language tokens and then pretrained on 50
-languages.
+## Notes
 
-According to the abstract
+- You can check the full list of language codes via `tokenizer.lang_code_to_id.keys()`.
+- mBART requires a special language id token in the source and target text during training. The source text format is `X [eos, src_lang_code]` where `X` is the source text. The target text format is `[tgt_lang_code] X [eos]`. The `bos` token is never used. The [`~PreTrainedTokenizerBase._call_`] encodes the source text format passed as the first argument or with the `text` keyword. The target text format is passed with the `text_label` keyword.
+- Set the `decoder_start_token_id` to the target language id for mBART.
 
-*Multilingual translation models can be created through multilingual finetuning. Instead of finetuning on one
-direction, a pretrained model is finetuned on many directions at the same time. It demonstrates that pretrained models
-can be extended to incorporate additional languages without loss of performance. Multilingual finetuning improves on
-average 1 BLEU over the strongest baselines (being either multilingual from scratch or bilingual finetuning) while
-improving 9.3 BLEU on average over bilingual baselines from scratch.*
+    ```py
+    import torch
+    from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
 
+    model = AutoModelForSeq2SeqLM.from_pretrained("facebook/mbart-large-en-ro", torch_dtype=torch.bfloat16, attn_implementation="sdpa", device_map="auto")
+    tokenizer = MBartTokenizer.from_pretrained("facebook/mbart-large-en-ro", src_lang="en_XX")
 
-### Training of MBart-50
+    article = "UN Chief Says There Is No Military Solution in Syria"
+    inputs = tokenizer(article, return_tensors="pt")
 
-The text format for MBart-50 is slightly different from mBART. For MBart-50 the language id token is used as a prefix
-for both source and target text i.e the text format is `[lang_code] X [eos]`, where `lang_code` is source
-language id for source text and target language id for target text, with `X` being the source or target text
-respectively.
+    translated_tokens = model.generate(**inputs, decoder_start_token_id=tokenizer.lang_code_to_id["ro_RO"])
+    tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
+    ```
 
+- mBART-50 has a different text format. The language id token is used as the prefix for the source and target text. The text format is `[lang_code] X [eos]` where `lang_code` is the source language id for the source text and target language id for the target text. `X` is the source or target text respectively.
+- Set the `eos_token_id` as the `decoder_start_token_id` for mBART-50. The target language id is used as the first generated token by passing `forced_bos_token_id` to [`~GenerationMixin.generate`].
 
-MBart-50 has its own tokenizer [`MBart50Tokenizer`].
+    ```py
+    import torch
+    from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
 
--  Supervised training
+    model = AutoModelForSeq2SeqLM.from_pretrained("facebook/mbart-large-50-many-to-many-mmt", torch_dtype=torch.bfloat16, attn_implementation="sdpa", device_map="auto")
+    tokenizer = MBartTokenizer.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
 
-```python
-from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
+    article_ar = "الأمين العام للأمم المتحدة يقول إنه لا يوجد حل عسكري في سوريا."
+    tokenizer.src_lang = "ar_AR"
 
-model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50")
-tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50", src_lang="en_XX", tgt_lang="ro_RO")
-
-src_text = " UN Chief Says There Is No Military Solution in Syria"
-tgt_text = "Şeful ONU declară că nu există o soluţie militară în Siria"
-
-model_inputs = tokenizer(src_text, text_target=tgt_text, return_tensors="pt")
-
-model(**model_inputs)  # forward pass
-```
-
-- Generation
-
-  To generate using the mBART-50 multilingual translation models, `eos_token_id` is used as the
-  `decoder_start_token_id` and the target language id is forced as the first generated token. To force the
-  target language id as the first generated token, pass the *forced_bos_token_id* parameter to the *generate* method.
-  The following example shows how to translate between Hindi to French and Arabic to English using the
-  *facebook/mbart-50-large-many-to-many* checkpoint.
-
-```python
-from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
-
-article_hi = "संयुक्त राष्ट्र के प्रमुख का कहना है कि सीरिया में कोई सैन्य समाधान नहीं है"
-article_ar = "الأمين العام للأمم المتحدة يقول إنه لا يوجد حل عسكري في سوريا."
-
-model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
-tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
-
-# translate Hindi to French
-tokenizer.src_lang = "hi_IN"
-encoded_hi = tokenizer(article_hi, return_tensors="pt")
-generated_tokens = model.generate(**encoded_hi, forced_bos_token_id=tokenizer.lang_code_to_id["fr_XX"])
-tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
-# => "Le chef de l 'ONU affirme qu 'il n 'y a pas de solution militaire en Syria."
-
-# translate Arabic to English
-tokenizer.src_lang = "ar_AR"
-encoded_ar = tokenizer(article_ar, return_tensors="pt")
-generated_tokens = model.generate(**encoded_ar, forced_bos_token_id=tokenizer.lang_code_to_id["en_XX"])
-tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
-# => "The Secretary-General of the United Nations says there is no military solution in Syria."
-```
-
-## Documentation resources
-
-- [Text classification task guide](../tasks/sequence_classification)
-- [Question answering task guide](../tasks/question_answering)
-- [Causal language modeling task guide](../tasks/language_modeling)
-- [Masked language modeling task guide](../tasks/masked_language_modeling)
-- [Translation task guide](../tasks/translation)
-- [Summarization task guide](../tasks/summarization)
+    encoded_ar = tokenizer(article_ar, return_tensors="pt")
+    generated_tokens = model.generate(**encoded_ar, forced_bos_token_id=tokenizer.lang_code_to_id["en_XX"])
+    tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
+    ```
 
 ## MBartConfig
 
@@ -253,4 +204,4 @@ tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
     - decode
 
 </jax>
-</frameworkcontent>
+</frameworkcontent>

docs/source/en/model_doc/mlcd.md

@@ -0,0 +1,81 @@
+<!--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.
+
+-->
+
+# MLCD
+
+<div class="flex flex-wrap space-x-1">
+<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+</div>
+
+## Overview
+
+The MLCD models were released by the DeepGlint-AI team in [unicom](https://github.com/deepglint/unicom), which focuses on building foundational visual models for large multimodal language models using large-scale datasets such as LAION400M and COYO700M, and employs sample-to-cluster contrastive learning to optimize performance. MLCD models are primarily used for multimodal visual large language models, such as LLaVA.
+
+🔥**MLCD-ViT-bigG**🔥 series is the state-of-the-art vision transformer model enhanced with 2D Rotary Position Embedding (RoPE2D), achieving superior performance on document understanding and visual question answering tasks. Developed by DeepGlint AI, this model demonstrates exceptional capabilities in processing complex visual-language interactions.
+
+Tips:
+
+- We adopted the official [LLaVA-NeXT](https://github.com/LLaVA-VL/LLaVA-NeXT) and the official training dataset [LLaVA-NeXT-Data](https://huggingface.co/datasets/lmms-lab/LLaVA-NeXT-Data) for evaluating the foundational visual models.
+
+- The language model is [Qwen2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct). 
+
+Result: 
+
+| Vision Tower                                                                                  | RoPE2D | ChartQA   | DocVQA    | InfoVQA   | OCRBench   | MMMU      |
+| :-------------------------------------------------------------------------------------------- | :----: | :-------- | :-------- | :-------- | :--------- | :-------- |
+| CLIP (ViT-L-14-336px)                                                                         |   ×    | 66.52     | 75.21     | 38.88     | 525.00     | 44.20     |
+| SigLIP (ViT-SO400M-384px)                                                                     |   ×    | 69.28     | 76.71     | 41.38     | 554.00     | 46.78     |
+| DFN5B (ViT-H-14-378px)                                                                        |   ×    | 64.36     | 70.87     | 38.59     | 473.00     | **48.00** |
+| **[MLCD (ViT-L-14-336px)](https://huggingface.co/DeepGlint-AI/mlcd-vit-large-patch14-336)**   |   ×    | 67.84     | 76.46     | 43.48     | 531.00     | 44.30     |
+| **[MLCD (ViT-bigG-14-336px)](https://huggingface.co/DeepGlint-AI/mlcd-vit-bigG-patch14-336)** |   √    | 71.07     | 79.63     | 44.38     | 572.00     | 46.78     |
+| **[MLCD (ViT-bigG-14-448px)](https://huggingface.co/DeepGlint-AI/mlcd-vit-bigG-patch14-448)** |   √    | **73.80** | **83.34** | **46.59** | **582.00** | 46.00     |
+
+
+## Usage
+
+```python
+import requests
+from PIL import Image
+from transformers import AutoProcessor, MLCDVisionModel
+
+# Load model and processor
+model = MLCDVisionModel.from_pretrained("DeepGlint-AI/mlcd-vit-bigG-patch14-448")
+processor = AutoProcessor.from_pretrained("DeepGlint-AI/mlcd-vit-bigG-patch14-448")
+
+# Process single image
+url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+image = Image.open(requests.get(url, stream=True).raw)
+inputs = processor(images=image, return_tensors="pt")
+
+# Generate outputs
+with torch.no_grad():
+    outputs = model(**inputs)
+
+# Get visual features
+features = outputs.last_hidden_state
+
+print(f"Extracted features shape: {features.shape}")
+```
+
+## MLCDVisionConfig
+
+[[autodoc]] MLCDVisionConfig
+
+## MLCDVisionModel
+
+[[autodoc]] MLCDVisionModel
+    - forward

docs/source/en/model_doc/mobilenet_v1.md

@@ -77,6 +77,11 @@ If you're interested in submitting a resource to be included here, please feel f
 [[autodoc]] MobileNetV1ImageProcessor
     - preprocess
 
+## MobileNetV1ImageProcessorFast
+
+[[autodoc]] MobileNetV1ImageProcessorFast
+    - preprocess
+
 ## MobileNetV1Model
 
 [[autodoc]] MobileNetV1Model

docs/source/en/model_doc/phi.md

@@ -13,166 +13,117 @@ specific language governing permissions and limitations under the License.
 rendered properly in your Markdown viewer.
 
 -->
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+        <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+        <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
+</div>
 
 # Phi
 
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
-<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
-</div>
+[Phi](https://huggingface.co/papers/2306.11644) is a 1.3B parameter transformer model optimized for Python code generation. It focuses on "textbook-quality" training data of code examples, exercises and synthetic Python problems rather than scaling the model size or compute.
 
-## Overview
+You can find all the original Phi checkpoints under the [Phi-1](https://huggingface.co/collections/microsoft/phi-1-6626e29134744e94e222d572) collection.
 
-The Phi-1 model was proposed in [Textbooks Are All You Need](https://arxiv.org/abs/2306.11644) by Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tauman Kalai, Yin Tat Lee and Yuanzhi Li.
+> [!TIP]
+> Click on the Phi models in the right sidebar for more examples of how to apply Phi to different language tasks.
 
-The Phi-1.5 model was proposed in [Textbooks Are All You Need II: phi-1.5 technical report](https://arxiv.org/abs/2309.05463) by Yuanzhi Li, Sébastien Bubeck, Ronen Eldan, Allie Del Giorno, Suriya Gunasekar and Yin Tat Lee.
+The example below demonstrates how to generate text with [`Pipeline`], [`AutoModel`] and from the command line.
 
-### Summary
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
-In Phi-1 and Phi-1.5 papers, the authors showed how important the quality of the data is in training relative to the model size.
-They selected high quality "textbook" data alongside with synthetically generated data for training their small sized Transformer
-based model Phi-1 with 1.3B parameters. Despite this small scale, phi-1 attains pass@1 accuracy 50.6% on HumanEval and 55.5% on MBPP.
-They follow the same strategy for Phi-1.5 and created another 1.3B parameter model with performance on natural language tasks comparable
-to models 5x larger, and surpassing most non-frontier LLMs. Phi-1.5 exhibits many of the traits of much larger LLMs such as the ability
-to “think step by step” or perform some rudimentary in-context learning.
-With these two experiments the authors successfully showed the huge impact of quality of training data when training machine learning models.
+```py
+import torch
+from transformers import pipeline
 
-The abstract from the Phi-1 paper is the following:
+pipeline = pipeline(task="text-generation", model="microsoft/phi-1.5", device=0, torch_dtype=torch.bfloat16)
+pipeline("pipeline('''def print_prime(n): """ Print all primes between 1 and n"""''')")
 
-*We introduce phi-1, a new large language model for code, with significantly smaller size than
-competing models: phi-1 is a Transformer-based model with 1.3B parameters, trained for 4 days on
-8 A100s, using a selection of “textbook quality” data from the web (6B tokens) and synthetically
-generated textbooks and exercises with GPT-3.5 (1B tokens). Despite this small scale, phi-1 attains
-pass@1 accuracy 50.6% on HumanEval and 55.5% on MBPP. It also displays surprising emergent
-properties compared to phi-1-base, our model before our finetuning stage on a dataset of coding
-exercises, and phi-1-small, a smaller model with 350M parameters trained with the same pipeline as
-phi-1 that still achieves 45% on HumanEval.*
-
-The abstract from the Phi-1.5 paper is the following:
-
-*We continue the investigation into the power of smaller Transformer-based language models as
-initiated by TinyStories – a 10 million parameter model that can produce coherent English – and
-the follow-up work on phi-1, a 1.3 billion parameter model with Python coding performance close
-to the state-of-the-art. The latter work proposed to use existing Large Language Models (LLMs) to
-generate “textbook quality” data as a way to enhance the learning process compared to traditional
-web data. We follow the “Textbooks Are All You Need” approach, focusing this time on common
-sense reasoning in natural language, and create a new 1.3 billion parameter model named phi-1.5,
-with performance on natural language tasks comparable to models 5x larger, and surpassing most
-non-frontier LLMs on more complex reasoning tasks such as grade-school mathematics and basic
-coding. More generally, phi-1.5 exhibits many of the traits of much larger LLMs, both good –such
-as the ability to “think step by step” or perform some rudimentary in-context learning– and bad,
-including hallucinations and the potential for toxic and biased generations –encouragingly though, we
-are seeing improvement on that front thanks to the absence of web data. We open-source phi-1.5 to
-promote further research on these urgent topics.*
-
-This model was contributed by [Susnato Dhar](https://huggingface.co/susnato).
-
-The original code for Phi-1, Phi-1.5 and Phi-2 can be found [here](https://huggingface.co/microsoft/phi-1), [here](https://huggingface.co/microsoft/phi-1_5) and [here](https://huggingface.co/microsoft/phi-2), respectively.
-
-## Usage tips
-
-- This model is quite similar to `Llama` with the main difference in [`PhiDecoderLayer`], where they used [`PhiAttention`] and [`PhiMLP`] layers in parallel configuration.
-- The tokenizer used for this model is identical to the [`CodeGenTokenizer`].
-
-## How to use Phi-2
-
-<Tip warning={true}>
-
-Phi-2 has been integrated in the development version (4.37.0.dev) of `transformers`. Until the official version is released through `pip`, ensure that you are doing one of the following:
-
-* When loading the model, ensure that `trust_remote_code=True` is passed as an argument of the `from_pretrained()` function.
-
-* Update your local `transformers` to the development version: `pip uninstall -y transformers && pip install git+https://github.com/huggingface/transformers`. The previous command is an alternative to cloning and installing from the source.
-
-</Tip>
-
-```python
->>> from transformers import AutoModelForCausalLM, AutoTokenizer
-
->>> model = AutoModelForCausalLM.from_pretrained("microsoft/phi-2")
->>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-2")
-
->>> inputs = tokenizer('Can you help me write a formal email to a potential business partner proposing a joint venture?', return_tensors="pt", return_attention_mask=False)
-
->>> outputs = model.generate(**inputs, max_length=30)
->>> text = tokenizer.batch_decode(outputs)[0]
->>> print(text)
-Can you help me write a formal email to a potential business partner proposing a joint venture?
-Input: Company A: ABC Inc.
-Company B
 ```
 
-### Example :
+</hfoption>
 
-```python
->>> from transformers import PhiForCausalLM, AutoTokenizer
+<hfoption id="AutoModel">
 
->>> # define the model and tokenizer.
->>> model = PhiForCausalLM.from_pretrained("microsoft/phi-1_5")
->>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-1_5")
+```py
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
->>> # feel free to change the prompt to your liking.
->>> prompt = "If I were an AI that had just achieved"
+tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-1")
+model = AutoModelForCausalLM.from_pretrained("microsoft/phi-1", torch_dtype=torch.float16, device_map="auto", attn_implementation="sdpa")
 
->>> # apply the tokenizer.
->>> tokens = tokenizer(prompt, return_tensors="pt")
+input_ids = tokenizer('''def print_prime(n):
+   """
+   Print all primes between 1 and n
+   """''', return_tensors="pt").to("cuda")
 
->>> # use the model to generate new tokens.
->>> generated_output = model.generate(**tokens, use_cache=True, max_new_tokens=10)
-
->>> tokenizer.batch_decode(generated_output)[0]
-'If I were an AI that had just achieved a breakthrough in machine learning, I would be thrilled'
+output = model.generate(**input_ids, cache_implementation="static")
+print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
-## Combining Phi and Flash Attention 2
-
-First, make sure to install the latest version of Flash Attention 2 to include the sliding window attention feature.
+</hfoption>
+<hfoption id="transformers-cli">
 
 ```bash
-pip install -U flash-attn --no-build-isolation
+echo -e "'''def print_prime(n): """ Print all primes between 1 and n"""'''" | transformers-cli run --task text-classification --model microsoft/phi-1.5 --device 0
 ```
 
-Make also sure that you have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of flash-attn repository. Make also sure to load your model in half-precision (e.g. `torch.float16``)
+</hfoption>
+</hfoptions>
 
-To load and run a model using Flash Attention 2, refer to the snippet below:
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
 
-```python
->>> import torch
->>> from transformers import PhiForCausalLM, AutoTokenizer
+The example below uses [bitsandbytes](https://huggingface.co/docs/transformers/en/quantization/bitsandbytes) to only quantize the weights to 4-bits.
 
->>> # define the model and tokenizer and push the model and tokens to the GPU.
->>> model = PhiForCausalLM.from_pretrained("microsoft/phi-1_5", torch_dtype=torch.float16, attn_implementation="flash_attention_2").to("cuda")  # doctest: +SKIP
->>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-1_5")
+```py
+import torch
+from transformers import BitsAndBytesConfig, AutoTokenizer, AutoModelForCausalLM
 
->>> # feel free to change the prompt to your liking.
->>> prompt = "If I were an AI that had just achieved"
+bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True)
+tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-1")
+model = AutoModelForCausalLM.from_pretrained("microsoft/phi-1", torch_dtype=torch.float16, device_map="auto", attn_implementation="sdpa", quantization_config=bnb_config)
 
->>> # apply the tokenizer.
->>> tokens = tokenizer(prompt, return_tensors="pt").to("cuda")
+input_ids = tokenizer('''def print_prime(n):
+   """
+   Print all primes between 1 and n
+   """''', return_tensors="pt").to("cuda")
 
->>> # use the model to generate new tokens.
->>> generated_output = model.generate(**tokens, use_cache=True, max_new_tokens=10)  # doctest: +SKIP
-
->>> tokenizer.batch_decode(generated_output)[0]  # doctest: +SKIP
-'If I were an AI that had just achieved a breakthrough in machine learning, I would be thrilled'
+output = model.generate(**input_ids, cache_implementation="static")
+print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
 
-### Expected speedups
+## Notes
 
-Below is an expected speedup diagram that compares pure inference time between the native implementation in transformers using `microsoft/phi-1` checkpoint and the Flash Attention 2 version of the model using a sequence length of 2048.
+- If you're using Transformers < 4.37.0.dev, set `trust_remote_code=True` in [`~AutoModel.from_pretrained`]. Otherwise, make sure you update Transformers to the latest stable version.
 
-<div style="text-align: center">
-<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/phi_1_speedup_plot.jpg">
-</div>
+    ```py
+    import torch
+    from transformers import AutoTokenizer, AutoModelForCausalLM
+    
+    tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-1")
+    model = AutoModelForCausalLM.from_pretrained(
+        "microsoft/phi-1",
+        torch_dtype=torch.float16,
+        device_map="auto",
+        trust_remote_code=True,
+        attn_implementation="sdpa")
+    
+    input_ids = tokenizer('''def print_prime(n):
+       """
+       Print all primes between 1 and n
+       """''', return_tensors="pt").to("cuda")
+    
+    output = model.generate(**input_ids, cache_implementation="static")
+    print(tokenizer.decode(output[0], skip_special_tokens=True))
+    ```
 
 ## PhiConfig
 
 [[autodoc]] PhiConfig
 
-<frameworkcontent>
-<pt>
-
 ## PhiModel
 
 [[autodoc]] PhiModel
@@ -193,6 +144,3 @@ Below is an expected speedup diagram that compares pure inference time between t
 
 [[autodoc]] PhiForTokenClassification
     - forward
-
-</pt>
-</frameworkcontent>

docs/source/en/model_doc/phi4_multimodal.md

@@ -64,7 +64,7 @@ inputs = processor.apply_chat_template(
     tokenize=True,
     return_dict=True,
     return_tensors="pt",
-).to(device, torch.float16)
+).to(device)
 
 # Generate response
 generate_ids = model.generate(
@@ -98,8 +98,7 @@ inputs = processor.apply_chat_template(
     tokenize=True,
     return_dict=True,
     return_tensors="pt",
-    sample_rate=sample_rate,
-).to(device, torch.float16)
+).to(device)
 
 generate_ids = model.generate(
     **inputs,

docs/source/en/model_doc/poolformer.md

@@ -73,6 +73,11 @@ If you're interested in submitting a resource to be included here, please feel f
 [[autodoc]] PoolFormerImageProcessor
     - preprocess
 
+## PoolFormerImageProcessorFast
+
+[[autodoc]] PoolFormerImageProcessorFast
+    - preprocess
+
 ## PoolFormerModel
 
 [[autodoc]] PoolFormerModel

docs/source/en/model_doc/pvt.md

@@ -64,6 +64,11 @@ This model was contributed by [Xrenya](https://huggingface.co/Xrenya). The origi
 [[autodoc]] PvtImageProcessor
     - preprocess
 
+## PvtImageProcessorFast
+
+[[autodoc]] PvtImageProcessorFast
+    - preprocess
+
 ## PvtForImageClassification
 
 [[autodoc]] PvtForImageClassification

docs/source/en/model_doc/qwen2_5_omni.md

@@ -59,7 +59,7 @@ model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
 )
 processor = Qwen2_5OmniProcessor.from_pretrained("Qwen/Qwen2.5-Omni-7B")
 
-conversation = [
+conversations = [
     {
         "role": "system",
         "content": [
@@ -115,7 +115,7 @@ model = Qwen2_5OmniThinkerForConditionalGeneration.from_pretrained(
 )
 processor = Qwen2_5OmniProcessor.from_pretrained("Qwen/Qwen2.5-Omni-7B")
 
-conversation = [
+conversations = [
     {
         "role": "system",
         "content": [

docs/source/en/model_doc/qwen2_5_vl.md

@@ -232,10 +232,15 @@ model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
 
 [[autodoc]] Qwen2_5_VLConfig
 
+## Qwen2_5_VLTextConfig
+
+[[autodoc]] Qwen2_5_VLTextConfig
+
 ## Qwen2_5_VLProcessor
 
 [[autodoc]] Qwen2_5_VLProcessor
 
+
 ## Qwen2_5_VLModel
 
 [[autodoc]] Qwen2_5_VLModel

docs/source/en/model_doc/qwen2_vl.md

@@ -278,6 +278,10 @@ model = Qwen2VLForConditionalGeneration.from_pretrained(
 
 [[autodoc]] Qwen2VLConfig
 
+## Qwen2VLTextConfig
+
+[[autodoc]] Qwen2VLTextConfig
+
 ## Qwen2VLImageProcessor
 
 [[autodoc]] Qwen2VLImageProcessor

docs/source/en/model_doc/siglip.md

@@ -14,184 +14,116 @@ rendered properly in your Markdown viewer.
 
 -->
 
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+            <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+            <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+            <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
+</div>
+
 # SigLIP
 
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
-<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
-</div>
+[SigLIP](https://huggingface.co/papers/2303.15343) is a multimodal image-text model similar to [CLIP](clip). It uses separate image and text encoders to generate representations for both modalities.
 
-## Overview
+Unlike CLIP, SigLIP employs a pairwise sigmoid loss on image-text pairs during training. This training loss eliminates the need for a global view of all pairwise similarities between images and texts within a batch. Consequently, it enables more efficient scaling to larger batch sizes while also delivering superior performance with smaller batch sizes.
 
-The SigLIP model was proposed in [Sigmoid Loss for Language Image Pre-Training](https://arxiv.org/abs/2303.15343) by Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, Lucas Beyer. SigLIP proposes to replace the loss function used in [CLIP](clip) by a simple pairwise sigmoid loss. This results in better performance in terms of zero-shot classification accuracy on ImageNet.
+You can find all the original SigLIP checkpoints under the [SigLIP](https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba) collection.
 
-The abstract from the paper is the following:
 
-*We propose a simple pairwise Sigmoid loss for Language-Image Pre-training (SigLIP). Unlike standard contrastive learning with softmax normalization, the sigmoid loss operates solely on image-text pairs and does not require a global view of the pairwise similarities for normalization. The sigmoid loss simultaneously allows further scaling up the batch size, while also performing better at smaller batch sizes. Combined with Locked-image Tuning, with only four TPUv4 chips, we train a SigLiT model that achieves 84.5% ImageNet zero-shot accuracy in two days. The disentanglement of the batch size from the loss further allows us to study the impact of examples vs pairs and negative to positive ratio. Finally, we push the batch size to the extreme, up to one million, and find that the benefits of growing batch size quickly diminish, with a more reasonable batch size of 32k being sufficient.*
+> [!TIP]
+> Click on the SigLIP models in the right sidebar for more examples of how to apply SigLIP to different image and text tasks.
 
-## Usage tips
+The example below demonstrates how to generate similarity scores between texts and image(s) with [`Pipeline`] or the [`AutoModel`] class.
 
-- Usage of SigLIP is similar to [CLIP](clip). The main difference is the training loss, which does not require a global view of all the pairwise similarities of images and texts within a batch. One needs to apply the sigmoid activation function to the logits, rather than the softmax.
-- Training is supported but does not use `torch.distributed` utilities which may limit the scalability of batch size. However, DDP and FDSP works on single-node multi-gpu setup.
-- When using the standalone [`SiglipTokenizer`] or [`SiglipProcessor`], make sure to pass `padding="max_length"` as that's how the model was trained.
-- To get the same results as the pipeline, a prompt template of "This is a photo of {label}." should be used.
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
-<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/siglip_table.jpeg"
-alt="drawing" width="600"/>
+```py
+import torch
+from transformers import pipeline
 
-<small> SigLIP evaluation results compared to CLIP. Taken from the <a href="https://arxiv.org/abs/2303.15343">original paper</a>.</small>
+image = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
+candidate_labels = ["a Pallas cat", "a lion", "a Siberian tiger"]
 
-This model was contributed by [nielsr](https://huggingface.co/nielsr).
-The original code can be found [here](https://github.com/google-research/big_vision/tree/main).
-
-## Usage example
-
-There are 2 main ways to use SigLIP: either using the pipeline API, which abstracts away all the complexity for you, or by using the `SiglipModel` class yourself.
-
-### Pipeline API
-
-The pipeline allows to use the model in a few lines of code:
-
-```python
->>> from transformers import pipeline
->>> from PIL import Image
->>> import requests
-
->>> # load pipe
->>> image_classifier = pipeline(task="zero-shot-image-classification", model="google/siglip-base-patch16-224")
-
->>> # load image
->>> url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
->>> image = Image.open(requests.get(url, stream=True).raw)
-
->>> # inference
->>> candidate_labels = ["2 cats", "a plane", "a remote"]
->>> outputs = image_classifier(image, candidate_labels=candidate_labels)
->>> outputs = [{"score": round(output["score"], 4), "label": output["label"] } for output in outputs]
->>> print(outputs)
-[{'score': 0.1979, 'label': '2 cats'}, {'score': 0.0, 'label': 'a remote'}, {'score': 0.0, 'label': 'a plane'}]
+pipeline = pipeline(task="zero-shot-image-classification", model="google/siglip-base-patch16-224", device=0, torch_dtype=torch.bfloat16)
+pipeline(image, candidate_labels=candidate_labels)
 ```
 
-### Using the model yourself
+</hfoption>
+<hfoption id="AutoModel">
 
-If you want to do the pre- and postprocessing yourself, here's how to do that:
+```py
+import torch
+import requests
+from PIL import Image
+from transformers import AutoProcessor, AutoModel
 
-```python
->>> from PIL import Image
->>> import requests
->>> from transformers import AutoProcessor, AutoModel
->>> import torch
+model = AutoModel.from_pretrained("google/siglip-base-patch16-224", torch_dtype=torch.float16, device_map="auto", attn_implementation="sdpa")
+processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
 
->>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224")
->>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
+url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
+image = Image.open(requests.get(url, stream=True).raw)
+candidate_labels = ["a Pallas cat", "a lion", "a Siberian tiger"]
+texts = [f'This is a photo of {label}.' for label in candidate_labels]
+inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt").to("cuda")
 
->>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
->>> image = Image.open(requests.get(url, stream=True).raw)
+with torch.no_grad():
+    outputs = model(**inputs)
 
->>> candidate_labels = ["2 cats", "2 dogs"]
-# follows the pipeline prompt template to get same results
->>> texts = [f'This is a photo of {label}.' for label in candidate_labels]
-# important: we pass `padding=max_length` since the model was trained with this
->>> inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt")
-
->>> with torch.no_grad():
-...     outputs = model(**inputs)
-
->>> logits_per_image = outputs.logits_per_image
->>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
->>> print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
-19.8% that image 0 is '2 cats'
+logits_per_image = outputs.logits_per_image
+probs = torch.sigmoid(logits_per_image)
+print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
 ```
 
-## Resources
+</hfoption>
+</hfoptions>
 
-A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with SigLIP.
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
 
-- [Zero-shot image classification task guide](../tasks/zero_shot_image_classification)
-- Demo notebooks for SigLIP can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/SigLIP). 🌎
+The example below uses [bitsandbytes](../quantization/bitsandbytes) to only quantize the weights to int4.
 
-If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
+```py
+import torch
+import requests
+from PIL import Image
+from transformers import AutoProcessor, AutoModel, BitsAndBytesConfig
 
+bnb_config = BitsAndBytesConfig(load_in_4bit=True)
+model = AutoModel.from_pretrained("google/siglip-base-patch16-224", quantization_config=bnb_config, device_map="auto", attn_implementation="sdpa")
+processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
 
-## Combining SigLIP and Flash Attention 2
+url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
+image = Image.open(requests.get(url, stream=True).raw)
+candidate_labels = ["a Pallas cat", "a lion", "a Siberian tiger"]
+texts = [f'This is a photo of {label}.' for label in candidate_labels]
+inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt").to("cuda")
 
-First, make sure to install the latest version of Flash Attention 2.
+with torch.no_grad():
+    outputs = model(**inputs)
 
-```bash
-pip install -U flash-attn --no-build-isolation
+logits_per_image = outputs.logits_per_image
+probs = torch.sigmoid(logits_per_image)
+print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
 ```
+## Notes
 
-Make also sure that you have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of flash-attn repository. Make also sure to load your model in half-precision (e.g. `torch.float16``)
+- Training is supported for DDP and FSDP on single-node multi-GPU setups. However, it does not use [torch.distributed](https://pytorch.org/tutorials/beginner/dist_overview.html) utilities which may limit the scalability of batch size.
+- When using the standalone [`SiglipTokenizer`] or [`SiglipProcessor`], make sure to pass `padding="max_length"` because that is how the model was trained.
+- To get the same results as the [`Pipeline`], a prompt template of `"This is a photo of {label}."` should be passed to the processor.
+- Toggle the `attn_implementation` parameter to either `"sdpa"` or `"flash_attention_2"` to use a more memory-efficient attention.
+    ```py
+    # pip install -U flash-attn --no-build-isolation
 
-To load and run a model using Flash Attention 2, refer to the snippet below:
+    from transformers import SiglipModel
 
-```python
->>> import torch
->>> import requests
->>> from PIL import Image
->>> from transformers import SiglipProcessor, SiglipModel
->>> device = "cuda" # the device to load the model onto
-
->>> model = SiglipModel.from_pretrained(
-...     "google/siglip-so400m-patch14-384",
-...     attn_implementation="flash_attention_2",
-...     torch_dtype=torch.float16,
-...     device_map=device,
-... )
->>> processor = SiglipProcessor.from_pretrained("google/siglip-so400m-patch14-384")
-
->>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
->>> image = Image.open(requests.get(url, stream=True).raw)
-
->>> candidate_labels = ["2 cats", "2 dogs"]
-# follows the pipeline prompt template to get same results
->>> texts = [f'This is a photo of {label}.' for label in candidate_labels]
-# important: we pass `padding=max_length` since the model was trained with this
->>> inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt").to(device)
-
->>> with torch.no_grad():
-...     with torch.autocast(device):
-...         outputs = model(**inputs)
-
->>> logits_per_image = outputs.logits_per_image
->>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
->>> print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
-19.8% that image 0 is '2 cats'
-```
-
-
-## Using Scaled Dot Product Attention (SDPA)
-
-PyTorch includes a native scaled dot-product attention (SDPA) operator as part of `torch.nn.functional`. This function 
-encompasses several implementations that can be applied depending on the inputs and the hardware in use. See the 
-[official documentation](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html) 
-or the [GPU Inference](https://huggingface.co/docs/transformers/main/en/perf_infer_gpu_one#pytorch-scaled-dot-product-attention)
-page for more information.
-
-You may set `attn_implementation="sdpa"` in `from_pretrained()` to explicitly request SDPA to be used. Make sure you have `torch>=2.1.1`.
-
-```python
->>> from transformers import SiglipModel
-
->>> model = SiglipModel.from_pretrained(
-...     "google/siglip-so400m-patch14-384",
-...     attn_implementation="sdpa",
-...     torch_dtype=torch.float16,
-...     device_map=device,
-... )
-```
-
-For the best speedups, we recommend loading the model in half-precision (e.g. `torch.float16` or `torch.bfloat16`).
-
-
-## Expected speedups
-
-Below is an expected speedup diagram that compares inference time between the native implementation in transformers using `google/siglip-so400m-patch14-384` checkpoint in `float16` precision and the Flash Attention 2 / SDPA version of the model using different batch sizes.
-
-<div style="text-align: center">
-<img src="https://i.imgur.com/cWm4rsn.png">
-</div>
+    model = SiglipModel.from_pretrained(
+        "google/siglip-so400m-patch14-384",
+        attn_implementation="flash_attention_2",
+        torch_dtype=torch.float16,
+        device_map=device,
+    )
+    ```
 
 
 ## SiglipConfig

docs/source/en/model_doc/siglip2.md

@@ -14,225 +14,160 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# SigLIP2
-
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
-<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+            <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+            <img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+            <img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
 </div>
 
+# SigLIP2
+
 ## Overview
 
-The SigLIP2 model was proposed in [SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features](https://huggingface.co/papers/2502.14786) by Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin,
-Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier Hénaff, Jeremiah Harmsen,
-Andreas Steiner and Xiaohua Zhai.
+[SigLIP2](https://huggingface.co/papers/2502.14786) is a family of multilingual vision-language encoders that builds on the [SigLIP](./siglip) training recipe. It includes decoder-based pretraining, self-distillation, and masked prediction to improve dense prediction tasks (segmentation, depth estimation, etc.). This model is available in two variants:
 
-The model comes in two variants
+- NaFlex supports different resolutions and maintains the native image aspect ratio
+- FixRes supports fixed resolutions and is backwards compatible with [SigLIP](./siglip)
 
- 1) FixRes - model works with fixed resolution images (backward compatible with SigLIP v1)
- 2) NaFlex - model works with variable image aspect ratios and resolutions (SigLIP2 in `transformers`)
 
-The abstract from the paper is the following:
+You can find all the original SigLIP2 checkpoints under the [SigLIP2](https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107) collection.
 
-*We introduce SigLIP 2, a family of new multilingual vision-language encoders that build on the success
-of the original SigLIP. In this second iteration, we extend the original image-text training objective with
-several prior, independently developed techniques into a unified recipe—this includes decoder-based
-pretraining, self-supervised losses (self-distillation, masked prediction) and online data curation. With
-these changes, SigLIP 2 models outperform their SigLIP counterparts at all model scales in core capabilities, 
-including zero-shot classification (best SigLIP 2 ViT-g/16 achieves 85.0% ImageNet zero-shot
-accuracy), image-text retrieval, and transfer performance when extracting visual representations for
-Vision-Language Models (VLMs). Furthermore, the new training recipe leads to significant improvements 
-on localization and dense prediction tasks. We also train variants which support multiple resolutions 
-and preserve the input’s native aspect ratio. Finally, we train on a more diverse data-mixture that
-includes de-biasing techniques, leading to much better multilingual understanding and improved fair-
-ness. To provide users with the ability to trade-off inference cost with performance, we release model
-checkpoints at four sizes (ViT-B/86M, L/303M, So400m/400M, and g/1B).*
+> [!TIP]
+> Click on the SigLIP2 models in the right sidebar for more examples of how to apply SigLIP2 to different image and text tasks.
 
-## Usage tips
+The example below demonstrates zero-shot classification with [`Pipeline`] or the [`AutoModel`] class.
 
-- Usage of SigLIP2 is similar to [SigLIP](siglip) and [CLIP](clip). The main difference from CLIP is the training loss, which does not require a global view of all the pairwise similarities of images and texts within a batch. One needs to apply the sigmoid activation function to the logits, rather than the softmax.
-- Training is supported but does not use `torch.distributed` utilities which may limit the scalability of batch size. However, DDP and FDSP works on single-node multi-gpu setup.
-- When using the standalone [`GemmaTokenizerFast`] make sure to pass `padding="max_length"` and `max_length=64` as that's how the model was trained.
-- Model was trained with *lowercased* text, make sure you make the same preprocessing for your text labels.
-- To get the same results as the pipeline, a prompt template of "this is a photo of {label}" should be used.
-- The NaFlex variant supports processing images at higher resolutions by adjusting the `max_num_patches` parameter in the `Processor`. The default value is `max_num_patches=256`. Increasing `max_num_patches` to 1024 (4x) will approximately double processed image height and width, while preserving the aspect ratio.
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
-<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/siglip2_metrics_table.png"
-alt="drawing" width="600"/>
+```py
+import torch
+from transformers import pipeline
 
-This model was contributed by [qubvel](https://huggingface.co/qubvel-hf).
-The original code can be found [here](https://github.com/google-research/big_vision/tree/main).
+image = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
+candidate_labels = ["a Pallas cat", "a lion", "a Siberian tiger"]
 
-## Usage example
-
-There are 2 main ways to use SigLIP2: either using the pipeline API, which abstracts away all the complexity for you, or by using the `Siglip2Model` class yourself.
-
-### FixRes variant
-
-**Pipeline API**
-
-The pipeline allows to use the model in a few lines of code:
-
-```python
->>> from transformers import pipeline
->>> from PIL import Image
->>> import requests
-
->>> # load pipe
->>> image_classifier = pipeline(
-...     task="zero-shot-image-classification",
-...     model="google/siglip2-base-patch16-224",
-... )
-
->>> # load image
->>> url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
->>> image = Image.open(requests.get(url, stream=True).raw)
-
->>> # inference
->>> candidate_labels = ["2 cats", "a plane", "a remote"]
->>> outputs = image_classifier(image, candidate_labels=candidate_labels)
->>> outputs = [{"score": round(output["score"], 4), "label": output["label"] } for output in outputs]
->>> print(outputs)
-[{'score': 0.1499, 'label': '2 cats'}, {'score': 0.0008, 'label': 'a remote'}, {'score': 0.0, 'label': 'a plane'}]
+pipeline = pipeline(task="zero-shot-image-classification", model="google/siglip2-base-patch16-224", device=0, torch_dtype=torch.bfloat16)
+pipeline(image, candidate_labels=candidate_labels)
 ```
 
-**Using the model yourself**
+</hfoption>
+<hfoption id="AutoModel (FixRes)">
 
-If you want to do the pre- and postprocessing yourself, here's how to do that:
+```py
+import torch
+import requests
+from PIL import Image
+from transformers import AutoProcessor, AutoModel
 
-```python
->>> from PIL import Image
->>> import requests
->>> from transformers import AutoProcessor, AutoModel
->>> import torch
+model = AutoModel.from_pretrained("google/siglip2-base-patch16-224", torch_dtype=torch.float16, device_map="auto", attn_implementation="sdpa")
+processor = AutoProcessor.from_pretrained("google/siglip2-base-patch16-224")
 
->>> model = AutoModel.from_pretrained("google/siglip2-base-patch16-224")
->>> processor = AutoProcessor.from_pretrained("google/siglip2-base-patch16-224")
+url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
+image = Image.open(requests.get(url, stream=True).raw)
+candidate_labels = ["a Pallas cat", "a lion", "a Siberian tiger"]
 
->>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
->>> image = Image.open(requests.get(url, stream=True).raw)
-
->>> candidate_labels = ["2 cats", "2 dogs"]
 # follows the pipeline prompt template to get same results
->>> texts = [f"This is a photo of {label}." for label in candidate_labels]
+texts = [f'This is a photo of {label}.' for label in candidate_labels]
 
 # IMPORTANT: we pass `padding=max_length` and `max_length=64` since the model was trained with this
->>> inputs = processor(text=texts, images=image, padding="max_length", max_length=64, return_tensors="pt")
+inputs = processor(text=texts, images=image, padding="max_length", max_length=64, return_tensors="pt").to("cuda")
 
->>> with torch.no_grad():
-...     outputs = model(**inputs)
+with torch.no_grad():
+    outputs = model(**inputs)
 
->>> logits_per_image = outputs.logits_per_image
->>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
->>> print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
-15.0% that image 0 is '2 cats'
+logits_per_image = outputs.logits_per_image
+probs = torch.sigmoid(logits_per_image)
+print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
 ```
 
-### NaFlex variant
+</hfoption>
+<hfoption id="AutoModel (NaFlex)">
 
-NaFlex combines ideas from FlexiViT, i.e. supporting multiple, predefined sequence lengths 
-with a single ViT model, and NaViT, namely processing images at their native aspect ratio.
-This enables processing different types of images at appropriate resolution, e.g. using a
-larger resolution to process document images, while at the same time minimizing the impact 
-of aspect ratio distortion on certain inference tasks, e.g. on OCR.
+```py
+import torch
+import requests
+from PIL import Image
+from transformers import AutoProcessor, AutoModel
 
-Given a patch size and target sequence length, NaFlex preprocesses the data by first resizing 
-the input image such that the height and width after resizing are multiples of the patch size,
-while 
-    
-    1. keeping the aspect ratio distortion as small as possible
-    2. producing a sequence length of at most the desired target sequence length (`max_num_patches`)
-    
-The resulting distortion in width and height is at most `(patch_size - 1) / width` and
-`(patch_size - 1) / height`, respectively, which tends to be small for common resolutions and aspect ratios. 
-After resizing, the image is split into a sequence of patches, and a mask with padding information is added.
+model = AutoModel.from_pretrained("google/siglip2-base-patch16-naflex", torch_dtype=torch.float16, device_map="auto", attn_implementation="sdpa")
+processor = AutoProcessor.from_pretrained("google/siglip2-base-patch16-naflex")
 
-```python
->>> from PIL import Image
->>> import requests
->>> from transformers import AutoProcessor, AutoModel
->>> import torch
+url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
+image = Image.open(requests.get(url, stream=True).raw)
+candidate_labels = ["a Pallas cat", "a lion", "a Siberian tiger"]
+texts = [f'This is a photo of {label}.' for label in candidate_labels]
 
->>> model = AutoModel.from_pretrained("google/siglip2-base-patch16-naflex")
->>> processor = AutoProcessor.from_pretrained("google/siglip2-base-patch16-naflex")
+# default value for `max_num_patches` is 256, but you can increase resulted image resolution providing higher values e.g. `max_num_patches=512`
+inputs = processor(text=texts, images=image, padding="max_length", max_num_patches=256, return_tensors="pt").to("cuda")
 
->>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
->>> image = Image.open(requests.get(url, stream=True).raw)
+with torch.no_grad():
+    outputs = model(**inputs)
+
+logits_per_image = outputs.logits_per_image
+probs = torch.sigmoid(logits_per_image)
+print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
+```
+
+</hfoption>
+</hfoptions>
+
+Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
+
+The example below uses [bitsandbytes](../quantization/bitsandbytes) to only quantize the weights to int4.
+
+```py
+import torch
+import requests
+from PIL import Image
+from transformers import AutoProcessor, AutoModel, BitsAndBytesConfig
+
+bnb_config = BitsAndBytesConfig(load_in_4bit=True)
+model = AutoModel.from_pretrained("google/siglip2-large-patch16-512", quantization_config=bnb_config, device_map="auto", attn_implementation="sdpa")
+processor = AutoProcessor.from_pretrained("google/siglip2-base-patch16-224")
+
+url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
+image = Image.open(requests.get(url, stream=True).raw)
+candidate_labels = ["a Pallas cat", "a lion", "a Siberian tiger"]
 
->>> candidate_labels = ["2 cats", "2 dogs"]
 # follows the pipeline prompt template to get same results
->>> texts = [f"This is a photo of {label}." for label in candidate_labels]
+texts = [f'This is a photo of {label}.' for label in candidate_labels]
 
-# default value for `max_num_patches` is 256, but you can increase resulted image resolution providing
-# higher values e.g. `max_num_patches=512`
->>> inputs = processor(text=texts, images=image, max_num_patches=256, return_tensors="pt")
+# IMPORTANT: we pass `padding=max_length` and `max_length=64` since the model was trained with this
+inputs = processor(text=texts, images=image, padding="max_length", max_length=64, return_tensors="pt").to("cuda")
 
->>> with torch.no_grad():
-...     outputs = model(**inputs)
+with torch.no_grad():
+    outputs = model(**inputs)
 
->>> logits_per_image = outputs.logits_per_image
->>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
->>> print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
-21.1% that image 0 is '2 cats'
+logits_per_image = outputs.logits_per_image
+probs = torch.sigmoid(logits_per_image)
+print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
 ```
 
-## Resources
+## Notes
 
-A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with SigLIP2.
+- Training is supported for DDP and FSDP on single-node multi-GPU setups. However, it does not use [torch.distributed](https://pytorch.org/tutorials/beginner/dist_overview.html) utilities which may limit the scalability of batch size.
+- When using the standalone [`GemmaTokenizerFast`] make sure to pass `padding="max_length"` and `max_length=64` as that's how the model was trained.
+- Model was trained with *lowercased* text, so make sure your text labels are preprocessed the same way.
+- To get the same results as the [`Pipeline`], a prompt template of `"This is a photo of {label}."` should be passed to the processor.
+- The NaFlex variant processes different types of images at the appropriate resolution (using a larger resolution to process document images for example), while also minimizing the impact of aspect ratio distortion for certain inference tasks like OCR.
 
-- [Zero-shot image classification task guide](../tasks/zero_shot_image_classification)
-- Demo notebook for SigLIP2 can be found [here](https://github.com/qubvel/transformers-notebooks/tree/master/notebooks/SigLIP2_inference.ipynb). 🌎
+   NaFlex resizes the input image so the height and width are multiples of the patch size after resizing. It keeps the aspect ratio distortion as low as possible and produces a sequence length of at most the desired target sequence length (`max_num_patches`). After resizing, the image is split into a sequence of patches and a mask with padding information is added.
+- Toggle the `attn_implementation` parameter to either `"sdpa"` or `"flash_attention_2"` to use a more memory-efficient attention.
+    ```py
+    # pip install -U flash-attn --no-build-isolation
 
-If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
-
-
-## Combining SigLIP2 and Flash Attention 2
-
-First, make sure to install the latest version of Flash Attention 2.
-
-```bash
-pip install -U flash-attn --no-build-isolation
-```
-
-Make also sure that you have a hardware that is compatible with Flash-Attention 2. Read more about it in the official documentation of flash-attn repository. Make also sure to load your model in half-precision (e.g. `torch.float16``)
-
-To load and run a model using Flash Attention 2, refer to the snippet below:
-
-```python
->>> import torch
->>> import requests
->>> from PIL import Image
->>> from transformers import AutoProcessor, AutoModel
->>> device = "cuda" # the device to load the model onto
-
->>> model = AutoModel.from_pretrained(
-...     "google/siglip2-so400m-patch14-384",
-...     attn_implementation="flash_attention_2",
-...     torch_dtype=torch.float16,
-...     device_map=device,
-... )
->>> processor = AutoProcessor.from_pretrained("google/siglip2-so400m-patch14-384")
-
->>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
->>> image = Image.open(requests.get(url, stream=True).raw)
-
->>> candidate_labels = ["2 cats", "2 dogs"]
-# follows the pipeline prompt template to get same results
->>> texts = [f'This is a photo of {label}.' for label in candidate_labels]
-# important: we pass `padding=max_length` since the model was trained with this
->>> inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt").to(device)
-
->>> with torch.no_grad():
-...     with torch.autocast(device):
-...         outputs = model(**inputs)
-
->>> logits_per_image = outputs.logits_per_image
->>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
->>> print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
-19.8% that image 0 is '2 cats'
-```
+    from transformers import SiglipModel
 
+    model = SiglipModel.from_pretrained(
+        "google/siglip2-so400m-patch14-384",
+        attn_implementation="flash_attention_2",
+        torch_dtype=torch.float16,
+        device_map=device,
+    )
+    ```
 ## Siglip2Config
 
 [[autodoc]] Siglip2Config

docs/source/en/model_doc/timesfm.md

@@ -0,0 +1,88 @@
+<!--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.
+
+-->
+
+# TimesFM
+
+<div class="flex flex-wrap space-x-1">
+<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+</div>
+
+## Overview
+
+TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model proposed in [A decoder-only foundation model for time-series forecasting](https://huggingface.co/papers/2310.10688) by Abhimanyu Das, Weihao Kong, Rajat Sen, and  Yichen Zhou. It is a decoder only model that uses non-overlapping patches of time-series data as input and outputs some output patch length prediction in an autoregressive fashion.
+
+
+The abstract from the paper is the following:
+
+*Motivated by recent advances in large language models for Natural Language Processing (NLP), we design a time-series foundation model for forecasting whose out-of-the-box zero-shot performance on a variety of public datasets comes close to the accuracy of state-of-the-art supervised forecasting models for each individual dataset. Our model is based on pretraining a patched-decoder style attention model on a large time-series corpus, and can work well across different forecasting history lengths, prediction lengths and temporal granularities.*
+
+
+This model was contributed by [kashif](https://huggingface.co/kashif).
+The original code can be found [here](https://github.com/google-research/timesfm).
+
+
+To use the model:
+
+```python
+import torch
+from transformers import TimesFmModelForPrediction
+
+
+model = TimesFmModelForPrediction.from_pretrained(
+    "google/timesfm-2.0-500m-pytorch",
+    torch_dtype=torch.bfloat16,
+    attn_implementation="sdpa",
+    device_map="cuda" if torch.cuda.is_available() else None
+)
+
+
+ # Create dummy inputs
+forecast_input = [
+    np.sin(np.linspace(0, 20, 100)),
+    np.sin(np.linspace(0, 20, 200)),
+    np.sin(np.linspace(0, 20, 400)),
+]
+frequency_input = [0, 1, 2]
+
+# Convert inputs to sequence of tensors
+forecast_input_tensor = [
+    torch.tensor(ts, dtype=torch.bfloat16).to("cuda" if torch.cuda.is_available() else "cpu")
+    for ts in forecast_input
+]
+frequency_input_tensor = torch.tensor(frequency_input, dtype=torch.long).to(
+    "cuda" if torch.cuda.is_available() else "cpu"
+)
+
+# Get predictions from the pre-trained model
+with torch.no_grad():
+    outputs = model(past_values=forecast_input_tensor, freq=frequency_input_tensor, return_dict=True)
+    point_forecast_conv = outputs.mean_predictions.float().cpu().numpy()
+    quantile_forecast_conv = outputs.full_predictions.float().cpu().numpy()
+```
+
+## TimesFmConfig
+
+[[autodoc]] TimesFmConfig
+
+## TimesFmModel
+
+[[autodoc]] TimesFmModel
+    - forward
+
+## TimesFmModelForPrediction
+
+[[autodoc]] TimesFmModelForPrediction
+    - forward

docs/source/en/model_doc/vits.md

@@ -7,168 +7,139 @@ 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.
--->
+specific language governing permissions and limitations under the License.-->
+
+<div style="float: right;">
+    <div class="flex flex-wrap space-x-1">
+        <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+    </div>
+</div>
 
 # VITS
 
-<div class="flex flex-wrap space-x-1">
-<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
-</div>
+[VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech)](https://hf.co/papers/2106.06103) is a end-to-end speech synthesis model, simplifying the traditional two-stage text-to-speech (TTS) systems. It's unique because it directly synthesizes speech from text using variational inference, adversarial learning, and normalizing flows to produce natural and expressive speech with diverse rhythms and intonations.
 
-## Overview
+You can find all the original VITS checkpoints under the [AI at Meta](https://huggingface.co/facebook?search_models=mms-tts) organization.
 
-The VITS model was proposed in [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
+> [!TIP]
+> Click on the VITS models in the right sidebar for more examples of how to apply VITS.
 
-VITS (**V**ariational **I**nference with adversarial learning for end-to-end **T**ext-to-**S**peech) is an end-to-end 
-speech synthesis model that predicts a speech waveform conditional on an input text sequence. It is a conditional variational 
-autoencoder (VAE) comprised of a posterior encoder, decoder, and conditional prior.
+The example below demonstrates how to generate text based on an image with [`Pipeline`] or the [`AutoModel`] class.
 
-A set of spectrogram-based acoustic features are predicted by the flow-based module, which is formed of a Transformer-based
-text encoder and multiple coupling layers. The spectrogram is decoded using a stack of transposed convolutional layers,
-much in the same style as the HiFi-GAN vocoder. Motivated by the one-to-many nature of the TTS problem, where the same text 
-input can be spoken in multiple ways, the model also includes a stochastic duration predictor, which allows the model to 
-synthesise speech with different rhythms from the same input text. 
-
-The model is trained end-to-end with a combination of losses derived from variational lower bound and adversarial training. 
-To improve the expressiveness of the model, normalizing flows are applied to the conditional prior distribution. During 
-inference, the text encodings are up-sampled based on the duration prediction module, and then mapped into the 
-waveform using a cascade of the flow module and HiFi-GAN decoder. Due to the stochastic nature of the duration predictor,
-the model is non-deterministic, and thus requires a fixed seed to generate the same speech waveform.
-
-The abstract from the paper is the following:
-
-*Several recent end-to-end text-to-speech (TTS) models enabling single-stage training and parallel sampling have been proposed, but their sample quality does not match that of two-stage TTS systems. In this work, we present a parallel end-to-end TTS method that generates more natural sounding audio than current two-stage models. Our method adopts variational inference augmented with normalizing flows and an adversarial training process, which improves the expressive power of generative modeling. We also propose a stochastic duration predictor to synthesize speech with diverse rhythms from input text. With the uncertainty modeling over latent variables and the stochastic duration predictor, our method expresses the natural one-to-many relationship in which a text input can be spoken in multiple ways with different pitches and rhythms. A subjective human evaluation (mean opinion score, or MOS) on the LJ Speech, a single speaker dataset, shows that our method outperforms the best publicly available TTS systems and achieves a MOS comparable to ground truth.*
-
-This model can also be used with TTS checkpoints from [Massively Multilingual Speech (MMS)](https://arxiv.org/abs/2305.13516) 
-as these checkpoints use the same architecture and a slightly modified tokenizer.
-
-This model was contributed by [Matthijs](https://huggingface.co/Matthijs) and [sanchit-gandhi](https://huggingface.co/sanchit-gandhi). The original code can be found [here](https://github.com/jaywalnut310/vits).
-
-## Usage examples
-
-Both the VITS and MMS-TTS checkpoints can be used with the same API. Since the flow-based model is non-deterministic, it 
-is good practice to set a seed to ensure reproducibility of the outputs. For languages with a Roman alphabet, 
-such as English or French, the tokenizer can be used directly to pre-process the text inputs. The following code example 
-runs a forward pass using the MMS-TTS English checkpoint:
+<hfoptions id="usage">
+<hfoption id="Pipeline">
 
 ```python
 import torch
-from transformers import VitsTokenizer, VitsModel, set_seed
+from transformers import pipeline, set_seed
+from scipy.io.wavfile import write
 
-tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")
-model = VitsModel.from_pretrained("facebook/mms-tts-eng")
+set_seed(555)
 
-inputs = tokenizer(text="Hello - my dog is cute", return_tensors="pt")
+pipe = pipeline(
+    task="text-to-speech",
+    model="facebook/mms-tts-eng",
+    torch_dtype=torch.float16,
+    device=0
+)
 
-set_seed(555)  # make deterministic
+speech = pipe("Hello, my dog is cute")
+
+# Extract audio data and sampling rate
+audio_data = speech["audio"]
+sampling_rate = speech["sampling_rate"]
+
+# Save as WAV file
+write("hello.wav", sampling_rate, audio_data.squeeze())
+```
+
+</hfoption>
+<hfoption id="AutoModel">
+
+```python
+import torch
+import scipy
+from IPython.display import Audio
+from transformers import AutoTokenizer, VitsModel, set_seed
+
+tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-eng")
+model = VitsModel.from_pretrained("facebook/mms-tts-eng", torch_dtype=torch.float16).to("cuda")
+inputs = tokenizer("Hello, my dog is cute", return_tensors="pt").to("cuda")
+
+set_seed(555)
 
 with torch.no_grad():
-   outputs = model(**inputs)
+    outputs = model(**inputs)
 
 waveform = outputs.waveform[0]
-```
-
-The resulting waveform can be saved as a `.wav` file:
-
-```python
-import scipy
-
-scipy.io.wavfile.write("techno.wav", rate=model.config.sampling_rate, data=waveform)
-```
-
-Or displayed in a Jupyter Notebook / Google Colab:
-
-```python
-from IPython.display import Audio
+scipy.io.wavfile.write("hello.wav", rate=model.config.sampling_rate, data=waveform)
 
+# display in Colab notebook
 Audio(waveform, rate=model.config.sampling_rate)
 ```
 
-For certain languages with a non-Roman alphabet, such as Arabic, Mandarin or Hindi, the [`uroman`](https://github.com/isi-nlp/uroman) 
-perl package is required to pre-process the text inputs to the Roman alphabet.
+</hfoption>
+</hfoptions>
 
-You can check whether you require the `uroman` package for your language by inspecting the `is_uroman` attribute of 
-the pre-trained `tokenizer`:
+## Notes
 
-```python
-from transformers import VitsTokenizer
+- Set a seed for reproducibility because VITS synthesizes speech non-deterministically.
+- For languages with non-Roman alphabets (Korean, Arabic, etc.), install the [uroman](https://github.com/isi-nlp/uroman) package to preprocess the text inputs to the Roman alphabet. You can check if the tokenizer requires uroman as shown below.
 
-tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")
-print(tokenizer.is_uroman)
-```
-If the is_uroman attribute is `True`, the tokenizer will automatically apply the `uroman` package to your text inputs, but you need to install uroman if not already installed using:  
-```
-pip install --upgrade uroman
-```
-Note: Python version required to use `uroman` as python package should be >= `3.10`. 
-You can use the tokenizer as usual without any additional preprocessing steps:
-```python
-import torch
-from transformers import VitsTokenizer, VitsModel, set_seed
-import os
-import subprocess
+   ```py
+   # pip install -U uroman
+   from transformers import VitsTokenizer
 
-tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-kor")
-model = VitsModel.from_pretrained("facebook/mms-tts-kor")
-text = "이봐 무슨 일이야"
-inputs = tokenizer(text=text, return_tensors="pt")
+   tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")
+   print(tokenizer.is_uroman)
+   ```
 
-set_seed(555)  # make deterministic
-with torch.no_grad():
-   outputs = model(inputs["input_ids"])
+   If your language requires uroman, the tokenizer automatically applies it to the text inputs. Python >= 3.10 doesn't require any additional preprocessing steps. For Python < 3.10, follow the steps below.
 
-waveform = outputs.waveform[0]
-```
-If you don't want to upgrade to python >= `3.10`, then you can use the `uroman` perl package to pre-process the text inputs to the Roman alphabet.
-To do this, first clone the uroman repository to your local machine and set the bash variable `UROMAN` to the local path:
+   ```bash
+   git clone https://github.com/isi-nlp/uroman.git
+   cd uroman
+   export UROMAN=$(pwd)
+   ```
 
+   Create a function to preprocess the inputs. You can either use the bash variable `UROMAN` or pass the directory path directly to the function.
 
-```bash
-git clone https://github.com/isi-nlp/uroman.git
-cd uroman
-export UROMAN=$(pwd)
-```
+   ```py
+   import torch
+   from transformers import VitsTokenizer, VitsModel, set_seed
+   import os
+   import subprocess
 
-You can then pre-process the text input using the following code snippet. You can either rely on using the bash variable 
-`UROMAN` to point to the uroman repository, or you can pass the uroman directory as an argument to the `uromanize` function:
+   tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-kor")
+   model = VitsModel.from_pretrained("facebook/mms-tts-kor")
 
-```python
-import torch
-from transformers import VitsTokenizer, VitsModel, set_seed
-import os
-import subprocess
+   def uromanize(input_string, uroman_path):
+       """Convert non-Roman strings to Roman using the `uroman` perl package."""
+       script_path = os.path.join(uroman_path, "bin", "uroman.pl")
 
-tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-kor")
-model = VitsModel.from_pretrained("facebook/mms-tts-kor")
+       command = ["perl", script_path]
 
-def uromanize(input_string, uroman_path):
-    """Convert non-Roman strings to Roman using the `uroman` perl package."""
-    script_path = os.path.join(uroman_path, "bin", "uroman.pl")
+       process = subprocess.Popen(command, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+       # Execute the perl command
+       stdout, stderr = process.communicate(input=input_string.encode())
 
-    command = ["perl", script_path]
+       if process.returncode != 0:
+           raise ValueError(f"Error {process.returncode}: {stderr.decode()}")
 
-    process = subprocess.Popen(command, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
-    # Execute the perl command
-    stdout, stderr = process.communicate(input=input_string.encode())
+       # Return the output as a string and skip the new-line character at the end
+       return stdout.decode()[:-1]
 
-    if process.returncode != 0:
-        raise ValueError(f"Error {process.returncode}: {stderr.decode()}")
+   text = "이봐 무슨 일이야"
+   uromanized_text = uromanize(text, uroman_path=os.environ["UROMAN"])
 
-    # Return the output as a string and skip the new-line character at the end
-    return stdout.decode()[:-1]
+   inputs = tokenizer(text=uromanized_text, return_tensors="pt")
 
-text = "이봐 무슨 일이야"
-uromanized_text = uromanize(text, uroman_path=os.environ["UROMAN"])
+   set_seed(555)  # make deterministic
+   with torch.no_grad():
+      outputs = model(inputs["input_ids"])
 
-inputs = tokenizer(text=uromanized_text, return_tensors="pt")
-
-set_seed(555)  # make deterministic
-with torch.no_grad():
-   outputs = model(inputs["input_ids"])
-
-waveform = outputs.waveform[0]
-```
+   waveform = outputs.waveform[0]
+   ```
 
 ## VitsConfig
 
@@ -177,10 +148,11 @@ waveform = outputs.waveform[0]
 ## VitsTokenizer
 
 [[autodoc]] VitsTokenizer
-    - __call__
-    - save_vocabulary
+- __call__
+- save_vocabulary
 
 ## VitsModel
 
 [[autodoc]] VitsModel
-    - forward
+- forward
+

docs/source/en/model_doc/yolos.md

@@ -92,6 +92,11 @@ Use [`YolosImageProcessor`] for preparing images (and optional targets) for the
 
 [[autodoc]] YolosImageProcessor
     - preprocess
+
+## YolosImageProcessorFast
+
+[[autodoc]] YolosImageProcessorFast
+    - preprocess
     - pad
     - post_process_object_detection
 

docs/source/en/perf_infer_gpu_one.md

@@ -244,7 +244,7 @@ model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B", device_m
 
 ### Benchmarks
 
-FlashAttention2 speeds up inference considerably especially for inputs with long sequences. However, since FlashAttention2 doesn't support computing attention scores with padding tokens, you must manually pad and unpad the attention scores for batched inference if a sequence contains padding tokens. The downside is batched generation is slower with padding tokens. 
+FlashAttention2 speeds up inference considerably especially for inputs with long sequences. However, since FlashAttention2 doesn't support computing attention scores with padding tokens, you must manually pad and unpad the attention scores for batched inference if a sequence contains padding tokens. The downside is batched generation is slower with padding tokens.
 
 <hfoptions id="padded">
 <hfoption id="short sequence length">

docs/source/en/perf_train_gpu_many.md

@@ -111,7 +111,7 @@ This approach optimizes parallel data processing by reducing idle GPU utilizatio
 
 Data, pipeline and model parallelism combine to form [3D parallelism](https://www.microsoft.com/en-us/research/blog/deepspeed-extreme-scale-model-training-for-everyone/) to optimize memory and compute efficiency.
 
-Memory effiiciency is achieved by splitting the model across GPUs and also dividing it into stages to create a pipeline. This allows GPUs to work in parallel on micro-batches of data, reducing the memory usage of the model, optimizer, and activations.
+Memory efficiency is achieved by splitting the model across GPUs and also dividing it into stages to create a pipeline. This allows GPUs to work in parallel on micro-batches of data, reducing the memory usage of the model, optimizer, and activations.
 
 Compute efficiency is enabled by ZeRO data parallelism where each GPU only stores a slice of the model, optimizer, and activations. This allows higher communication bandwidth between data parallel nodes because communication can occur independently or in parallel with the other pipeline stages.
 

docs/source/en/pipeline_webserver.md

@@ -52,10 +52,10 @@ async def homepage(request):
     return JSONResponse(output)
 
 async def server_loop(q):
-    pipeline = pipeline(task="fill-mask",model="google-bert/bert-base-uncased")
+    pipe = pipeline(task="fill-mask",model="google-bert/bert-base-uncased")
     while True:
         (string, response_q) = await q.get()
-        out = pipeline(string)
+        out = pipe(string)
         await response_q.put(out)
 
 app = Starlette(
@@ -81,6 +81,10 @@ Query the server with a POST request.
 
 ```bash
 curl -X POST -d "Paris is the [MASK] of France." http://localhost:8000/
+```
+This should return the output below.
+
+```bash
 [{'score': 0.9969332218170166,
   'token': 3007,
   'token_str': 'capital',
@@ -112,23 +116,27 @@ The example below is written in pseudocode for readability rather than performan
 1. There is no batch size limit.
 2. The timeout is reset on every queue fetch, so you could end up waiting much longer than the `timeout` value before processing a request. This would also delay the first inference request by that amount of time. The web server always waits 1ms even if the queue is empty, which is inefficient, because that time can be used to start inference. It could make sense though if batching is essential to your use case.
 
-    It would be better to have a single 1ms deadline, instead of resetting it on every fetch.
+    It would be better to have a single 1ms deadline, instead of resetting it on every fetch, as shown below.
 
 ```py
-(string, rq) = await q.get()
-strings = []
-queues = []
-while True:
-    try:
-        (string, rq) = await asyncio.wait_for(q.get(), timeout=0.001)
-    except asyncio.exceptions.TimeoutError:
-        break
-    strings.append(string)
-    queues.append(rq)
-strings
-outs = pipeline(strings, batch_size=len(strings))
-for rq, out in zip(queues, outs):
-    await rq.put(out)
+async def server_loop(q):
+    pipe = pipeline(task="fill-mask", model="google-bert/bert-base-uncased")
+    while True:
+        (string, rq) = await q.get()
+        strings = []
+        queues = []
+        strings.append(string)
+        queues.append(rq)
+        while True:
+            try:
+                (string, rq) = await asyncio.wait_for(q.get(), timeout=1)
+            except asyncio.exceptions.TimeoutError:
+                break
+            strings.append(string)
+            queues.append(rq)
+        outs = pipe(strings, batch_size=len(strings))
+        for rq, out in zip(queues, outs):
+            await rq.put(out)
 ```
 
 ## Error checking

docs/source/en/quantization/auto_round.md

@@ -0,0 +1,286 @@
+<!--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.
+-->
+
+# AutoRound
+
+[AutoRound](https://github.com/intel/auto-round) is an advanced quantization algorithm that delivers strong accuracy, even at 2-bit precision. 
+It leverages sign gradient descent to fine-tune both rounding values and min-max clipping thresholds in just 200 steps. Designed for broad compatibility, it seamlessly supports a wide range of LLMs and is actively expanding to cover more VLMs as well. 
+It also supports quantization and inference across multiple hardware platforms, including CPU, XPU, and CUDA.
+
+AutoRound also offers a variety of useful features, including mixed-bit tuning and inference, lm-head quantization, support for exporting to formats like GPTQ/AWQ/GGUF, and flexible tuning recipes. 
+For a comprehensive overview and the latest updates, check out the AutoRound [README](https://github.com/intel/auto-round).
+
+AutoRound was originally developed as part of the [Intel Neural Compressor](https://github.com/intel/neural-compressor), serving as a general-purpose model compression library for deep learning. 
+It has since evolved into a standalone library focused specifically on low-precision optimization for large language models (LLMs). 
+AutoRound remains fully integrated with the Intel Neural Compressor, and you can explore the repository for more details.
+
+
+## Installation
+
+```bash
+pip install auto-round
+```
+
+## Supported Quantization Configurations
+
+AutoRound supports several quantization configurations:
+
+- **Int8 Weight Only**
+- **Int4 Weight Only**
+- **Int3 Weight Only**
+- **Int2 Weight Only**
+- **Mixed bits Weight only**
+
+## Hardware Compatibility
+
+CPU, XPU, and CUDA for both quantization and inference.
+
+## Quantization and Serialization (offline)
+
+Currently, only offline mode is supported to generate quantized models.
+
+<hfoptions id="quantization">
+<hfoption id="quantization cmd">
+
+### Command Line Usage
+```bash
+auto-round \
+    --model facebook/opt-125m \
+    --bits 4 \
+    --group_size 128 \
+    --output_dir ./tmp_autoround
+```
+
+AutoRound also offer another two recipes, `auto-round-best` and `auto-round-light`, designed for optimal accuracy and improved speed, respectively. 
+For 2 bits, we recommend using `auto-round-best` or `auto-round`.
+</hfoption>
+
+<hfoption id="quantization auto-round api">
+
+### AutoRound API Usage
+This setting offers a better trade-off between accuracy and tuning cost, and is recommended in all scenarios.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from auto_round import AutoRound
+
+model_name = "facebook/opt-125m"
+model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+bits, group_size, sym = 4, 128, True
+# mixed bits config
+# layer_config = {"model.decoder.layers.6.self_attn.out_proj": {"bits": 2, "group_size": 32}}
+autoround = AutoRound(
+    model,
+    tokenizer,
+    bits=bits,
+    group_size=group_size,
+    sym=sym,
+    # enable_torch_compile=True,
+    # layer_config=layer_config,
+)
+
+output_dir = "./tmp_autoround"
+# format= 'auto_round'(default), 'auto_gptq', 'auto_awq'
+autoround.quantize_and_save(output_dir, format='auto_round') 
+```
+
+</hfoption>
+
+<hfoption id="quantization auto-round-best">
+
+### AutoRoundBest recipe
+This setting provides the best accuracy in most scenarios but is 4–5× slower than the standard AutoRound recipe. It is especially recommended for 2-bit quantization and is a good choice if sufficient resources are available.
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from auto_round import AutoRound
+
+model_name = "facebook/opt-125m"
+model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+bits, group_size, sym = 4, 128, True
+autoround = AutoRound(
+    model,
+    tokenizer,
+    bits=bits,
+    group_size=group_size,
+    sym=sym,
+    nsamples=512,
+    iters=1000,
+    low_gpu_mem_usage=True
+)
+
+output_dir = "./tmp_autoround"
+autoround.quantize_and_save(output_dir, format='auto_round') 
+```
+</hfoption>
+
+<hfoption id="quantization auto-round-light">
+
+### AutoRoundLight recipe
+This setting offers the best speed (2 - 3X faster than AutoRound), but it may cause a significant accuracy drop for small models and 2-bit quantization. It is recommended for 4-bit settings and models larger than 3B.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from auto_round import AutoRound
+
+model_name = "facebook/opt-125m"
+model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+bits, group_size, sym = 4, 128, True
+autoround = AutoRound(
+    model,
+    tokenizer,
+    bits=bits,
+    group_size=group_size,
+    sym=sym,
+    iters=50,
+    lr=5e-3,
+)
+
+output_dir = "./tmp_autoround"
+autoround.quantize_and_save(output_dir, format='auto_round') 
+```
+
+</hfoption>
+
+</hfoptions>
+
+W4G128 Average Accuracy of 13 tasks (mmlu-pro, if_eval, gsm8k, etc) and Time Cost Results (Testing was conducted on the Nvidia A100 80G using the version of PyTorch 2.6.0 with enable_torch_compile):
+
+| Model   | Qwen2.5-0.5B-Instruct | Falcon3-3B    | Qwen2.5-7B-Instruct | Meta-Llama-3.1-8B-Instruct | Falcon3-10B   | Qwen2.5-72B-Instruct |
+|---------|--------------------|---------------|------------------|----------------------------|---------------|-------------------|
+| 16bits  | 0.4192             | 0.5203        | 0.6470           | 0.6212                     | 0.6151        | 0.7229            |
+| Best    | **0.4137**(7m)     | **0.5142**(23m) | 0.6426(58m)      | **0.6116**(65m)            | **0.6092**(81m) | 0.7242(575m)      |
+| Default | 0.4129(2m)         | 0.5133(6m)    | 0.6441(13m)      | 0.6106(13m)                | 0.6080(18m)   | **0.7252**(118m)  |
+| Light   | 0.4052(2m)         | 0.5108(3m)    | **0.6453**(5m)   | 0.6104(6m)                 | 0.6063(6m)    | 0.7243(37m)       |
+
+## Inference
+
+AutoRound automatically selects the best available backend based on the installed libraries and prompts the user to install additional libraries when a better backend is found.
+<hfoptions id="inference">
+<hfoption id="inference cpu">
+
+### CPU
+
+Supports 2, 4, and 8 bits. We recommend using intel-extension-for-pytorch (IPEX) for 4 bits inference.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_name = "OPEA/Qwen2.5-1.5B-Instruct-int4-sym-inc"
+model = AutoModelForCausalLM.from_pretrained(model_name, device_map="cpu", torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+text = "There is a girl who likes adventure,"
+inputs = tokenizer(text, return_tensors="pt").to(model.device)
+print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50, do_sample=False)[0]))
+```
+
+</hfoption>
+
+<hfoption id="inference xpu">
+
+### XPU
+
+Supports 4 bits only. We recommend using intel-extension-for-pytorch (IPEX) for inference.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_name = "OPEA/Qwen2.5-1.5B-Instruct-int4-sym-inc"
+model = AutoModelForCausalLM.from_pretrained(model_name, device_map="xpu", torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+text = "There is a girl who likes adventure,"
+inputs = tokenizer(text, return_tensors="pt").to(model.device)
+print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50, do_sample=False)[0]))
+```
+
+</hfoption>
+
+<hfoption id="inference cuda">
+
+### CUDA
+
+Supports 2, 3, 4, and 8 bits. We recommend using GPTQModel for 4 and 8 bits inference.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_name = "OPEA/Qwen2.5-1.5B-Instruct-int4-sym-inc"
+model = AutoModelForCausalLM.from_pretrained(model_name, device_map="cuda", torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+text = "There is a girl who likes adventure,"
+inputs = tokenizer(text, return_tensors="pt").to(model.device)
+print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50, do_sample=False)[0]))
+```
+
+</hfoption>
+
+<hfoption id="inference backend">
+
+### Specify Inference Backend
+
+AutoRound automatically selects the backend for each layer based on compatibility. In general, the priority order is Marlin > ExLLaMAV2 > Triton, but the final choice depends on factors such as group size, bit width, packing format, hardware device, and other implementation details. For more details, please refer to [backends](https://github.com/intel/auto-round?tab=readme-ov-file#specify-backend),
+
+The backend may not always be the most suitable for certain devices. 
+You can specify your preferred backend such as "ipex" for CPU and CPU, "marlin/exllamav2/triton" for CUDA, according to your needs or hardware compatibility. Please note that additional corresponding libraries may be required.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer, AutoRoundConfig
+
+model_name = "OPEA/Qwen2.5-1.5B-Instruct-int4-sym-inc"
+quantization_config = AutoRoundConfig(backend="ipex")
+model = AutoModelForCausalLM.from_pretrained(model_name, device_map="cpu", quantization_config=quantization_config, torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+text = "There is a girl who likes adventure,"
+inputs = tokenizer(text, return_tensors="pt").to(model.device)
+print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50, do_sample=False)[0]))
+```
+
+</hfoption>
+
+
+<hfoption id="format convert">
+
+### Convert GPTQ/AWQ to AutoRound
+
+Most GPTQ/AWQ models can be converted to the AutoRound format for better compatibility and support with Intel devices. Please note that the quantization config will be changed if the model is serialized.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer, AutoRoundConfig
+
+model_name = "ybelkada/opt-125m-gptq-4bit"
+quantization_config = AutoRoundConfig()
+model = AutoModelForCausalLM.from_pretrained(model_name, device_map="cpu", quantization_config=quantization_config, torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+text = "There is a girl who likes adventure,"
+inputs = tokenizer(text, return_tensors="pt").to(model.device)
+print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50, do_sample=False)[0]))
+```
+
+</hfoption>
+
+</hfoptions>
+
+## Issues
+
+If you encounter any issues with the transformers integration, please open an issue on
+the [transformers](https://github.com/huggingface/transformers/issues) repository.  
+If you encounter any issues with auto-round, please open an issue on
+the [AutoRound](https://github.com/intel/auto-round/issues) repository.
+
+
+## Acknowledgement
+Special thanks to open-source low precision libraries such as AutoGPTQ, AutoAWQ, GPTQModel, Triton, Marlin, and ExLLaMAV2 for providing low-precision CUDA kernels, which are leveraged in AutoRound.
+
+## Contribution
+Contributions to [AutoRound](https://github.com/intel/auto-round/pulls) are welcome and greatly appreciated!
+Whether it's fixing bugs, improving documentation, adding new features, or suggesting improvements, your help is always valued.

docs/source/en/quantization/bitsandbytes.md

@@ -14,13 +14,21 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# bitsandbytes
+# Bitsandbytes
 
-[bitsandbytes](https://github.com/bitsandbytes-foundation/bitsandbytes) features the LLM.int8 and QLoRA quantization to enable accessible large language model inference and training.
+The [bitsandbytes](https://github.com/bitsandbytes-foundation/bitsandbytes) library provides quantization tools for LLMs through a lightweight Python wrapper around CUDA functions. It enables working with large models using limited computational resources by reducing their memory footprint.
 
-[LLM.int8()](https://hf.co/papers/2208.07339) is a quantization method that aims to make large language model inference more accessible without significant degradation. Unlike naive 8-bit quantization, which can result in loss of critical information and accuracy, LLM.int8() dynamically adapts to ensure sensitive components of the computation retain higher precision when needed.
+At its core, bitsandbytes provides:
 
-QLoRA, or 4-bit quantization, compresses a model even further to 4-bits and inserts a small set of trainable low-rank adaptation (LoRA) weights to allowing training. 
+- **Quantized Linear Layers**: `Linear8bitLt` and `Linear4bit` layers that replace standard PyTorch linear layers with memory-efficient quantized alternatives
+- **Optimized Optimizers**: 8-bit versions of common optimizers through its `optim` module, enabling training of large models with reduced memory requirements
+- **Matrix Multiplication**: Optimized matrix multiplication operations that leverage the quantized format
+
+bitsandbytes offers two main quantization features:
+
+1. **LLM.int8()** - An 8-bit quantization method that makes inference more accessible without significant performance degradation. Unlike naive quantization, [LLM.int8()](https://hf.co/papers/2208.07339) dynamically preserves higher precision for critical computations, preventing information loss in sensitive parts of the model.
+
+2. **QLoRA** - A 4-bit quantization technique that compresses models even further while maintaining trainability by inserting a small set of trainable low-rank adaptation (LoRA) weights.
 
 > **Note:** For a user-friendly quantization experience, you can use the `bitsandbytes` [community space](https://huggingface.co/spaces/bnb-community/bnb-my-repo).
 
@@ -30,12 +38,38 @@ Run the command below to install bitsandbytes.
 ```bash
 pip install --upgrade transformers accelerate bitsandbytes
 ```
+To compile from source, follow the instructions in the [bitsandbytes installation guide](https://huggingface.co/docs/bitsandbytes/main/en/installation).
+
+## Hardware Compatibility
+bitsandbytes is currently only supported on CUDA GPUs for CUDA versions 11.0 - 12.8. However, there's an ongoing multi-backend effort under development, which is currently in alpha. If you're interested in providing feedback or testing, check out the [bitsandbytes repository](https://github.com/bitsandbytes-foundation/bitsandbytes) for more information.
+
+### CUDA
+
+| Feature | Minimum Hardware Requirement |
+|---------|-------------------------------|
+| 8-bit optimizers | NVIDIA Maxwell (GTX 900 series, TITAN X, M40) or newer GPUs * |
+| LLM.int8() | NVIDIA Turing (RTX 20 series, T4) or newer GPUs |
+| NF4/FP4 quantization | NVIDIA Maxwell (GTX 900 series, TITAN X, M40) or newer GPUs * |
+
+### Multi-backend
+
+| Backend | Supported Versions | Python versions | Architecture Support | Status |
+|---------|-------------------|----------------|---------------------|---------|
+| AMD ROCm | 6.1+ | 3.10+ | minimum CDNA - gfx90a, RDNA - gfx1100 | Alpha |
+| Apple Silicon (MPS) | WIP | 3.10+ | M1/M2 chips | Planned |
+| Intel CPU | v2.4.0+ (ipex) | 3.10+ | Intel CPU | Alpha |
+| Intel GPU | v2.4.0+ (ipex) | 3.10+ | Intel GPU | Experimental |
+| Ascend NPU | 2.1.0+ (torch_npu) | 3.10+ | Ascend NPU | Experimental |
+
+> **Note:** Bitsandbytes is moving away from the multi-backend approach towards using [Pytorch Custom Operators](https://pytorch.org/tutorials/advanced/custom_ops_landing_page.html), as the main mechanism for supporting new hardware, and dispatching to the correct backend.
+
+## Quantization Examples
 
 Quantize a model by passing a [`BitsAndBytesConfig`] to [`~PreTrainedModel.from_pretrained`]. This works for any model in any modality, as long as it supports [Accelerate](https://huggingface.co/docs/accelerate/index) and contains [torch.nn.Linear](https://pytorch.org/docs/stable/generated/torch.nn.Linear.html) layers.
 
 <hfoptions id="bnb">
 <hfoption id="8-bit">
-
+<div class="bnb-container" style="border: 1px solid #ddd; border-radius: 8px; padding: 20px; margin: 20px 0">
 Quantizing a model in 8-bit halves the memory-usage, and for large models, set `device_map="auto"` to efficiently distribute the weights across all available GPUs.
 
 ```py
@@ -45,6 +79,7 @@ quantization_config = BitsAndBytesConfig(load_in_8bit=True)
 
 model_8bit = AutoModelForCausalLM.from_pretrained(
     "bigscience/bloom-1b7", 
+    device_map="auto",
     quantization_config=quantization_config
 )
 ```
@@ -59,6 +94,7 @@ quantization_config = BitsAndBytesConfig(load_in_8bit=True)
 
 model_8bit = AutoModelForCausalLM.from_pretrained(
     "facebook/opt-350m", 
+    device_map="auto",
     quantization_config=quantization_config, 
     torch_dtype="auto"
 )
@@ -74,16 +110,16 @@ quantization_config = BitsAndBytesConfig(load_in_8bit=True)
 
 model = AutoModelForCausalLM.from_pretrained(
     "bigscience/bloom-560m", 
+    device_map="auto",
     quantization_config=quantization_config
 )
-tokenizer = AutoTokenizer.from_pretrained("bigscience/bloom-560m")
 
 model.push_to_hub("bloom-560m-8bit")
 ```
-
+</div>
 </hfoption>
 <hfoption id="4-bit">
-
+<div class="bnb-container" style="border: 1px solid #ddd; border-radius: 8px; padding: 20px; margin: 20px 0">
 Quantizing a model in 4-bit reduces your memory-usage by 4x, and for large models, set `device_map="auto"` to efficiently distribute the weights across all available GPUs.
 
 ```py
@@ -93,6 +129,7 @@ quantization_config = BitsAndBytesConfig(load_in_4bit=True)
 
 model_4bit = AutoModelForCausalLM.from_pretrained(
     "bigscience/bloom-1b7",
+    device_map="auto",
     quantization_config=quantization_config
 )
 ```
@@ -107,6 +144,7 @@ quantization_config = BitsAndBytesConfig(load_in_4bit=True)
 
 model_4bit = AutoModelForCausalLM.from_pretrained(
     "facebook/opt-350m",
+    device_map="auto",
     quantization_config=quantization_config, 
     torch_dtype="auto"
 )
@@ -115,6 +153,20 @@ model_4bit.model.decoder.layers[-1].final_layer_norm.weight.dtype
 
 Make sure you have the latest bitsandbytes version so you can serialize 4-bit models and push them to the Hub with [`~PreTrainedModel.push_to_hub`]. Use [`~PreTrainedModel.save_pretrained`] to save the 4-bit model locally.  
 
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+
+quantization_config = BitsAndBytesConfig(load_in_4bit=True)
+
+model = AutoModelForCausalLM.from_pretrained(
+    "bigscience/bloom-560m", 
+    device_map="auto",
+    quantization_config=quantization_config
+)
+
+model.push_to_hub("bloom-560m-4bit")
+```
+</div>
 </hfoption>
 </hfoptions>
 

docs/source/en/quantization/concept_guide.md

@@ -0,0 +1,178 @@
+<!--Copyright 2024 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.
+
+-->
+
+# Quantization concepts
+
+Quantization reduces the memory footprint and computational cost of large machine learning models like those found in the Transformers library. It achieves this by representing the model's weights and or activations with lower-precision data types (like 8-bit integers or int8) instead of the standard 32-bit floating-point (float32).
+
+
+Reducing a model's precision offers several significant benefits:
+
+-  Smaller model size: Lower-precision data types require less storage space. An int8 model, for example, is roughly 4 times smaller than its float32 counterpart.
+-  Faster inference: Operations on lower-precision data types, especially integers, can be significantly faster on compatible hardware (CPUs and GPUs often have specialized instructions for int8 operations). This leads to lower latency.
+-  Reduced energy consumption: Faster computations and smaller memory transfers often translate to lower power usage.
+
+The primary trade-off in quantization is *efficiency* vs. *accuracy*. Reducing precision saves resources but inevitably introduces small errors (quantization noise). The goal is to minimize this error using appropriate schemes (affine/symmetric), granularity (per-tensor/channel), and techniques (PTQ/QAT) so that the model's performance on its target task degrades as little as possible.
+
+The sections below cover quantization schemes, granularity, and techniques.
+
+## Quantization schemes
+
+The core idea is to map the range of values found in the original float32 weights and activations to the much smaller range represented by int8 (typically \\([-128, 127]\\)).
+
+This section covers how some quantization techniques work.
+
+<div class="flex justify-center">
+    <img width="606" alt="quant_visual" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/quant_visual.png" />
+</div>
+
+### Affine quantization
+
+The most common method is *affine quantization*. For a given float32 tensor (like a layer's weights), it finds the minimum \\(val_{min}\\) and maximum \\(val_{max}\\) values. This range \\([val_{min}, val_{max}]\\) is mapped to the int8 range \\([q_{min}, q_{max}]\\), which is typically \\([-128, 127]\\).
+
+There are two main ways to perform this mapping, *symmetric* and *asymmetric*. The choice between symmetric and asymmetric quantization determines how the float32 range is mapped to the int8 range.
+
+- Symmetric: This method assumes the original float32 range is symmetric around zero ( \\([ -a, a ]\\) ). This range is mapped symmetrically to the int8 range, for example, \\([-127, 127]\\). A key characteristic is that the float32 value \\(0.0\\) maps directly to the int8 value \\(0\\). This only requires one parameter, the **scale ( \\(S\\) )**, to define the mapping. It can simplify computations, but it might be less accurate if the original data distribution isn't naturally centered around zero.
+- Asymmetric (Affine): This method does not assume the data is centered around zero. It maps the exact range \\([val_{min}, val_{max}]\\) from float32 to the full int8 range, like \\([-128, 127]\\). This requires two parameters, a **scale ( \\(S\\) )** and a **zero-point ( \\(Z\\) )**. 
+
+
+    scale ( \\(S\\) ): A positive float32 number representing the ratio between the float32 and the int8 range.
+
+$$
+S = \frac{val_{max} - val_{min}}{q_{max} - q_{min}}
+$$
+
+zero-Point ( \\(Z\\) ): An int8 value that corresponds to the float32 value \\(0.0\\).
+
+$$
+Z = q_{min} - round\left(\frac{val_{min}}{S}\right)
+$$
+
+> [!TIP]
+> In symmetric quantization, Z would typically be fixed at 0.
+
+With these parameters, a float32 value, \\(x\\). can be quantized to int8 ( \\(q\\) ) with the formula below.
+
+$$
+q = round\left(\frac{x}{S} + Z\right)
+$$
+
+The int8 value, \\(q\\), can be dequantized back to approximate float32 with the formula below.
+
+$$
+x \approx S \cdot (q - Z)
+$$
+
+<div class="flex justify-center">
+    <img width="606" alt="dequant" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/dequant.png" />
+</div>
+
+During inference, computations like matrix multiplication are performed using the int8 values ( \\(q\\) ), and the result is dequantized back to float32 (often using a higher-precision accumulation type like int32 internally) before it is passed to the next layer.
+
+### int4 and weight packing
+
+<div class="flex justify-center">
+    <img width="606" alt="weight packing" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/weight_packing.png" />
+</div>
+
+int4 quantization further reduces the model size and memory usage (halving it compared to int8). The same affine or symmetric quantization principles apply, mapping the float32 range to the 16 possible values representable by int4 ( \\([-8, 7]\\) for signed int4).
+
+A key aspect of int4 quantization is **weight packing**. Since most hardware can't natively handle 4-bit data types in memory, two int4 values are typically packed together into a single int8 byte for storage and transfer. For example, the first value might occupy the lower 4 bits and the second value the upper 4 bits of the byte (`packed_byte = (val1 & 0x0F) | (val2 << 4)`).
+
+int4 is still beneficial even without native int4 compute because the primary benefit comes from reduced memory bandwidth. Loading packed int4 weights (stored as int8) from memory (RAM or VRAM) to the compute units is twice as fast as loading int8 weights. For large models, memory access is often a significant bottleneck. The speed up from faster data transfer can outweigh the computational overhead of unpacking and dequantizing on the fly, leading to overall faster inference, especially in memory-bound scenarios.
+
+However, int4 quantization typically results in a larger accuracy drop compared to int8. Advanced quantization techniques like [GPTQ](./gptq) or [AWQ](./awq) are often necessary for good performance with int4.
+
+### FP8 Quantization (A8W8)
+
+<div class="flex justify-center">
+    <img width="606" alt="fp8" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/fp8.png" />
+</div>
+A newer datatype, 8-bit floating-point (FP8), offers another way to reduce precision while retaining more accuracy than int8 in certain scenarios. FP8 keeps the floating-point structure (sign, exponent, mantissa) but uses fewer bits.
+
+There are two common FP8 variants.
+
+- E4M3: 1 sign bit, 4 exponent bits, 3 mantissa bits. Offers higher precision (more mantissa bits) but a smaller dynamic range (fewer exponent bits).
+- E5M2: 1 sign bit, 5 exponent bits, 2 mantissa bits. Offers a wider dynamic range but lower precision.
+
+FP8 is used in the *A8W8* quantization scheme, which quantizes both activations (A) and weights (W) to 8-bit precision.
+
+While int8 has broad support, efficient FP8 computation requires specific hardware capabilities found in newer GPUs like NVIDIA H100/H200/B100 and AMD Instinct MI300 series. Without native hardware acceleration, the benefits of FP8 might not be fully realized.
+
+Transformers supports FP8 through specific backends like [FBGEMM](./fbgemm_fp8), [FineGrainedFP8](./finegrained_fp8), and [compressed-tensors](./compressed_tensors). These backends handle the underlying FP8 conversion and computation when the appropriate hardware and configurations are used.
+
+## Granularity
+
+Quantization parameters ( \\(S\\) and \\(Z\\)) can be calculated in one of two ways.
+
+- Per-Tensor: One set of \\(S\\) and \\(Z\\) for the entire tensor. Simpler, but less accurate if data values vary greatly within the tensor.
+- Per-Channel (or Per-Group/Block): Separate \\(S\\) and \\(Z\\) for each channel or group. More accurate and better performance at the cost of slightly more complexity and memory.
+
+<div class="flex justify-center">
+    <img width="625" alt="Granularities" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/Granularities.png" />
+</div>
+
+## Quantization techniques
+
+There are two main types of quantization techniques.
+
+- Post-Training Quantization (PTQ): Quantization is applied  *after* the model is fully trained.
+- Quantization-Aware Training (QAT): Quantization effects are simulated *during* training by inserting "fake quantization" ops that simulate the rounding errors of quantization. This lets the model adapt to quantization, and usually results in better accuracy, especially at lower bit-widths.
+
+## Quantization in Transformers
+
+Transformers integrates several quantization backends such as bitsandbytes, torchao, compressed-tensors, and more (refer to the quantization [overview](./overview) for more backends). 
+
+
+All backends are unified under the [`HfQuantizer`] API and associated [`QuantizationConfig`] classes. You can integrate your own custom quantization backends by implementing a custom [`HfQuantizer`] and [`QuantizationConfig`], as shown in the [Contribution](./contribute) guide.
+
+The typical workflow for quantization in Transformers is to:
+
+1. Choose a quantization method suitable for your hardware and use case (see the [Overview](./overview) or [Selecting a quantization method](./selecting) guide to help you).
+2. Load a pre-quantized model from the Hugging Face Hub or load a float32/float16/bfloat16 model and apply a specific quantization method with [`QuantizationConfig`].
+
+The example below demonstrates loading a 8B parameter model and quantizing it to 4-bits with bitsandbytes.
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+
+model_id = "meta-llama/Llama-3.1-8B-Instruct"
+
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.bfloat16
+)
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_id,
+    quantization_config=quantization_config,
+    torch_dtype=torch.bfloat16,
+    device_map="auto"
+)
+```
+
+
+## Resources
+
+To explore quantization and related performance optimization concepts more deeply, check out the following resources.
+
+- [Quantization Fundamentals with Hugging Face](https://www.deeplearning.ai/short-courses/quantization-fundamentals-with-hugging-face/)
+- [Quantization in Depth](https://www.deeplearning.ai/short-courses/quantization-in-depth)
+- [Introduction to Quantization cooked in 🤗 with 💗🧑‍🍳](https://huggingface.co/blog/merve/quantization)
+- [EfficientML.ai Lecture 5 - Quantization Part I](https://www.youtube.com/watch?v=RP23-dRVDWM)
+- [Making Deep Learning Go Brrrr From First Principles](https://horace.io/brrr_intro.html)
+- [Accelerating Generative AI with PyTorch Part 2: LLM Optimizations](https://pytorch.org/blog/accelerating-generative-ai-2/)

docs/source/en/quantization/overview.md

@@ -22,25 +22,26 @@ Transformers supports many quantization methods, each with their pros and cons,
 
 Use the Space below to help you pick a quantization method depending on your hardware and number of bits to quantize to.
 
-| Quantization Method                           | On the fly quantization | CPU             | CUDA GPU | ROCm GPU  | Metal (Apple Silicon)              | Intel GPU       | Torch compile() | Bits          | PEFT Fine Tuning | Serializable with 🤗Transformers | 🤗Transformers Support  | Link to library                             |
-|-----------------------------------------------|----------------------|-----------------|----------|-----------|------------------------------------|-----------------|-----------------|---------------|------------------|-----------------------------|-------------------------|---------------------------------------------|
-| [AQLM](./aqlm)                             | 🔴                   | 🟢              |     🟢     | 🔴        | 🔴                                 | 🔴              | 🟢              | 1/2         | 🟢               | 🟢                          | 🟢                      | https://github.com/Vahe1994/AQLM            |
-| [AWQ](./awq)                               | 🔴                   | 🟢              | 🟢        | 🟢        | 🔴                                 | 🟢              | ?               | 4             | 🟢               | 🟢                          | 🟢                      | https://github.com/casper-hansen/AutoAWQ    |
-| [bitsandbytes](./bitsandbytes)             | 🟢                   | 🟡 |     🟢     | 🟡 | 🔴                    | 🟡 | 🔴 | 4/8         | 🟢               | 🟢                          | 🟢                      | https://github.com/bitsandbytes-foundation/bitsandbytes |
-| [compressed-tensors](./compressed_tensors) | 🔴                   | 🟢              |     🟢     | 🟢        | 🔴                                 | 🔴              | 🔴              | 1/8         | 🟢               | 🟢                          | 🟢                      | https://github.com/neuralmagic/compressed-tensors |
-| [EETQ](./eetq)                             | 🟢                   | 🔴              | 🟢        | 🔴        | 🔴                                 | 🔴              | ?               | 8             | 🟢               | 🟢                          | 🟢                      | https://github.com/NetEase-FuXi/EETQ        |
-| [GGUF / GGML (llama.cpp)](../gguf)         | 🟢                   | 🟢              | 🟢        | 🔴        | 🟢                                 | 🔴              | 🔴              | 1/8         | 🔴               | [See Notes](../gguf)     | [See Notes](../gguf) | https://github.com/ggerganov/llama.cpp      |
-| [GPTQModel](./gptq)                        | 🔴                   | 🟢 | 🟢        | 🟢        | 🟢                                 | 🟢 | 🔴              | 2/3/4/8 | 🟢               | 🟢                          | 🟢                      | https://github.com/ModelCloud/GPTQModel        |
-| [AutoGPTQ](./gptq)                         | 🔴                   | 🔴              | 🟢        | 🟢        | 🔴                                 | 🔴              | 🔴              | 2/3/4/8 | 🟢               | 🟢                          | 🟢                      | https://github.com/AutoGPTQ/AutoGPTQ        |
-| [HIGGS](./higgs)                           | 🟢                   | 🔴              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🟢              | 2/4         | 🔴               | 🟢                          | 🟢                      | https://github.com/HanGuo97/flute           |       
-| [HQQ](./hqq)                               | 🟢                   | 🟢              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🟢              | 1/8         | 🟢               | 🔴                          | 🟢                      | https://github.com/mobiusml/hqq/            |
-| [optimum-quanto](./quanto)                 | 🟢                   | 🟢              | 🟢        | 🔴        | 🟢                                 | 🔴              | 🟢              | 2/4/8     | 🔴               | 🔴                          | 🟢                      | https://github.com/huggingface/optimum-quanto       |
-| [FBGEMM_FP8](./fbgemm_fp8)                 | 🟢                   | 🔴              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🔴              | 8             | 🔴               | 🟢                          | 🟢                      | https://github.com/pytorch/FBGEMM       |
-| [torchao](./torchao)                       | 🟢                   | 🟢               | 🟢        | 🔴        | 🟡 | 🔴              |                 | 4/8         |                  | 🟢🔴                        | 🟢                      | https://github.com/pytorch/ao       |
-| [VPTQ](./vptq)                             | 🔴                   | 🔴              |     🟢     | 🟡        | 🔴                                 | 🔴              | 🟢              | 1/8         | 🔴               | 🟢                          | 🟢                      | https://github.com/microsoft/VPTQ            |
-| [FINEGRAINED_FP8](./finegrained_fp8)                 | 🟢                   | 🔴              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🔴              | 8             | 🔴               | 🟢                          | 🟢                      |        |
-| [SpQR](./spqr)                          | 🔴                       |  🔴   | 🟢        | 🔴              |    🔴    | 🔴         |         🟢              | 3              |              🔴                     | 🟢           | 🟢                      | https://github.com/Vahe1994/SpQR/       |
-| [Quark](./quark)                           | 🔴                       | 🟢 | 🟢      | 🟢      | 🟢                   | 🟢       | ?               | 2/4/6/8/9/16 | 🔴                | 🔴                               | 🟢                       | https://quark.docs.amd.com/latest/                      |
+| Quantization Method                       | On the fly quantization | CPU             | CUDA GPU | ROCm GPU  | Metal (Apple Silicon)              | Intel GPU       | Torch compile() | Bits         | PEFT Fine Tuning | Serializable with 🤗Transformers | 🤗Transformers Support  | Link to library                             |
+|-------------------------------------------|----------------------|-----------------|----------|-----------|------------------------------------|-----------------|-----------------|--------------|------------------|-----------------------------|-------------------------|---------------------------------------------|
+| [AQLM](./aqlm)                            | 🔴                   | 🟢              |     🟢     | 🔴        | 🔴                                 | 🔴              | 🟢              | 1/2          | 🟢               | 🟢                          | 🟢                      | https://github.com/Vahe1994/AQLM            |
+| [AutoRound](./auto_round)                 | 🔴                   | 🟢               | 🟢          |   🔴        |   🔴                                |   🟢              |   🔴               | 2/3/4/8      |    🔴              |       🟢                      |    🟢                       |      https://github.com/intel/auto-round                                       |
+| [AWQ](./awq)                              | 🔴                   | 🟢              | 🟢        | 🟢        | 🔴                                 | 🟢              | ?               | 4            | 🟢               | 🟢                          | 🟢                      | https://github.com/casper-hansen/AutoAWQ    |
+| [bitsandbytes](./bitsandbytes)            | 🟢                   | 🟡 |     🟢     | 🟡 | 🔴                    | 🟡 | 🔴 | 4/8          | 🟢               | 🟢                          | 🟢                      | https://github.com/bitsandbytes-foundation/bitsandbytes |
+| [compressed-tensors](./compressed_tensors) | 🔴                   | 🟢              |     🟢     | 🟢        | 🔴                                 | 🔴              | 🔴              | 1/8          | 🟢               | 🟢                          | 🟢                      | https://github.com/neuralmagic/compressed-tensors |
+| [EETQ](./eetq)                            | 🟢                   | 🔴              | 🟢        | 🔴        | 🔴                                 | 🔴              | ?               | 8            | 🟢               | 🟢                          | 🟢                      | https://github.com/NetEase-FuXi/EETQ        |
+| [GGUF / GGML (llama.cpp)](../gguf)        | 🟢                   | 🟢              | 🟢        | 🔴        | 🟢                                 | 🔴              | 🔴              | 1/8          | 🔴               | [See Notes](../gguf)     | [See Notes](../gguf) | https://github.com/ggerganov/llama.cpp      |
+| [GPTQModel](./gptq)                       | 🔴                   | 🟢 | 🟢        | 🟢        | 🟢                                 | 🟢 | 🔴              | 2/3/4/8      | 🟢               | 🟢                          | 🟢                      | https://github.com/ModelCloud/GPTQModel        |
+| [AutoGPTQ](./gptq)                        | 🔴                   | 🔴              | 🟢        | 🟢        | 🔴                                 | 🔴              | 🔴              | 2/3/4/8      | 🟢               | 🟢                          | 🟢                      | https://github.com/AutoGPTQ/AutoGPTQ        |
+| [HIGGS](./higgs)                          | 🟢                   | 🔴              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🟢              | 2/4          | 🔴               | 🟢                          | 🟢                      | https://github.com/HanGuo97/flute           |       
+| [HQQ](./hqq)                              | 🟢                   | 🟢              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🟢              | 1/8          | 🟢               | 🔴                          | 🟢                      | https://github.com/mobiusml/hqq/            |
+| [optimum-quanto](./quanto)                | 🟢                   | 🟢              | 🟢        | 🔴        | 🟢                                 | 🔴              | 🟢              | 2/4/8        | 🔴               | 🔴                          | 🟢                      | https://github.com/huggingface/optimum-quanto       |
+| [FBGEMM_FP8](./fbgemm_fp8)                | 🟢                   | 🔴              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🔴              | 8            | 🔴               | 🟢                          | 🟢                      | https://github.com/pytorch/FBGEMM       |
+| [torchao](./torchao)                      | 🟢                   | 🟢               | 🟢        | 🔴        | 🟡 | 🔴              |                 | 4/8          |                  | 🟢🔴                        | 🟢                      | https://github.com/pytorch/ao       |
+| [VPTQ](./vptq)                            | 🔴                   | 🔴              |     🟢     | 🟡        | 🔴                                 | 🔴              | 🟢              | 1/8          | 🔴               | 🟢                          | 🟢                      | https://github.com/microsoft/VPTQ            |
+| [FINEGRAINED_FP8](./finegrained_fp8)      | 🟢                   | 🔴              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🔴              | 8            | 🔴               | 🟢                          | 🟢                      |        |
+| [SpQR](./spqr)                            | 🔴                     |  🔴   | 🟢        | 🔴              |    🔴    | 🔴         |         🟢              | 3            |              🔴                     | 🟢           | 🟢                      | https://github.com/Vahe1994/SpQR/       |
+| [Quark](./quark)                          | 🔴                     | 🟢 | 🟢      | 🟢      | 🟢                   | 🟢       | ?               | 2/4/6/8/9/16 | 🔴                | 🔴                               | 🟢                       | https://quark.docs.amd.com/latest/                      |
 
 ## Resources
 

docs/source/en/quantization/selecting.md

@@ -130,6 +130,28 @@ Methods like [AQLM](./aqlm), [SpQR](./spqr), [VPTQ](./vptq), [HIGGS](./higgs), e
     *   You have significant compute resources available for potentially complex quantization procedures.
 We recommend consulting each methods documentation and associated papers carefully before choosing one for use in production.
 
+## Benchmark Comparison
+
+To provide a quantitative comparison of different quantization methods, we benchmarked several popular techniques on the Llama 3.1 8B and 70B models. The following tables show results for accuracy (higher is better), inference throughput measured in tokens/second (higher is better), peak VRAM usage measured in GB (lower is better), and quantization time.
+
+Performance metrics were measured on 2 NVIDIA A100 80GB GPU for Llama 3.1 70B (bfloat16), 1 NVIDIA H100 80GB GPU for FP8 methods, and 1 NVIDIA A100 80GB GPU for all other methods. Throughput was measured with a batch size of 1 and generating 64 tokens.
+Results for `torch.compile` and Marlin kernels are included where applicable and supported.
+
+<iframe
+  src="https://huggingface.co/datasets/derekl35/quantization-benchmarks/embed/viewer/default/train"
+  frameborder="0"
+  width="100%"
+  height="560px"
+  title="benchmarking results dataset"
+></iframe>
+
+The key takeaways are:
+
+| Quantization & Methods                      | Memory Savings (vs bf16) | Accuracy             | Other Notes                                                        |
+|-------------------------------------------- |------------------------- |--------------------- |------------------------------------------------------------------- |
+| **8-bit** (bnb-int8, HQQ, Quanto, torchao, fp8) | ~2x             | Very close to baseline bf16 model   |                                                                    |
+| **4-bit** (AWQ, GPTQ, HQQ, bnb-nf4)    | ~4x                      | Relatively high accuracy            | AWQ/GPTQ often lead in accuracy but need calibration. HQQ/bnb-nf4 are easy on-the-fly. |
+| **Sub-4-bit** (VPTQ, AQLM, 2-bit GPTQ) | Extreme (>4x)            | Noticeable drop, especially at 2-bit | Quantization times can be very long (AQLM, VPTQ). Performance varies. |
 
 > [!TIP]
 > Always benchmark the performance (accuracy and speed) of the quantized model on your specific task and hardware to ensure it meets your requirements. Refer to the individual documentation pages linked above for detailed usage instructions.

docs/source/en/quantization/torchao.md

@@ -11,50 +11,100 @@ rendered properly in your Markdown viewer.
 
 # torchao
 
+[![Open In Colab: Torchao Demo](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/transformers_doc/en/quantization/torchao.ipynb)
+
 [torchao](https://github.com/pytorch/ao) is a PyTorch architecture optimization library with support for custom high performance data types, quantization, and sparsity. It is composable with native PyTorch features such as [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) for even faster inference and training.
 
-Install torchao with the following command.
+See the table below for additional torchao features.
+
+| Feature | Description |
+|--------|-------------|
+| **Quantization Aware Training (QAT)** | Train quantized models with minimal accuracy loss (see [QAT README](https://github.com/pytorch/ao/blob/main/torchao/quantization/qat/README.md)) |
+| **Float8 Training** | High-throughput training with float8 formats (see [torchtitan](https://github.com/pytorch/torchtitan/blob/main/docs/float8.md) and [Accelerate](https://huggingface.co/docs/accelerate/usage_guides/low_precision_training#configuring-torchao) docs) |
+| **Sparsity Support** | Semi-structured (2:4) sparsity for faster inference (see [Accelerating Neural Network Training with Semi-Structured (2:4) Sparsity](https://pytorch.org/blog/accelerating-neural-network-training/) blog post) |
+| **Optimizer Quantization** | Reduce optimizer state memory with 4 and 8-bit variants of Adam |
+| **KV Cache Quantization** | Enables long context inference with lower memory (see [KV Cache Quantization](https://github.com/pytorch/ao/blob/main/torchao/_models/llama/README.md)) |
+| **Custom Kernels Support** | use your own `torch.compile` compatible ops |
+| **FSDP2** | Composable with FSDP2 for training|
+
+> [!TIP]
+> Refer to the torchao [README.md](https://github.com/pytorch/ao#torchao-pytorch-architecture-optimization) for more details about the library.
+
+
+torchao supports the [quantization techniques](https://github.com/pytorch/ao/blob/main/torchao/quantization/README.md) below.
+
+- A16W8 Float8 Dynamic Quantization
+- A16W8 Float8 WeightOnly Quantization
+- A8W8 Int8 Dynamic Quantization
+- A16W8 Int8 Weight Only Quantization
+- A16W4 Int4 Weight Only Quantization
+- Autoquantization
+
+
+Check the table below to see if your hardware is compatible.
+
+| Component | Compatibility |
+|----------|----------------|
+| CUDA Versions | ✅ cu118, cu126, cu128 |
+| CPU | ✅ change `device_map="cpu"` (see examples below) |
+
+
+
+Install torchao from PyPi or the PyTorch index with the following commands.
+
+<hfoptions id="install torchao">
+<hfoption id="PyPi">
 
 ```bash
 # Updating 🤗 Transformers to the latest version, as the example script below uses the new auto compilation
-pip install --upgrade torch torchao transformers
+# Stable release from Pypi which will default to CUDA 12.6
+pip install --upgrade torchao transformers
 ```
+</hfoption> 
+<hfoption id="PyTorch Index">
+Stable Release from the PyTorch index
+```bash
+pip install torchao --index-url https://download.pytorch.org/whl/cu126 # options are cpu/cu118/cu126/cu128
+```
+</hfoption>
+</hfoptions>
 
-torchao supports many quantization types for different data types (int4, float8, weight only, etc.).
-Starting with version 0.10.0, torchao provides enhanced flexibility through the `AOBaseConfig` API, allowing for more customized quantization configurations.
-And full access to the techniques offered in the torchao library.
+If your torcha version is below 0.10.0, you need to upgrade it, please refer to the [deprecation notice](#deprecation-notice) for more details.
+
+## Quantization examples
+
+TorchAO provides a variety of quantization configurations. Each configuration can be further customized with parameters such as `group_size`, `scheme`, and `layout` to optimize for specific hardware and model architectures.
+
+For a complete list of available configurations, see the [quantization API documentation](https://github.com/pytorch/ao/blob/main/torchao/quantization/quant_api.py).
 
 You can manually choose the quantization types and settings or automatically select the quantization types.
 
-<hfoptions id="torchao">
-<hfoption id="manual">
+Create a [`TorchAoConfig`] and specify the quantization type and `group_size` of the weights to quantize (for int8 weight only and int4 weight only). Set the `cache_implementation` to `"static"` to automatically [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) the forward method.
 
+We'll show examples for recommended quantization methods based on hardwares, e.g. A100 GPU, H100 GPU, CPU.
 
-Create a [`TorchAoConfig`] and specify the quantization type and `group_size` of the weights to quantize. Set the `cache_implementation` to `"static"` to automatically [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) the forward method.
-
-> [!TIP]
-> Run the quantized model on a CPU by changing `device_map` to `"cpu"` and `layout` to `Int4CPULayout()`. This is only available in torchao 0.8.0+.
-
-In torchao 0.10.0+, you can use the more flexible `AOBaseConfig` approach instead of string identifiers:
-
+### H100 GPU
+<hfoptions id="examples-H100-GPU">
+<hfoption id="float8-dynamic-and-weight-only">
 ```py
 import torch
 from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
-from torchao.quantization import Int4WeightOnlyConfig
+from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
 
-# Using AOBaseConfig instance (torchao >= 0.10.0)
-quant_config = Int4WeightOnlyConfig(group_size=128)
+quant_config = Float8DynamicActivationFloat8WeightConfig()
+# or float8 weight only quantization
+# quant_config = Float8WeightOnlyConfig()
 quantization_config = TorchAoConfig(quant_type=quant_config)
 
 # Load and quantize the model
 quantized_model = AutoModelForCausalLM.from_pretrained(
-    "meta-llama/Meta-Llama-3-8B",
+    "meta-llama/Llama-3.1-8B-Instruct",
     torch_dtype="auto",
     device_map="auto",
     quantization_config=quantization_config
 )
 
-tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B")
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
 input_text = "What are we having for dinner?"
 input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
 
@@ -62,22 +112,302 @@ input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
 output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
 print(tokenizer.decode(output[0], skip_special_tokens=True))
 ```
+</hfoption>
+<hfoption id="int4-weight-only">
 
-## Available Quantization Schemes
+```py
+import torch
+from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+from torchao.quantization import GemliteUIntXWeightOnlyConfig
 
-TorchAO provides a variety of quantization configurations:
+# We integrated with gemlite, which optimizes for batch size N on A100 and H100
+quant_config = GemliteUIntXWeightOnlyConfig(group_size=128)
+quantization_config = TorchAoConfig(quant_type=quant_config)
 
-- `Int4WeightOnlyConfig`
-- `Int8WeightOnlyConfig`
-- `Int8DynamicActivationInt8WeightConfig`
-- `Float8WeightOnlyConfig`
+# Load and quantize the model
+quantized_model = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    torch_dtype="auto",
+    device_map="auto",
+    quantization_config=quantization_config
+)
 
-Each configuration can be further customized with parameters such as `group_size`, `scheme`, and `layout` to optimize for specific hardware and model architectures.
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
 
-For a complete list of available configurations, see our [quantization API documentation](https://github.com/pytorch/ao/blob/main/torchao/quantization/quant_api.py).
+# auto-compile the quantized model with `cache_implementation="static"` to get speed up
+output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+</hfoption>
+</hfoptions>
+
+### A100 GPU
+<hfoptions id="examples-A100-GPU">
+<hfoption id="int8-dynamic-and-weight-only">
+```py
+import torch
+from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+from torchao.quantization import Int8DynamicActivationInt8WeightConfig
+
+quant_config = Int8DynamicActivationInt8WeightConfig()
+# or int8 weight only quantization
+# quant_config = Int8WeightOnlyConfig()
+quantization_config = TorchAoConfig(quant_type=quant_config)
+
+# Load and quantize the model
+quantized_model = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    torch_dtype="auto",
+    device_map="auto",
+    quantization_config=quantization_config
+)
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+# auto-compile the quantized model with `cache_implementation="static"` to get speed up
+output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+</hfoption>
+
+<hfoption id="int4-weight-only">
+
+```py
+import torch
+from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+from torchao.quantization import GemliteUIntXWeightOnlyConfig
+
+# For batch size N, we recommend gemlite, which may require autotuning
+# default is 4 bit, 8 bit is also supported by passing `bit_width=8`
+quant_config = GemliteUIntXWeightOnlyConfig(group_size=128)
+
+# For batch size 1, we also have custom tinygemm kernel that's only optimized for this
+# We can set `use_hqq` to `True` for better accuracy
+# quant_config = Int4WeightOnlyConfig(group_size=128, use_hqq=True)
+
+quantization_config = TorchAoConfig(quant_type=quant_config)
+
+# Load and quantize the model
+quantized_model = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    torch_dtype="auto",
+    device_map="auto",
+    quantization_config=quantization_config
+)
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+# auto-compile the quantized model with `cache_implementation="static"` to get speed up
+output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+</hfoption>
+</hfoptions>
+
+### CPU
+<hfoptions id="examples-CPU">
+<hfoption id="int8-dynamic-and-weight-only">
+```py
+import torch
+from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+from torchao.quantization import Int8DynamicActivationInt8WeightConfig
+
+quant_config = Int8DynamicActivationInt8WeightConfig()
+# quant_config = Int8WeightOnlyConfig()
+quantization_config = TorchAoConfig(quant_type=quant_config)
+
+# Load and quantize the model
+quantized_model = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    torch_dtype="auto",
+    device_map="cpu",
+    quantization_config=quantization_config
+)
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt")
+
+# auto-compile the quantized model with `cache_implementation="static"` to get speed up
+output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+</hfoption>
+<hfoption id="int4-weight-only">
+
+> [!TIP]
+> Run the quantized model on a CPU by changing `device_map` to `"cpu"` and `layout` to `Int4CPULayout()`.
+
+```py
+import torch
+from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+from torchao.quantization import Int4WeightOnlyConfig
+from torchao.dtypes import Int4CPULayout
+
+quant_config = Int4WeightOnlyConfig(group_size=128, layout=Int4CPULayout())
+quantization_config = TorchAoConfig(quant_type=quant_config)
+
+# Load and quantize the model
+quantized_model = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    torch_dtype="auto",
+    device_map="cpu",
+    quantization_config=quantization_config
+)
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt")
+
+# auto-compile the quantized model with `cache_implementation="static"` to get speed up
+output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+</hfoption>
+</hfoptions>
+
+### Autoquant
+
+If you want to automatically choose a quantization type for quantizable layers (`nn.Linear`) you can use the [autoquant](https://pytorch.org/ao/stable/generated/torchao.quantization.autoquant.html#torchao.quantization.autoquant) API.
+
+The `autoquant` API automatically chooses a quantization type by micro-benchmarking on input type and shape and compiling a single linear layer.
+
+Note: autoquant is for GPU only right now.
+
+Create a [`TorchAoConfig`] and set to `"autoquant"`. Set the `cache_implementation` to `"static"` to automatically [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) the forward method. Finally, call `finalize_autoquant` on the quantized model to finalize the quantization and log the input shapes.
+
+
+```py
+import torch
+from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+
+quantization_config = TorchAoConfig("autoquant", min_sqnr=None)
+quantized_model = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    torch_dtype="auto",
+    device_map="auto",
+    quantization_config=quantization_config
+)
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+# auto-compile the quantized model with `cache_implementation="static"` to get speed up
+output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
+# explicitly call `finalize_autoquant` (may be refactored and removed in the future)
+quantized_model.finalize_autoquant()
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+
+
+## Serialization
+
+torchao implements [torch.Tensor subclasses](https://pytorch.org/docs/stable/notes/extending.html#subclassing-torch-tensor) for maximum flexibility in supporting new quantized torch.Tensor formats. [Safetensors](https://huggingface.co/docs/safetensors/en/index) serialization and deserialization does not work with torchao.
+
+To avoid arbitrary user code execution, torchao sets `weights_only=True` in [torch.load](https://pytorch.org/docs/stable/generated/torch.load.html) to ensure only tensors are loaded. Any known user functions can be whitelisted with [add_safe_globals](https://pytorch.org/docs/stable/notes/serialization.html#torch.serialization.add_safe_globals).
+
+<hfoptions id="serialization-examples">
+<hfoption id="save-locally">
+```py
+# don't serialize model with Safetensors
+output_dir = "llama3-8b-int4wo-128"
+quantized_model.save_pretrained("llama3-8b-int4wo-128", safe_serialization=False)
+```
+</hfoption>
+<hfoption id="push-to-huggingface-hub">
+```py
+# don't serialize model with Safetensors
+USER_ID = "your_huggingface_user_id"
+REPO_ID = "llama3-8b-int4wo-128"
+quantized_model.push_to_hub(f"{USER_ID}/llama3-8b-int4wo-128", safe_serialization=False)
+tokenizer.push_to_hub(f"{USER_ID}/llama3-8b-int4wo-128")
+```
+</hfoption>
+</hfoptions>
+
+
+## Loading quantized models
+
+Loading a quantized model depends on the quantization scheme. For quantization schemes, like int8 and float8, you can quantize the model on any device and also load it on any device. The example below demonstrates quantizing a model on the CPU and then loading it on CUDA.
+```py
+import torch
+from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+from torchao.quantization import Int8WeightOnlyConfig
+
+quant_config = Int8WeightOnlyConfig(group_size=128)
+quantization_config = TorchAoConfig(quant_type=quant_config)
+
+# Load and quantize the model
+quantized_model = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    torch_dtype="auto",
+    device_map="cpu",
+    quantization_config=quantization_config
+)
+# save the quantized model
+output_dir = "llama-3.1-8b-torchao-int8-cuda"
+quantized_model.save_pretrained(output_dir, safe_serialization=False)
+
+# reload the quantized model
+reloaded_model = AutoModelForCausalLM.from_pretrained(
+    output_dir, 
+    device_map="auto", 
+    torch_dtype=torch.bfloat16
+)
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+output = reloaded_model.generate(**input_ids, max_new_tokens=10)
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+
+```
+For int4, the model can only be loaded on the same device it was quantized on because the layout is specific to the device. The example below demonstrates quantizing and loading a model on the CPU.
+
+```py
+import torch
+from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+from torchao.quantization import Int4WeightOnlyConfig
+from torchao.dtypes import Int4CPULayout
+
+quant_config = Int4WeightOnlyConfig(group_size=128, layout=Int4CPULayout())
+quantization_config = TorchAoConfig(quant_type=quant_config)
+
+# Load and quantize the model
+quantized_model = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    torch_dtype="auto",
+    device_map="cpu",
+    quantization_config=quantization_config
+)
+# save the quantized model
+output_dir = "llama-3.1-8b-torchao-int4-cpu"
+quantized_model.save_pretrained(output_dir, safe_serialization=False)
+
+# reload the quantized model
+reloaded_model = AutoModelForCausalLM.from_pretrained(
+    output_dir, 
+    device_map="cpu", 
+    torch_dtype=torch.bfloat16
+)
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt")
+
+output = reloaded_model.generate(**input_ids, max_new_tokens=10)
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+
+```
+
+## ⚠️ Deprecation Notice
 
-> **⚠️ DEPRECATION WARNING**
->
 > Starting with version 0.10.0, the string-based API for quantization configuration (e.g., `TorchAoConfig("int4_weight_only", group_size=128)`) is **deprecated** and will be removed in a future release.
 >
 > Please use the new `AOBaseConfig`-based approach instead:
@@ -94,7 +424,7 @@ For a complete list of available configurations, see our [quantization API docum
 >
 > The new API offers greater flexibility, better type safety, and access to the full range of features available in torchao.
 >
-> ## Migration Guide
+> [Migration Guide](#migration-guide)
 >
 > Here's how to migrate from common string identifiers to their `AOBaseConfig` equivalents:
 >
@@ -107,30 +437,10 @@ For a complete list of available configurations, see our [quantization API docum
 > All configuration objects accept parameters for customization (e.g., `group_size`, `scheme`, `layout`).
 
 
-Below is the API for for torchao < `0.9.0`
 
-```py
-import torch
-from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+## Resources
 
-quantization_config = TorchAoConfig("int4_weight_only", group_size=128)
-quantized_model = AutoModelForCausalLM.from_pretrained(
-    "meta-llama/Meta-Llama-3-8B",
-    torch_dtype="auto",
-    device_map="auto",
-    quantization_config=quantization_config
-)
-
-tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B")
-input_text = "What are we having for dinner?"
-input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
-
-# auto-compile the quantized model with `cache_implementation="static"` to get speed up
-output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
-print(tokenizer.decode(output[0], skip_special_tokens=True))
-```
-
-Run the code below to benchmark the quantized models performance.
+For a better sense of expected performance, view the [benchmarks](https://github.com/pytorch/ao/tree/main/torchao/quantization#benchmarks) for various models with CUDA and XPU backends. You can also run the code below to benchmark a model yourself.
 
 ```py
 from torch._inductor.utils import do_bench_using_profiling
@@ -153,76 +463,8 @@ print("bf16 model:", benchmark_fn(bf16_model.generate, **input_ids, max_new_toke
 > [!TIP]
 > For best performance, you can use recommended settings by calling `torchao.quantization.utils.recommended_inductor_config_setter()`
 
-</hfoption>
-<hfoption id="automatic">
-
-The [autoquant](https://pytorch.org/ao/stable/generated/torchao.quantization.autoquant.html#torchao.quantization.autoquant) API automatically chooses a quantization type for quantizable layers (`nn.Linear`) by micro-benchmarking on input type and shape and compiling a single linear layer.
-
-Create a [`TorchAoConfig`] and set to `"autoquant"`. Set the `cache_implementation` to `"static"` to automatically [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) the forward method. Finally, call `finalize_autoquant` on the quantized model to finalize the quantization and log the input shapes.
-
-> [!TIP]
-> Run the quantized model on a CPU by changing `device_map` to `"cpu"` and `layout` to `Int4CPULayout()`. This is only available in torchao 0.8.0+.
-
-```py
-import torch
-from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
-
-quantization_config = TorchAoConfig("autoquant", min_sqnr=None)
-quantized_model = AutoModelForCausalLM.from_pretrained(
-    "meta-llama/Meta-Llama-3-8B",
-    torch_dtype="auto",
-    device_map="auto",
-    quantization_config=quantization_config
-)
-
-tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B")
-input_text = "What are we having for dinner?"
-input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
-
-# auto-compile the quantized model with `cache_implementation="static"` to get speed up
-output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
-# explicitly call `finalize_autoquant` (may be refactored and removed in the future)
-quantized_model.finalize_autoquant()
-print(tokenizer.decode(output[0], skip_special_tokens=True))
-```
-
-Run the code below to benchmark the quantized models performance.
-
-```py
-from torch._inductor.utils import do_bench_using_profiling
-from typing import Callable
-
-def benchmark_fn(func: Callable, *args, **kwargs) -> float:
-    """Thin wrapper around do_bench_using_profiling"""
-    no_args = lambda: func(*args, **kwargs)
-    time = do_bench_using_profiling(no_args)
-    return time * 1e3
-
-MAX_NEW_TOKENS = 1000
-print("autoquantized model:", benchmark_fn(quantized_model.generate, **input_ids, max_new_tokens=MAX_NEW_TOKENS, cache_implementation="static"))
-
-bf16_model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype=torch.bfloat16)
-output = bf16_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static") # auto-compile
-print("bf16 model:", benchmark_fn(bf16_model.generate, **input_ids, max_new_tokens=MAX_NEW_TOKENS, cache_implementation="static"))
-```
-
-</hfoption>
-</hfoptions>
-
-## Serialization
-
-torchao implements [torch.Tensor subclasses](https://pytorch.org/docs/stable/notes/extending.html#subclassing-torch-tensor) for maximum flexibility in supporting new quantized torch.Tensor formats. [Safetensors](https://huggingface.co/docs/safetensors/en/index) serialization and deserialization does not work with torchao.
-
-To avoid arbitrary user code execution, torchao sets `weights_only=True` in [torch.load](https://pytorch.org/docs/stable/generated/torch.load.html) to ensure only tensors are loaded. Any known user functions can be whitelisted with [add_safe_globals](https://pytorch.org/docs/stable/notes/serialization.html#torch.serialization.add_safe_globals).
-
-```py
-# don't serialize model with Safetensors
-output_dir = "llama3-8b-int4wo-128"
-quantized_model.save_pretrained("llama3-8b-int4wo-128", safe_serialization=False)
-```
-
-## Resources
-
-For a better sense of expected performance, view the [benchmarks](https://github.com/pytorch/ao/tree/main/torchao/quantization#benchmarks) for various models with CUDA and XPU backends.
-
 Refer to [Other Available Quantization Techniques](https://github.com/pytorch/ao/tree/main/torchao/quantization#other-available-quantization-techniques) for more examples and documentation.
+
+## Issues
+
+If you encounter any issues with the Transformers integration, please open an issue on the [Transformers](https://github.com/huggingface/transformers/issues) repository. For issues directly related to torchao, please open an issue on the [torchao](https://github.com/pytorch/ao/issues) repository.

docs/source/en/tasks/image_text_to_text.md

@@ -160,7 +160,48 @@ outputs[0]["generated_text"]
 #  with a yellow center in the foreground. The flower is surrounded by red and white flowers with green stems
 ```
 
-## Streaming
+If you prefer, you can also load the images separately and pass them to the pipeline like so:
+
+```python
+pipe = pipeline("image-text-to-text", model="HuggingFaceTB/SmolVLM-256M-Instruct")
+
+img_urls = [
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.png",
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg",
+]
+images = [
+    Image.open(requests.get(img_urls[0], stream=True).raw),
+    Image.open(requests.get(img_urls[1], stream=True).raw),
+]
+
+messages = [
+    {
+        "role": "user",
+        "content": [
+            {"type": "image"},
+            {"type": "image"},
+            {"type": "text", "text": "What do you see in these images?"},
+        ],
+    }
+]
+outputs = pipe(text=messages, images=images, max_new_tokens=50, return_full_text=False)
+outputs[0]["generated_text"]
+" In the first image, there are two cats sitting on a plant. In the second image, there are flowers with a pinkish hue."
+```
+
+The images will still be included in the `"input_text"` field of the output:
+
+```python
+outputs[0]['input_text']
+"""
+[{'role': 'user',
+  'content': [{'type': 'image',
+    'image': <PIL.PngImagePlugin.PngImageFile image mode=RGBA size=622x412>},
+   {'type': 'image',
+    'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=5184x3456>},
+   {'type': 'text', 'text': 'What do you see in these images?'}]}]## Streaming
+"""
+```
 
 We can use [text streaming](./generation_strategies#streaming) for a better generation experience. Transformers supports streaming with the [`TextStreamer`] or [`TextIteratorStreamer`] classes. We will use the [`TextIteratorStreamer`] with IDEFICS-8B.
 

docs/source/en/tasks/visual_document_retrieval.md

@@ -0,0 +1,144 @@
+<!--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.
+
+-->
+# Visual document retrieval
+
+Documents can contain multimodal data if they include charts, tables, and visuals in addition to text. Retrieving information from these documents is challenging because text retrieval models alone can't handle visual data and image retrieval models lack the granularity and document processing capabilities.
+
+Visual document retrieval can help retrieve information from all types of documents, including multimodal retrieval augmented generation (RAG). These models accept documents (as images) and texts and calculates the similarity scores between them.
+
+This guide demonstrates how to index and retrieve documents with [ColPali](../model_doc/colpali).  
+
+> [!TIP]
+> For large scale use cases, you may want to index and retrieve documents with a vector database.
+
+Make sure Transformers and Datasets is installed.
+
+```bash
+pip install -q datasets transformers
+```
+
+We will index a dataset of documents related to UFO sightings. We filter the examples where our column of interest is missing. It contains several columns, we are interested in the column `specific_detail_query` where it contains short summary of the document, and `image` column that contains our documents.
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("davanstrien/ufo-ColPali")
+dataset = dataset["train"]
+dataset = dataset.filter(lambda example: example["specific_detail_query"] is not None)
+dataset
+```
+```
+Dataset({
+    features: ['image', 'raw_queries', 'broad_topical_query', 'broad_topical_explanation', 'specific_detail_query', 'specific_detail_explanation', 'visual_element_query', 'visual_element_explanation', 'parsed_into_json'],
+    num_rows: 2172
+})
+```
+
+Let's load the model and the tokenizer.
+
+```python
+import torch
+from transformers import ColPaliForRetrieval, ColPaliProcessor
+
+model_name = "vidore/colpali-v1.2-hf"
+
+processor = ColPaliProcessor.from_pretrained(model_name)
+
+model = ColPaliForRetrieval.from_pretrained(
+    model_name,
+    torch_dtype=torch.bfloat16,
+    device_map="cuda",
+).eval()
+```
+
+Pass the text query to the processor and return the indexed text embeddings from the model. For image-to-text search, replace the `text` parameter in [`ColPaliProcessor`] with the `images` parameter to pass images.
+
+```python
+inputs = processor(text="a document about Mars expedition").to("cuda")
+with torch.no_grad():
+  text_embeds = model(**inputs, return_tensors="pt").embeddings
+```
+
+Index the images offline, and during inference, return the query text embeddings to get its closest image embeddings.
+
+Store the image and image embeddings by writing them to the dataset with [`~datasets.Dataset.map`] as shown below. Add an `embeddings` column that contains the indexed embeddings. ColPali embeddings take up a lot of storage, so remove them from the GPU and store them in the CPU as NumPy vectors.
+
+```python
+ds_with_embeddings = dataset.map(lambda example: {'embeddings': model(**processor(images=example["image"]).to("cuda"), return_tensors="pt").embeddings.to(torch.float32).detach().cpu().numpy()})
+```
+
+For online inference, create a function to search the image embeddings in batches and retrieve the k-most relevant images. The function below returns the indices in the dataset and their scores for a given indexed dataset, text embeddings, number of top results, and the batch size.
+
+```python
+def find_top_k_indices_batched(dataset, text_embedding, processor, k=10, batch_size=4):
+    scores_and_indices = []
+
+    for start_idx in range(0, len(dataset), batch_size):
+
+        end_idx = min(start_idx + batch_size, len(dataset))
+        batch = dataset[start_idx:end_idx]        
+        batch_embeddings = [torch.tensor(emb[0], dtype=torch.float32) for emb in batch["embeddings"]]
+        scores = processor.score_retrieval(text_embedding.to("cpu").to(torch.float32), batch_embeddings)
+
+        if hasattr(scores, "tolist"):
+            scores = scores.tolist()[0]
+
+        for i, score in enumerate(scores):
+            scores_and_indices.append((score, start_idx + i))
+
+    sorted_results = sorted(scores_and_indices, key=lambda x: -x[0])
+
+    topk = sorted_results[:k]
+    indices = [idx for _, idx in topk]
+    scores = [score for score, _ in topk]
+
+    return indices, scores
+```
+
+Generate the text embeddings and pass them to the function above to return the dataset indices and scores.
+
+```python
+with torch.no_grad():
+  text_embeds = model(**processor(text="a document about Mars expedition").to("cuda"), return_tensors="pt").embeddings
+indices, scores = find_top_k_indices_batched(ds_with_embeddings, text_embeds, processor, k=3, batch_size=4)
+print(indices, scores)
+```
+
+```
+([440, 442, 443],
+ [14.370786666870117,
+  13.675487518310547,
+  12.9899320602417])
+```
+
+Display the images to view the Mars related documents.
+
+```python
+for i in indices:
+  display(dataset[i]["image"])
+```
+
+<div style="display: flex; align-items: center;">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/doc_1.png" 
+         alt="Document 1" 
+         style="height: 200px; object-fit: contain; margin-right: 10px;">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/doc_2.png" 
+         alt="Document 2" 
+         style="height: 200px; object-fit: contain;">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/doc_3.png" 
+         alt="Document 3" 
+         style="height: 200px; object-fit: contain;">
+</div>

docs/source/ja/internal/modeling_utils.md

@@ -25,23 +25,6 @@ rendered properly in your Markdown viewer.
 
 [[autodoc]] pytorch_utils.Conv1D
 
-[[autodoc]] modeling_utils.PoolerStartLogits
-    - forward
-
-[[autodoc]] modeling_utils.PoolerEndLogits
-    - forward
-
-[[autodoc]] modeling_utils.PoolerAnswerClass
-    - forward
-
-[[autodoc]] modeling_utils.SquadHeadOutput
-
-[[autodoc]] modeling_utils.SQuADHead
-    - forward
-
-[[autodoc]] modeling_utils.SequenceSummary
-    - forward
-
 ## PyTorch Helper Functions
 
 [[autodoc]] pytorch_utils.apply_chunking_to_forward

docs/source/ja/model_doc/bridgetower.md

@@ -144,6 +144,11 @@ BridgeTower は、ビジュアル エンコーダー、テキスト エンコー
 [[autodoc]] BridgeTowerImageProcessor
     - preprocess
 
+## BridgeTowerImageProcessorFast
+
+[[autodoc]] BridgeTowerImageProcessorFast
+    - preprocess
+
 ## BridgeTowerProcessor
 
 [[autodoc]] BridgeTowerProcessor

docs/source/ja/model_doc/chinese_clip.md

@@ -86,6 +86,11 @@ Chinese-CLIP モデルは、[OFA-Sys](https://huggingface.co/OFA-Sys) によっ
 [[autodoc]] ChineseCLIPImageProcessor
     - preprocess
 
+## ChineseCLIPImageProcessorFast
+
+[[autodoc]] ChineseCLIPImageProcessorFast
+    - preprocess
+
 ## ChineseCLIPFeatureExtractor
 
 [[autodoc]] ChineseCLIPFeatureExtractor

docs/source/ja/model_doc/conditional_detr.md

@@ -43,6 +43,11 @@ alt="描画" width="600"/>
 
 [[autodoc]] ConditionalDetrImageProcessor
     - preprocess
+
+## ConditionalDetrImageProcessorFast
+
+[[autodoc]] ConditionalDetrImageProcessorFast
+    - preprocess
     - post_process_object_detection
     - post_process_instance_segmentation
     - post_process_semantic_segmentation

docs/source/ko/_toctree.yml

@@ -77,6 +77,8 @@
         title: 이미지 특징 추출
       - local: tasks/mask_generation
         title: 마스크 생성
+      - local: tasks/keypoint_detection
+        title: 키포인트 탐지
       - local: tasks/knowledge_distillation_for_image_classification
         title: 컴퓨터 비전(이미지 분류)를 위한 지식 증류(knowledge distillation)
     title: 컴퓨터 비전
@@ -480,8 +482,8 @@
         title: (번역중) RemBERT
       - local: in_translation
         title: (번역중) RetriBERT
-      - local: in_translation
-        title: (번역중) RoBERTa
+      - local: model_doc/roberta
+        title: RoBERTa
       - local: in_translation
         title: (번역중) RoBERTa-PreLayerNorm
       - local: in_translation
@@ -720,6 +722,8 @@
         title: Qwen2VL
       - local: in_translation
         title: (번역중) Segment Anything
+      - local: model_doc/siglip
+        title: SigLIP
       - local: in_translation
         title: (번역중) Speech Encoder Decoder Models
       - local: in_translation

docs/source/ko/internal/modeling_utils.md

@@ -25,23 +25,6 @@ rendered properly in your Markdown viewer.
 
 [[autodoc]] pytorch_utils.Conv1D
 
-[[autodoc]] modeling_utils.PoolerStartLogits
-   - forward
-
-[[autodoc]] modeling_utils.PoolerEndLogits
-   - forward
-
-[[autodoc]] modeling_utils.PoolerAnswerClass
-   - forward
-
-[[autodoc]] modeling_utils.SquadHeadOutput
-
-[[autodoc]] modeling_utils.SQuADHead
-   - forward
-
-[[autodoc]] modeling_utils.SequenceSummary
-   - forward
-
 ## PyTorch 헬퍼(helper) 함수 [[transformers.apply_chunking_to_forward]]
 
 [[autodoc]] pytorch_utils.apply_chunking_to_forward

docs/source/ko/model_doc/roberta.md

@@ -0,0 +1,230 @@
+<!--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.
+
+-->
+
+# RoBERTa[[roberta]]
+
+<div class="flex flex-wrap space-x-1">
+<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+<img alt="TensorFlow" src="https://img.shields.io/badge/TensorFlow-FF6F00?style=flat&logo=tensorflow&logoColor=white">
+<img alt="Flax" src="https://img.shields.io/badge/Flax-29a79b.svg?style=flat&logo=data:image/png;base64,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
+">
+<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+</div>
+
+## 개요[[overview]]
+
+RoBERTa 모델은 Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov가 제안한 논문 [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692)에서 소개되었습니다. 이 모델은 2018년에 구글에서 발표한 BERT 모델을 기반으로 합니다.
+
+RoBERTa는 BERT를 기반으로 하며, 주요 하이퍼파라미터를 수정하고, 사전 학습 단계에서 다음 문장 예측(Next Sentence Prediction)을 제거했으며, 훨씬 더 큰 미니 배치 크기와 학습률을 사용하여 학습을 진행했습니다.
+
+해당 논문의 초록입니다:
+
+*언어 모델 사전 학습은 성능을 크게 향상시켰지만, 서로 다른 접근 방식을 면밀히 비교하는 것은 어렵습니다. 학습은 계산 비용이 많이 들고, 종종 크기가 서로 다른 비공개 데이터셋에서 수행되며, 본 논문에서 보여주듯이 하이퍼파라미터 선택이 최종 성능에 큰 영향을 미칩니다. 우리는 BERT 사전 학습(Devlin et al., 2019)에 대한 재현 연구를 수행하여, 여러 핵심 하이퍼파라미터와 학습 데이터 크기의 영향을 면밀히 측정하였습니다. 그 결과, BERT는 충분히 학습되지 않았으며, 이후 발표된 모든 모델의 성능을 맞추거나 능가할 수 있음을 발견했습니다. 우리가 제안한 최상의 모델은 GLUE, RACE, SQuAD에서 최고 성능(state-of-the-art)을 달성했습니다. 이 결과는 지금까지 간과되어 온 설계 선택의 중요성을 강조하며, 최근 보고된 성능 향상의 근원이 무엇인지에 대한 의문을 제기합니다. 우리는 본 연구에서 사용한 모델과 코드를 공개합니다.*
+
+이 모델은 [julien-c](https://huggingface.co/julien-c)가 기여하였습니다. 원본 코드는 [여기](https://github.com/pytorch/fairseq/tree/master/examples/roberta)에서 확인할 수 있습니다.
+
+## 사용 팁[[usage-tips]]
+
+- 이 구현은 [`BertModel`]과 동일하지만, 임베딩 부분에 약간의 수정이 있으며 RoBERTa 사전학습 모델에 맞게 설정되어 있습니다.
+- RoBERTa는 BERT와 동일한 아키텍처를 가지고 있지만, 토크나이저로 바이트 수준 BPE(Byte-Pair Encoding, GPT-2와 동일)를 사용하고, 사전학습 방식이 다릅니다.
+- RoBERTa는 `token_type_ids`를 사용하지 않기 때문에, 어떤 토큰이 어떤 문장(segment)에 속하는지 별도로 표시할 필요가 없습니다. 문장 구분은 분리 토큰 `tokenizer.sep_token`(또는 `</s>`)을 사용해 나누면 됩니다.
+- RoBERTa는 BERT와 유사하지만, 더 나은 사전학습 기법을 사용합니다:
+
+    * 동적 마스킹: RoBERTa는 매 에폭마다 토큰을 다르게 마스킹하는 반면, BERT는 한 번만 마스킹합니다.
+    * 문장 패킹: 여러 문장을 최대 512 토큰까지 함께 패킹하여, 문장이 여러 문서에 걸쳐 있을 수도 있습니다.
+    * 더 큰 배치 사이즈: 학습 시 더 큰 미니배치를 사용합니다.
+    * 바이트 수준 BPE 어휘: 문자를 단위로 하지 않고 바이트 단위로 BPE를 적용하여 유니코드 문자를 더 유연하게 처리할 수 있습니다.
+
+- [CamemBERT](camembert)은 RoBERTa를 기반으로 한 래퍼 모델입니다. 사용 예제는 해당 모델 페이지를 참고하세요.
+
+## 자료[[resources]]
+
+RoBERTa를 처음 다룰 때 도움이 되는 Hugging Face 공식 자료와 커뮤니티 자료(🌎 아이콘으로 표시됨) 목록입니다. 이 목록에 자료를 추가하고 싶다면 언제든지 Pull Request를 보내주세요! 저희가 검토 후 반영하겠습니다. 추가하려는 자료는 기존 자료를 단순히 복제하는 것이 아닌, 새롭거나 유의미한 내용을 포함하고 있는 것이 좋습니다.
+
+<PipelineTag pipeline="text-classification"/>
+
+- RoBERTa와 [Inference API](https://huggingface.co/inference-api)를 활용한 [트위터 감성 분석 시작하기](https://huggingface.co/blog/sentiment-analysis-twitter) 블로그 포스트.
+- RoBERTa를 활용한 [Kili 및 Hugging Face AutoTrain을 이용한 의견 분류](https://huggingface.co/blog/opinion-classification-with-kili)에 관한 블로그 포스트.
+- [감성 분석을 위한 RoBERTa 미세조정](https://colab.research.google.com/github/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb)을 하는 방법에 대한 노트북.🌎
+- ['RobertaForSequenceClassification']은 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/pytorch/text-classification)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification.ipynb)에서 지원됩니다.
+- [`TFRobertaForSequenceClassification`]는 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/text-classification)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification-tf.ipynb)에서 지원됩니다.
+- [`FlaxRobertaForSequenceClassification`]는 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/flax/text-classification)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification_flax.ipynb)에서 지원됩니다.
+- [텍스트 분류 작업 가이드](../tasks/sequence_classification)
+
+<PipelineTag pipeline="token-classification"/>
+
+- [`RobertaForTokenClassification`]은 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/pytorch/token-classification)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/token_classification.ipynb)에서 지원됩니다.
+- [`TFRobertaForTokenClassification`]은 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/token-classification)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/token_classification-tf.ipynb)에서 지원됩니다.
+- [`FlaxRobertaForTokenClassification`]는 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/flax/token-classification)에서 지원됩니다.
+- 🤗 Hugging Face 코스의 [토큰 분류 챕터](https://huggingface.co/course/chapter7/2?fw=pt)
+- [토큰 분류 작업 가이드](../tasks/token_classification)
+
+<PipelineTag pipeline="fill-mask"/>
+
+- RoBERTa를 활용한 [Transformers와 Tokenizers를 활용한 새로운 언어 모델을 처음부터 학습하는 방법](https://huggingface.co/blog/how-to-train)에 대한 블로그 포스트.
+- [`RobertaForMaskedLM`]은 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling#robertabertdistilbert-and-masked-language-modeling)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling.ipynb)에서 지원됩니다.
+- [`TFRobertaForMaskedLM`]은 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/language-modeling#run_mlmpy)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling-tf.ipynb)에서 지원됩니다.
+- [`FlaxRobertaForMaskedLM`]은 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/flax/language-modeling#masked-language-modeling)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/masked_language_modeling_flax.ipynb)에서 지원됩니다.
+- 🤗 Hugging Face 코스의 [마스킹 언어 모델링 챕터](https://huggingface.co/course/chapter7/3?fw=pt)
+- [마스킹 언어 모델링 작업 가이드](../tasks/masked_language_modeling)
+
+<PipelineTag pipeline="question-answering"/>
+
+- RoBERTa를 활용한 질문 응답 작업에서의 [Optimum과 Transformers 파이프라인을 이용한 추론 가속화](https://huggingface.co/blog/optimum-inference)에 대한 블로그 포스트.
+- [`RobertaForQuestionAnswering`]은 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/pytorch/question-answering)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering.ipynb)에서 지원됩니다.
+- [`TFRobertaForQuestionAnswering`]은 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/question-answering)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering-tf.ipynb)에서 지원됩니다.
+- [`FlaxRobertaForQuestionAnswering`]은 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/flax/question-answering)에서 지원됩니다.
+- 🤗 Hugging Face 코스의 [질의응답 챕터](https://huggingface.co/course/chapter7/7?fw=pt)
+- [질의응답 작업 가이드](../tasks/question_answering)
+
+**다중 선택**
+- [`RobertaForMultipleChoice`]는 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/pytorch/multiple-choice)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/multiple_choice.ipynb)에서 지원됩니다.
+- [`TFRobertaForMultipleChoice`]는 [예제 스크립트](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/multiple-choice)와 [노트북](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/multiple_choice-tf.ipynb)에서 지원됩니다.
+- [다중 선택 작업 가이드](../tasks/multiple_choice)
+
+## RobertaConfig
+
+[[autodoc]] RobertaConfig
+
+## RobertaTokenizer
+
+[[autodoc]] RobertaTokenizer
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - save_vocabulary
+
+## RobertaTokenizerFast
+
+[[autodoc]] RobertaTokenizerFast
+    - build_inputs_with_special_tokens
+
+<frameworkcontent>
+<pt>
+
+## RobertaModel
+
+[[autodoc]] RobertaModel
+    - forward
+
+## RobertaForCausalLM
+
+[[autodoc]] RobertaForCausalLM
+    - forward
+
+## RobertaForMaskedLM
+
+[[autodoc]] RobertaForMaskedLM
+    - forward
+
+## RobertaForSequenceClassification
+
+[[autodoc]] RobertaForSequenceClassification
+    - forward
+
+## RobertaForMultipleChoice
+
+[[autodoc]] RobertaForMultipleChoice
+    - forward
+
+## RobertaForTokenClassification
+
+[[autodoc]] RobertaForTokenClassification
+    - forward
+
+## RobertaForQuestionAnswering
+
+[[autodoc]] RobertaForQuestionAnswering
+    - forward
+
+</pt>
+<tf>
+
+## TFRobertaModel
+
+[[autodoc]] TFRobertaModel
+    - call
+
+## TFRobertaForCausalLM
+
+[[autodoc]] TFRobertaForCausalLM
+    - call
+
+## TFRobertaForMaskedLM
+
+[[autodoc]] TFRobertaForMaskedLM
+    - call
+
+## TFRobertaForSequenceClassification
+
+[[autodoc]] TFRobertaForSequenceClassification
+    - call
+
+## TFRobertaForMultipleChoice
+
+[[autodoc]] TFRobertaForMultipleChoice
+    - call
+
+## TFRobertaForTokenClassification
+
+[[autodoc]] TFRobertaForTokenClassification
+    - call
+
+## TFRobertaForQuestionAnswering
+
+[[autodoc]] TFRobertaForQuestionAnswering
+    - call
+
+</tf>
+<jax>
+
+## FlaxRobertaModel
+
+[[autodoc]] FlaxRobertaModel
+    - __call__
+
+## FlaxRobertaForCausalLM
+
+[[autodoc]] FlaxRobertaForCausalLM
+    - __call__
+
+## FlaxRobertaForMaskedLM
+
+[[autodoc]] FlaxRobertaForMaskedLM
+    - __call__
+
+## FlaxRobertaForSequenceClassification
+
+[[autodoc]] FlaxRobertaForSequenceClassification
+    - __call__
+
+## FlaxRobertaForMultipleChoice
+
+[[autodoc]] FlaxRobertaForMultipleChoice
+    - __call__
+
+## FlaxRobertaForTokenClassification
+
+[[autodoc]] FlaxRobertaForTokenClassification
+    - __call__
+
+## FlaxRobertaForQuestionAnswering
+
+[[autodoc]] FlaxRobertaForQuestionAnswering
+    - __call__
+
+</jax>
+</frameworkcontent>

docs/source/ko/model_doc/siglip.md

@@ -0,0 +1,253 @@
+<!--Copyright 2023 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.
+
+-->
+
+# SigLIP[[siglip]]
+
+<div class="flex flex-wrap space-x-1">
+<img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-DE3412?style=flat&logo=pytorch&logoColor=white">
+<img alt="FlashAttention" src="https://img.shields.io/badge/%E2%9A%A1%EF%B8%8E%20FlashAttention-eae0c8?style=flat">
+<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
+</div>
+
+## 개요[[overview]]
+
+SigLIP 모델은 Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, Lucas Beyer의 [Sigmoid Loss for Language Image Pre-Training](https://arxiv.org/abs/2303.15343) 논문에서 제안되었습니다. SigLIP은 [CLIP](clip)에서 사용된 손실 함수를 간단한 쌍별 시그모이드 손실(pairwise sigmoid loss)로 대체할 것을 제안합니다. 이는 ImageNet에서 제로샷 분류 정확도 측면에서 더 나은 성능을 보입니다.
+
+논문의 초록은 다음과 같습니다:
+
+*우리는 언어-이미지 사전 학습(Language-Image Pre-training, SigLIP)을 위한 간단한 쌍별 시그모이드 손실을 제안합니다. 소프트맥스 정규화를 사용하는 표준 대조 학습과 달리, 시그모이드 손실은 이미지-텍스트 쌍에만 작용하며 정규화를 위해 쌍별 유사성의 전역적 관점을 필요로 하지 않습니다. 시그모이드 손실은 배치 크기를 더욱 확장할 수 있게 하는 동시에 작은 배치 크기에서도 더 나은 성능을 보입니다. Locked-image Tuning과 결합하여, 단 4개의 TPUv4 칩만으로 이틀 만에 84.5%의 ImageNet 제로샷 정확도를 달성하는 SigLiT 모델을 학습했습니다. 손실 함수에서 배치 크기를 분리함으로써 예제 대 쌍의 영향과 Negative 대 Positive 비율을 연구할 수 있게 되었습니다. 마지막으로, 우리는 배치 크기를 100만 개까지 극단적으로 늘려보았고, 배치 크기 증가의 이점이 빠르게 감소하며 32k의 더 합리적인 배치 크기로도 충분하다는 것을 발견했습니다.*
+
+## 사용 팁[[usage-tips]]
+
+- SigLIP의 사용법은 [CLIP](clip)과 유사합니다. 주요 차이점은 학습 손실 함수로, 배치 내 모든 이미지와 텍스트 간의 쌍별 유사성에 대한 전역적 관점이 필요하지 않습니다. 소프트맥스 대신 로짓에 시그모이드 활성화 함수를 적용해야 합니다.
+- 학습은 지원되지만 `torch.distributed` 유틸리티를 사용하지 않아 배치 크기의 확장성이 제한될 수 있습니다. 그러나 단일 노드 다중 GPU 설정에서는 DDP와 FDSP가 작동합니다.
+- 독립형 [`SiglipTokenizer`] 또는 [`SiglipProcessor`]를 사용할 때는 모델이 그렇게 학습되었으므로 `padding="max_length"`를 전달해야 합니다.
+- 파이프라인과 동일한 결과를 얻으려면 "This is a photo of {label}."의 프롬프트 템플릿을 사용해야 합니다.
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/siglip_table.jpeg"
+alt="drawing" width="600"/>
+
+<small> CLIP과 비교한 SigLIP 평가 결과. <a href="https://arxiv.org/abs/2303.15343">원본 논문</a>에서 발췌.</small>
+
+이 모델은 [nielsr](https://huggingface.co/nielsr)가 기여했습니다.
+원본 코드는 [여기](https://github.com/google-research/big_vision/tree/main)에서 찾을 수 있습니다.
+
+## 사용 예시[[usage-example]]
+
+SigLIP을 사용하는 방법에는 두 가지 주요 방법이 있습니다: 모든 복잡성을 추상화하는 파이프라인 API를 사용하거나, 직접 `SiglipModel` 클래스를 사용하는 방법입니다.
+
+### 파이프라인 API[[pipeline-API]]
+
+파이프라인을 사용하면 몇 줄의 코드로 모델을 사용할 수 있습니다:
+
+```python
+>>> from transformers import pipeline
+>>> from PIL import Image
+>>> import requests
+
+>>> # 파이프라인 로드
+>>> image_classifier = pipeline(task="zero-shot-image-classification", model="google/siglip-base-patch16-224")
+
+>>> # 이미지 로드
+>>> url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
+>>> image = Image.open(requests.get(url, stream=True).raw)
+
+>>> # 추론
+>>> candidate_labels = ["2 cats", "a plane", "a remote"]
+>>> outputs = image_classifier(image, candidate_labels=candidate_labels)
+>>> outputs = [{"score": round(output["score"], 4), "label": output["label"] } for output in outputs]
+>>> print(outputs)
+[{'score': 0.1979, 'label': '2 cats'}, {'score': 0.0, 'label': 'a remote'}, {'score': 0.0, 'label': 'a plane'}]
+```
+
+### 직접 모델 사용하기[[using-the-model-yourself]]
+
+전처리와 후처리를 직접 수행하려면 다음과 같이 하면 됩니다:
+
+```python
+>>> from PIL import Image
+>>> import requests
+>>> from transformers import AutoProcessor, AutoModel
+>>> import torch
+
+>>> model = AutoModel.from_pretrained("google/siglip-base-patch16-224")
+>>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
+
+>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+>>> image = Image.open(requests.get(url, stream=True).raw)
+
+>>> candidate_labels = ["2 cats", "2 dogs"]
+# 파이프라인 프롬프트 템플릿을 따라 동일한 결과를 얻습니다
+>>> texts = [f'This is a photo of {label}.' for label in candidate_labels]
+# 중요: 모델이 이렇게 학습되었으므로 `padding=max_length`를 전달합니다
+>>> inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt")
+
+>>> with torch.no_grad():
+...     outputs = model(**inputs)
+
+>>> logits_per_image = outputs.logits_per_image
+>>> probs = torch.sigmoid(logits_per_image) # 시그모이드 활성화 함수를 적용한 확률입니다
+>>> print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
+19.8% that image 0 is '2 cats'
+```
+
+## 리소스[[resources]]
+
+SigLIP을 시작하는 데 도움이 되는 공식 Hugging Face 및 커뮤니티(🌎로 표시) 리소스 목록입니다.
+
+- [제로샷 이미지 분류 작업 가이드](../tasks/zero_shot_image_classification)
+- SigLIP에 대한 데모 노트북은 [여기](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/SigLIP)에서 찾을 수 있습니다. 🌎
+
+여기에 포함될 리소스를 제출하는 데 관심이 있으시면 Pull Request를 열어주시면 검토하겠습니다! 리소스는 이상적으로 기존 리소스를 복제하는 대신 새로운 것을 보여주어야 합니다.
+
+
+## SigLIP과 Flash Attention 2 결합하기[[combining-siglip-with-flash-attention-2]]
+
+먼저 Flash Attention 2의 최신 버전을 설치해야 합니다.
+
+```bash
+pip install -U flash-attn --no-build-isolation
+```
+
+또한 Flash-Attention 2와 호환되는 하드웨어가 있는지 확인하세요. flash-attn 저장소의 공식 문서에서 자세히 알아보세요. 또한 모델을 반정밀도(예: `torch.float16`)로 로드해야 합니다.
+
+Flash Attention 2를 사용하여 모델을 로드하고 실행하려면 아래 코드를 참조하세요:
+
+```python
+>>> import torch
+>>> import requests
+>>> from PIL import Image
+>>> from transformers import SiglipProcessor, SiglipModel
+>>> device = "cuda" # 모델을 로드할 장치
+
+>>> model = SiglipModel.from_pretrained(
+...     "google/siglip-so400m-patch14-384",
+...     attn_implementation="flash_attention_2",
+...     torch_dtype=torch.float16,
+...     device_map=device,
+... )
+>>> processor = SiglipProcessor.from_pretrained("google/siglip-so400m-patch14-384")
+
+>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+>>> image = Image.open(requests.get(url, stream=True).raw)
+
+>>> candidate_labels = ["2 cats", "2 dogs"]
+# 파이프라인 프롬프트 템플릿을 따라 동일한 결과를 얻습니다
+>>> texts = [f'This is a photo of {label}.' for label in candidate_labels]
+# 중요: 모델이 이렇게 학습되었으므로 `padding=max_length`를 전달합니다
+>>> inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt").to(device)
+
+>>> with torch.no_grad():
+...     with torch.autocast(device):
+...         outputs = model(**inputs)
+
+>>> logits_per_image = outputs.logits_per_image
+>>> probs = torch.sigmoid(logits_per_image) # 시그모이드 활성화 함수를 적용한 확률입니다
+>>> print(f"{probs[0][0]:.1%} that image 0 is '{candidate_labels[0]}'")
+19.8% that image 0 is '2 cats'
+```
+
+
+## Scaled Dot Product Attention(SDPA) 사용하기[using-scaled-dot-product-attention(SDPA)]]
+
+PyTorch는 `torch.nn.functional`의 일부로 스케일된 점곱 어텐션(SDPA) 연산자를 포함합니다. 이 함수는 
+입력과 사용 중인 하드웨어에 따라 적용할 수 있는 여러 구현을 포함합니다. 자세한 내용은 
+[공식 문서](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html) 
+또는 [GPU 추론](https://huggingface.co/docs/transformers/main/en/perf_infer_gpu_one#pytorch-scaled-dot-product-attention) 
+페이지를 참조하세요.
+
+`from_pretrained()`에서 `attn_implementation="sdpa"`를 설정하여 SDPA를 명시적으로 요청할 수 있습니다. `torch>=2.1.1`이 설치되어 있는지 확인하세요.
+
+```python
+>>> from transformers import SiglipModel
+
+>>> model = SiglipModel.from_pretrained(
+...     "google/siglip-so400m-patch14-384",
+...     attn_implementation="sdpa",
+...     torch_dtype=torch.float16,
+...     device_map=device,
+... )
+```
+
+최상의 속도 향상을 위해 모델을 반정밀도(예: `torch.float16` 또는 `torch.bfloat16`)로 로드하는 것이 좋습니다.
+
+
+## 예상 속도 향상[[expected-speedups]]
+
+아래는 `google/siglip-so400m-patch14-384` 체크포인트를 `float16` 정밀도로 사용하는 transformers의 네이티브 구현과 Flash Attention 2 / SDPA 버전의 모델을 다양한 배치 크기로 비교한 추론 시간의 예상 속도 향상 다이어그램입니다.
+
+<div style="text-align: center">
+<img src="https://i.imgur.com/cWm4rsn.png">
+</div>
+
+
+## SiglipConfig
+
+[[autodoc]] SiglipConfig
+    - from_text_vision_configs
+
+## SiglipTextConfig
+
+[[autodoc]] SiglipTextConfig
+
+## SiglipVisionConfig
+
+[[autodoc]] SiglipVisionConfig
+
+## SiglipTokenizer
+
+[[autodoc]] SiglipTokenizer
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - save_vocabulary
+
+## SiglipImageProcessor
+
+[[autodoc]] SiglipImageProcessor
+    - preprocess
+
+## SiglipImageProcessorFast
+
+[[autodoc]] SiglipImageProcessorFast
+    - preprocess
+
+## SiglipProcessor
+
+[[autodoc]] SiglipProcessor
+
+## SiglipModel
+
+[[autodoc]] SiglipModel
+    - forward
+    - get_text_features
+    - get_image_features
+
+## SiglipTextModel
+
+[[autodoc]] SiglipTextModel
+    - forward
+
+## SiglipVisionModel
+
+[[autodoc]] SiglipVisionModel
+    - forward
+
+
+## SiglipForImageClassification
+
+[[autodoc]] SiglipForImageClassification
+    - forward 

docs/source/ko/tasks/keypoint_detection.md

@@ -0,0 +1,155 @@
+<!--Copyright 2023 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.
+
+-->
+
+# 키포인트 탐지 [[keypoint-detection]]
+
+[[open-in-colab]]
+
+키포인트 감지(Keypoint detection)은 이미지 내의 특정 포인트를 식별하고 위치를 탐지합니다. 이러한 키포인트는 랜드마크라고도 불리며 얼굴 특징이나 물체의 일부와 같은 의미 있는 특징을 나타냅니다.
+키포인트 감지 모델들은 이미지를 입력으로 받아 아래와 같은 출력을 반환합니다.
+
+- **키포인트들과 점수**: 관심 포인트들과 해당 포인트에 대한 신뢰도 점수
+- **디스크립터(Descriptors)**: 각 키포인트를 둘러싼 이미지 영역의 표현으로 텍스처, 그라데이션, 방향 및 기타 속성을 캡처합니다.
+
+이번 가이드에서는 이미지에서 키포인트를 추출하는 방법을 다루어 보겠습니다.
+
+이번 튜토리얼에서는 키포인트 감지의 기본이 되는 모델인 [SuperPoint](./model_doc/superpoint)를 사용해보겠습니다.
+
+```python
+from transformers import AutoImageProcessor, SuperPointForKeypointDetection
+processor = AutoImageProcessor.from_pretrained("magic-leap-community/superpoint")
+model = SuperPointForKeypointDetection.from_pretrained("magic-leap-community/superpoint")
+```
+아래의 이미지로 모델을 테스트 해보겠습니다.
+
+<div style="display: flex; align-items: center;">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg" 
+         alt="Bee" 
+         style="height: 200px; object-fit: contain; margin-right: 10px;">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.png" 
+         alt="Cats" 
+         style="height: 200px; object-fit: contain;">
+</div>
+
+
+```python
+import torch
+from PIL import Image
+import requests
+import cv2
+
+
+url_image_1 = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"
+image_1 = Image.open(requests.get(url_image_1, stream=True).raw)
+url_image_2 = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.png"
+image_2 = Image.open(requests.get(url_image_2, stream=True).raw)
+
+images = [image_1, image_2]
+```
+
+이제 입력을 처리하고 추론을 할 수 있습니다.
+
+
+```python
+inputs = processor(images,return_tensors="pt").to(model.device, model.dtype)
+outputs = model(**inputs)
+```
+모델 출력에는 배치 내의 각 항목에 대한 상대적인 키포인트, 디스크립터, 마스크와 점수가 있습니다. 마스크는 이미지에서 키포인트가 있는 영역을 강조하는 역할을 합니다.
+
+```python
+SuperPointKeypointDescriptionOutput(loss=None, keypoints=tensor([[[0.0437, 0.0167],
+         [0.0688, 0.0167],
+         [0.0172, 0.0188],
+         ...,
+         [0.5984, 0.9812],
+         [0.6953, 0.9812]]]), 
+         scores=tensor([[0.0056, 0.0053, 0.0079,  ..., 0.0125, 0.0539, 0.0377],
+        [0.0206, 0.0058, 0.0065,  ..., 0.0000, 0.0000, 0.0000]],
+       grad_fn=<CopySlices>), descriptors=tensor([[[-0.0807,  0.0114, -0.1210,  ..., -0.1122,  0.0899,  0.0357],
+         [-0.0807,  0.0114, -0.1210,  ..., -0.1122,  0.0899,  0.0357],
+         [-0.0807,  0.0114, -0.1210,  ..., -0.1122,  0.0899,  0.0357],
+         ...],
+       grad_fn=<CopySlices>), mask=tensor([[1, 1, 1,  ..., 1, 1, 1],
+        [1, 1, 1,  ..., 0, 0, 0]], dtype=torch.int32), hidden_states=None)
+```
+
+이미지에 실제 키포인트를 표시하기 위해선 결과값을 후처리 해야합니다. 이를 위해 실제 이미지 크기를 결과값과 함께 `post_process_keypoint_detection`에 전달해야 합니다.
+
+```python
+image_sizes = [(image.size[1], image.size[0]) for image in images]
+outputs = processor.post_process_keypoint_detection(outputs, image_sizes)
+```
+
+위 코드를 통해 결과값은 딕셔너리를 갖는 리스트가 되고, 각 딕셔너리들은 후처리된 키포인트, 점수 및 디스크립터로 이루어져있습니다.
+
+
+```python
+[{'keypoints': tensor([[ 226,   57],
+          [ 356,   57],
+          [  89,   64],
+          ...,
+          [3604, 3391]], dtype=torch.int32),
+  'scores': tensor([0.0056, 0.0053, ...], grad_fn=<IndexBackward0>),
+  'descriptors': tensor([[-0.0807,  0.0114, -0.1210,  ..., -0.1122,  0.0899,  0.0357],
+          [-0.0807,  0.0114, -0.1210,  ..., -0.1122,  0.0899,  0.0357]],
+         grad_fn=<IndexBackward0>)},
+    {'keypoints': tensor([[ 46,   6],
+          [ 78,   6],
+          [422,   6],
+          [206, 404]], dtype=torch.int32),
+  'scores': tensor([0.0206, 0.0058, 0.0065, 0.0053, 0.0070, ...,grad_fn=<IndexBackward0>),
+  'descriptors': tensor([[-0.0525,  0.0726,  0.0270,  ...,  0.0389, -0.0189, -0.0211],
+          [-0.0525,  0.0726,  0.0270,  ...,  0.0389, -0.0189, -0.0211]}]
+```
+
+이제 위 딕셔너리를 사용하여 키포인트를 표시할 수 있습니다.
+
+```python
+import matplotlib.pyplot as plt
+import torch
+
+for i in range(len(images)):
+  keypoints = outputs[i]["keypoints"]
+  scores = outputs[i]["scores"]
+  descriptors = outputs[i]["descriptors"]
+  keypoints = outputs[i]["keypoints"].detach().numpy()
+  scores = outputs[i]["scores"].detach().numpy()
+  image = images[i]
+  image_width, image_height = image.size
+
+  plt.axis('off')
+  plt.imshow(image)
+  plt.scatter(
+      keypoints[:, 0],
+      keypoints[:, 1],
+      s=scores * 100,
+      c='cyan',
+      alpha=0.4
+  )
+  plt.show()
+```
+
+아래에서 결과를 확인할 수 있습니다.
+
+<div style="display: flex; align-items: center;">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee_keypoint.png" 
+         alt="Bee" 
+         style="height: 200px; object-fit: contain; margin-right: 10px;">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats_keypoint.png" 
+         alt="Cats" 
+         style="height: 200px; object-fit: contain;">
+</div>
+

docs/source/zh/internal/modeling_utils.md

@@ -25,23 +25,6 @@ rendered properly in your Markdown viewer.
 
 [[autodoc]] pytorch_utils.Conv1D
 
-[[autodoc]] modeling_utils.PoolerStartLogits
-    - forward
-
-[[autodoc]] modeling_utils.PoolerEndLogits
-    - forward
-
-[[autodoc]] modeling_utils.PoolerAnswerClass
-    - forward
-
-[[autodoc]] modeling_utils.SquadHeadOutput
-
-[[autodoc]] modeling_utils.SQuADHead
-    - forward
-
-[[autodoc]] modeling_utils.SequenceSummary
-    - forward
-
 ## PyTorch帮助函数
 
 [[autodoc]] pytorch_utils.apply_chunking_to_forward

examples/pytorch/continuous_batching.py

@@ -1,20 +0,0 @@
-import datasets
-import torch
-
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-
-model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", attn_implementation="eager", torch_dtype=torch.float16)
-tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.float16)
-
-train_dataset = datasets.load_dataset('imdb', split='test')
-
-def tokenize_function(examples):
-    return tokenizer(examples["text"])
-
-tokenized_datasets = train_dataset.map(tokenize_function, batched=True)
-
-for batch in model.fast_generate(tokenized_datasets):
-    print(tokenizer.batch_decode(batch))
-
-

examples/pytorch/image-classification/run_image_classification_no_trainer.py

@@ -617,9 +617,6 @@ def main():
                 output_dir = os.path.join(args.output_dir, output_dir)
             accelerator.save_state(output_dir)
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -640,6 +637,9 @@ def main():
             with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
                 json.dump(all_results, f)
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/image-pretraining/run_mim_no_trainer.py

@@ -778,9 +778,6 @@ def main():
                 output_dir = os.path.join(args.output_dir, output_dir)
             accelerator.save_state(output_dir)
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -798,6 +795,9 @@ def main():
                     token=args.hub_token,
                 )
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/instance-segmentation/run_instance_segmentation_no_trainer.py

@@ -714,9 +714,6 @@ def main():
 
     logger.info(f"Test metrics: {metrics}")
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -739,6 +736,9 @@ def main():
                     ignore_patterns=["epoch_*"],
                 )
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/language-modeling/run_clm_no_trainer.py

@@ -697,9 +697,6 @@ def main():
                 output_dir = os.path.join(args.output_dir, output_dir)
             accelerator.save_state(output_dir)
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -719,6 +716,9 @@ def main():
             with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
                 json.dump({"perplexity": perplexity}, f)
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/language-modeling/run_fim_no_trainer.py

@@ -891,9 +891,6 @@ def main():
                 output_dir = os.path.join(args.output_dir, output_dir)
             accelerator.save_state(output_dir)
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -908,6 +905,9 @@ def main():
             with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
                 json.dump({"perplexity": perplexity}, f)
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/language-modeling/run_mlm_no_trainer.py

@@ -735,9 +735,6 @@ def main():
                 output_dir = os.path.join(args.output_dir, output_dir)
             accelerator.save_state(output_dir)
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -757,6 +754,9 @@ def main():
             with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
                 json.dump({"perplexity": perplexity}, f)
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/multiple-choice/run_swag_no_trainer.py

@@ -622,9 +622,6 @@ def main():
                 output_dir = os.path.join(args.output_dir, output_dir)
             accelerator.save_state(output_dir)
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -645,6 +642,9 @@ def main():
             with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
                 json.dump(all_results, f)
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/object-detection/run_object_detection_no_trainer.py

@@ -759,9 +759,6 @@ def main():
 
     logger.info(f"Test metrics: {metrics}")
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -784,6 +781,9 @@ def main():
                     ignore_patterns=["epoch_*"],
                 )
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/semantic-segmentation/run_semantic_segmentation_no_trainer.py

@@ -602,9 +602,6 @@ def main():
                 output_dir = os.path.join(args.output_dir, output_dir)
             accelerator.save_state(output_dir)
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -628,6 +625,9 @@ def main():
             with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
                 json.dump(all_results, f, indent=2)
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/text-classification/run_glue.py

@@ -19,6 +19,7 @@ import logging
 import os
 import random
 import sys
+from collections import Counter
 from dataclasses import dataclass, field
 from typing import Optional
 
@@ -467,6 +468,14 @@ def main():
             load_from_cache_file=not data_args.overwrite_cache,
             desc="Running tokenizer on dataset",
         )
+
+    def print_class_distribution(dataset, split_name):
+        label_counts = Counter(dataset["label"])
+        total = sum(label_counts.values())
+        logger.info(f"Class distribution in {split_name} set:")
+        for label, count in label_counts.items():
+            logger.info(f"  Label {label}: {count} ({count / total:.2%})")
+
     if training_args.do_train:
         if "train" not in raw_datasets:
             raise ValueError("--do_train requires a train dataset")
@@ -474,6 +483,7 @@ def main():
         if data_args.max_train_samples is not None:
             max_train_samples = min(len(train_dataset), data_args.max_train_samples)
             train_dataset = train_dataset.select(range(max_train_samples))
+        print_class_distribution(train_dataset, "train")
 
     if training_args.do_eval:
         if "validation" not in raw_datasets and "validation_matched" not in raw_datasets:
@@ -482,6 +492,7 @@ def main():
         if data_args.max_eval_samples is not None:
             max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
             eval_dataset = eval_dataset.select(range(max_eval_samples))
+        print_class_distribution(eval_dataset, "validation")
 
     if training_args.do_predict or data_args.task_name is not None or data_args.test_file is not None:
         if "test" not in raw_datasets and "test_matched" not in raw_datasets:
@@ -490,6 +501,7 @@ def main():
         if data_args.max_predict_samples is not None:
             max_predict_samples = min(len(predict_dataset), data_args.max_predict_samples)
             predict_dataset = predict_dataset.select(range(max_predict_samples))
+        print_class_distribution(predict_dataset, "test")
 
     # Log a few random samples from the training set:
     if training_args.do_train:
@@ -508,8 +520,12 @@ def main():
     # predictions and label_ids field) and has to return a dictionary string to float.
     def compute_metrics(p: EvalPrediction):
         preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions
+        labels = p.label_ids
+        if not training_args.eval_do_concat_batches:
+            preds = np.concatenate(preds, axis=0)
+            labels = np.concatenate(p.label_ids, axis=0)
         preds = np.squeeze(preds) if is_regression else np.argmax(preds, axis=1)
-        result = metric.compute(predictions=preds, references=p.label_ids)
+        result = metric.compute(predictions=preds, references=labels)
         if len(result) > 1:
             result["combined_score"] = np.mean(list(result.values())).item()
         return result

examples/pytorch/text-classification/run_glue_no_trainer.py

@@ -634,9 +634,6 @@ def main():
                 output_dir = os.path.join(args.output_dir, output_dir)
             accelerator.save_state(output_dir)
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -679,6 +676,9 @@ def main():
         with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
             json.dump(all_results, f)
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/text-generation/run_generation.py

@@ -38,8 +38,6 @@ from transformers import (
     OpenAIGPTLMHeadModel,
     OpenAIGPTTokenizer,
     OPTForCausalLM,
-    TransfoXLLMHeadModel,
-    TransfoXLTokenizer,
     XLMTokenizer,
     XLMWithLMHeadModel,
     XLNetLMHeadModel,
@@ -62,7 +60,6 @@ MODEL_CLASSES = {
     "ctrl": (CTRLLMHeadModel, CTRLTokenizer),
     "openai-gpt": (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
     "xlnet": (XLNetLMHeadModel, XLNetTokenizer),
-    "transfo-xl": (TransfoXLLMHeadModel, TransfoXLTokenizer),
     "xlm": (XLMWithLMHeadModel, XLMTokenizer),
     "gptj": (GPTJForCausalLM, AutoTokenizer),
     "bloom": (BloomForCausalLM, BloomTokenizerFast),
@@ -368,10 +365,7 @@ def main():
         prepare_input = PREPROCESSING_FUNCTIONS.get(args.model_type)
         preprocessed_prompt_text = prepare_input(args, model, tokenizer, prompt_text)
 
-        if model.__class__.__name__ in ["TransfoXLLMHeadModel"]:
-            tokenizer_kwargs = {"add_space_before_punct_symbol": True}
-        else:
-            tokenizer_kwargs = {}
+        tokenizer_kwargs = {}
 
         encoded_prompt = tokenizer.encode(
             preprocessed_prompt_text, add_special_tokens=False, return_tensors="pt", **tokenizer_kwargs

examples/pytorch/token-classification/run_ner.py

@@ -529,6 +529,9 @@ def main():
 
     def compute_metrics(p):
         predictions, labels = p
+        if not training_args.eval_do_concat_batches:
+            predictions = np.hstack(predictions)
+            labels = np.hstack(labels)
         predictions = np.argmax(predictions, axis=2)
 
         # Remove ignored index (special tokens)

examples/pytorch/token-classification/run_ner_no_trainer.py

@@ -794,9 +794,6 @@ def main():
                 output_dir = os.path.join(args.output_dir, output_dir)
             accelerator.save_state(output_dir)
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -826,6 +823,9 @@ def main():
                         all_results[key] = int(value)
                 json.dump(all_results, f)
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

examples/pytorch/translation/run_translation_no_trainer.py

@@ -762,9 +762,6 @@ def main():
                 output_dir = os.path.join(args.output_dir, output_dir)
             accelerator.save_state(output_dir)
 
-    if args.with_tracking:
-        accelerator.end_training()
-
     if args.output_dir is not None:
         accelerator.wait_for_everyone()
         unwrapped_model = accelerator.unwrap_model(model)
@@ -784,6 +781,9 @@ def main():
         with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
             json.dump({"eval_bleu": eval_metric["score"]}, f)
 
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+
 
 if __name__ == "__main__":
     main()

model_cards/README.md

@@ -1,22 +0,0 @@
-## 🔥 Model cards now live inside each huggingface.co model repo 🔥
-
-
-For consistency, ease of use and scalability, `README.md` model cards now live directly inside each model repo on the HuggingFace model hub.
-
-### How to update a model card
-
-You can directly update a model card inside any model repo you have **write access** to, i.e.:
-- a model under your username namespace
-- a model under any organization you are a part of.
-
-You can either:
-- update it, commit and push using your usual git workflow (command line, GUI, etc.)
-- or edit it directly from the website's UI.
-
-**What if you want to create or update a model card for a model you don't have write access to?**
-
-In that case, you can open a [Hub pull request](https://huggingface.co/docs/hub/repositories-pull-requests-discussions)! Check out the [announcement](https://huggingface.co/blog/community-update) of this feature for more details 🤗.
-
-### What happened to the model cards here?
-
-We migrated every model card from the repo to its corresponding huggingface.co model repo. Individual commits were preserved, and they link back to the original commit on GitHub.

pyproject.toml

@@ -55,3 +55,4 @@ markers = [
 ]
 log_cli = 1
 log_cli_level = "WARNING"
+asyncio_default_fixture_loop_scope = "function"

scripts/benchmark/trainer-benchmark.py

@@ -1,448 +0,0 @@
-#!/usr/bin/env python
-
-# HF Trainer benchmarking tool
-#
-# This tool can be used to run and compare multiple dimensions of the HF Trainers args.
-#
-# It then prints a report once in github format with all the information that needs to be shared
-# with others and second time in a console-friendly format, so it's easier to use for tuning things up.
-#
-# The main idea is:
-#
-#     ./trainer-benchmark.py --base-cmd '<cmd args that don't change>' \
-#     --variations '--tf32 0|--tf32 1' '--fp16 0|--fp16 1|--bf16 1' \
-#     --target-metric-key train_samples_per_second
-#
-# The variations can be any command line argument that you want to compare and not just dtype as in
-# the example.
-#
-# --variations allows you to compare variations in multiple dimensions.
-#
-# as the first dimension has 2 options and the second 3 in our example, this will run the trainer 6
-# times adding one of:
-#
-#    1. --tf32 0 --fp16 0
-#    2. --tf32 0 --fp16 1
-#    3. --tf32 0 --bf16 1
-#    4. --tf32 1 --fp16 0
-#    5. --tf32 1 --fp16 1
-#    6. --tf32 1 --bf16 1
-#
-# and print the results. This is just a cartesian product - and more than 2 dimensions can be used.
-#
-# If you want to rely on defaults, this:
-#    --variations '--tf32 0|--tf32 1' '--fp16 0|--fp16 1|--bf16 1'
-# is identical to this:
-#    --variations '--tf32 0|--tf32 1' '|--fp16|--bf16'
-#
-# the leading empty variation in the 2nd dimension is a valid variation.
-#
-# So here we get the following 6 variations:
-#
-#    1. --tf32 0
-#    2. --tf32 0 --fp16
-#    3. --tf32 0 --bf16
-#    4. --tf32 1
-#    5. --tf32 1 --fp16
-#    6. --tf32 1 --bf16
-#
-# In this particular case we don't know what the default tf32 setting is as it's normally
-# pytorch-version dependent). That's why it's best to do an explicit setting of each variation:
-#    `--tf32 0|--tf32 1`
-#
-# Here is a full example of a train:
-#
-# CUDA_VISIBLE_DEVICES=0 python ./scripts/benchmark/trainer-benchmark.py \
-# --base-cmd \
-# ' examples/pytorch/translation/run_translation.py --model_name_or_path google-t5/t5-small \
-# --output_dir output_dir --do_train --label_smoothing 0.1 --logging_strategy no \
-# --save_strategy no --per_device_train_batch_size 32 --max_source_length 512 \
-# --max_target_length 512 --num_train_epochs 1 --overwrite_output_dir \
-# --source_lang en --target_lang ro --dataset_name wmt16 --dataset_config "ro-en" \
-# --source_prefix "translate English to Romanian: " --warmup_steps 50 \
-# --max_train_samples 20000 --dataloader_num_workers 2 ' \
-# --target-metric-key train_samples_per_second --repeat-times 1 --variations \
-# '|--fp16|--bf16' '--tf32 0|--tf32 1' --report-metric-keys train_loss \
-# --repeat-times 1 --base-variation '--tf32 0'
-#
-# and here is a possible output:
-#
-#
-# | Variation       |     Train |   Diff |   Train |
-# |                 |   samples |      % |    loss |
-# |                 |       per |        |         |
-# |                 |    second |        |         |
-# |:----------------|----------:|-------:|--------:|
-# | --tf32 0        |    285.11 |      0 |    2.51 |
-# | --tf32 1        |    342.09 |     20 |    2.51 |
-# | --fp16 --tf32 0 |    423.49 |     49 |    2.51 |
-# | --fp16 --tf32 1 |    423.13 |     48 |    2.51 |
-# | --bf16 --tf32 0 |    416.80 |     46 |    2.52 |
-# | --bf16 --tf32 1 |    415.87 |     46 |    2.52 |
-#
-#
-# So you can quickly compare the different outcomes.
-#
-# Typically running each experiment once is enough, but if the environment is unstable you can
-# re-run each multiple times, e.g., 3 using --repeat-times 3 and it will report the averaged results.
-#
-# By default it'll use the lowest result as the base line to use as 100% and then compare the rest to
-# it as can be seen from the table above, but you can also specify which combination is the one to use as
-# the baseline, e.g., to change to another entry use: --base-variation '--tf32 1 --fp16 0'
-#
-# --target-metric-key is there to tell the program which metrics to compare - the different metric keys are
-# inside output_dir/all_results.json. e.g., to measure eval performance instead of train use:
-#    --target-metric-key eval_samples_per_second
-# but of course you will need to adjust the --base-cmd value in the example to perform evaluation as
-# well (as currently it doesn't)
-#
-
-import argparse
-import datetime
-import io
-import itertools
-import json
-import math
-import os
-import platform
-import re
-import shlex
-import subprocess
-import sys
-from pathlib import Path
-from statistics import fmean
-
-import pandas as pd
-import torch
-from tqdm import tqdm
-
-import transformers
-
-
-nan = float("nan")
-
-
-class Tee:
-    """
-    A helper class to tee print's output into a file.
-    Usage:
-    sys.stdout = Tee(filename)
-    """
-
-    def __init__(self, filename):
-        self.stdout = sys.stdout
-        self.file = open(filename, "a")
-
-    def __getattr__(self, attr):
-        return getattr(self.stdout, attr)
-
-    def write(self, msg):
-        self.stdout.write(msg)
-        # strip tqdm codes
-        self.file.write(re.sub(r"^.*\r", "", msg, 0, re.M))
-
-
-def get_original_command(max_width=80, full_python_path=False):
-    """
-    Return the original command line string that can be replayed nicely and wrapped for 80 char width.
-
-    Args:
-        max_width (`int`, *optional*, defaults to 80):
-            The width to wrap for.
-        full_python_path (`bool`, `optional`, defaults to `False`):
-             Whether to replicate the full path or just the last segment (i.e. `python`).
-    """
-
-    cmd = []
-
-    # deal with critical env vars
-    env_keys = ["CUDA_VISIBLE_DEVICES"]
-    for key in env_keys:
-        val = os.environ.get(key, None)
-        if val is not None:
-            cmd.append(f"{key}={val}")
-
-    # python executable (not always needed if the script is executable)
-    python = sys.executable if full_python_path else sys.executable.split("/")[-1]
-    cmd.append(python)
-
-    # now the normal args
-    cmd += list(map(shlex.quote, sys.argv))
-
-    # split up into up to MAX_WIDTH lines with shell multi-line escapes
-    lines = []
-    current_line = ""
-    while len(cmd) > 0:
-        current_line += f"{cmd.pop(0)} "
-        if len(cmd) == 0 or len(current_line) + len(cmd[0]) + 1 > max_width - 1:
-            lines.append(current_line)
-            current_line = ""
-    return "\\\n".join(lines)
-
-
-def get_base_command(args, output_dir):
-
-    # unwrap multi-line input
-    args.base_cmd = re.sub(r"[\\\n]+", " ", args.base_cmd)
-
-    # remove --output_dir if any and set our own
-    args.base_cmd = re.sub("--output_dir\s+[^\s]+", "", args.base_cmd)
-    args.base_cmd += f" --output_dir {output_dir}"
-
-    # ensure we have --overwrite_output_dir
-    args.base_cmd = re.sub("--overwrite_output_dir\s+", "", args.base_cmd)
-    args.base_cmd += " --overwrite_output_dir"
-
-    return [sys.executable] + shlex.split(args.base_cmd)
-
-
-def process_run_single(id, cmd, variation, output_dir, target_metric_key, metric_keys, verbose):
-
-    # Enable to debug everything but the run itself, to do it fast and see the progress.
-    # This is useful for debugging the output formatting quickly - we can remove it later once
-    # everybody is happy with the output
-    if 0:
-        import random
-        from time import sleep
-
-        sleep(0)
-        return dict(
-            {k: random.uniform(0, 100) for k in metric_keys},
-            **{target_metric_key: random.choice([nan, 10.31, 100.2, 55.6666, 222.22222222])},
-        )
-
-    result = subprocess.run(cmd, capture_output=True, text=True)
-
-    if verbose:
-        print("STDOUT", result.stdout)
-        print("STDERR", result.stderr)
-
-    # save the streams
-    prefix = variation.replace(" ", "-")
-    with open(Path(output_dir) / f"log.{prefix}.stdout.txt", "w") as f:
-        f.write(result.stdout)
-    with open(Path(output_dir) / f"log.{prefix}.stderr.txt", "w") as f:
-        f.write(result.stderr)
-
-    if result.returncode != 0:
-        if verbose:
-            print("failed")
-        return {target_metric_key: nan}
-
-    with io.open(f"{output_dir}/all_results.json", "r", encoding="utf-8") as f:
-        metrics = json.load(f)
-
-    # filter out just the keys we want
-    return {k: v for k, v in metrics.items() if k in metric_keys}
-
-
-def process_run(
-    id,
-    cmd,
-    variation_key,
-    variation,
-    longest_variation_len,
-    target_metric_key,
-    report_metric_keys,
-    repeat_times,
-    output_dir,
-    verbose,
-):
-    results = []
-    metrics = []
-    preamble = f"{id}: {variation:<{longest_variation_len}}"
-    outcome = f"{preamble}: "
-    metric_keys = set(report_metric_keys + [target_metric_key])
-    for i in tqdm(range(repeat_times), desc=preamble, leave=False):
-        single_run_metrics = process_run_single(
-            id, cmd, variation, output_dir, target_metric_key, metric_keys, verbose
-        )
-        result = single_run_metrics[target_metric_key]
-        if not math.isnan(result):
-            metrics.append(single_run_metrics)
-            results.append(result)
-            outcome += "✓"
-        else:
-            outcome += "✘"
-    outcome = f"\33[2K\r{outcome}"
-    if len(metrics) > 0:
-        mean_metrics = {k: fmean([x[k] for x in metrics]) for k in metrics[0].keys()}
-        mean_target = round(mean_metrics[target_metric_key], 2)
-        results_str = f"{outcome} {mean_target}"
-        if len(metrics) > 1:
-            results_str += f" {tuple(round(x, 2) for x in results)}"
-        print(results_str)
-        mean_metrics[variation_key] = variation
-        return mean_metrics
-    else:
-        print(outcome)
-        return {variation_key: variation, target_metric_key: nan}
-
-
-def get_versions():
-    properties = torch.cuda.get_device_properties(torch.device("cuda"))
-    return f"""
-Datetime    : {datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')}
-
-Software:
-transformers: {transformers.__version__}
-torch       : {torch.__version__}
-cuda        : {torch.version.cuda}
-python      : {platform.python_version()}
-
-Hardware:
-{torch.cuda.device_count()} GPUs      : {properties.name}, {properties.total_memory/2**30:0.2f}GB
-"""
-
-
-def process_results(results, target_metric_key, report_metric_keys, base_variation, output_dir):
-
-    df = pd.DataFrame(results)
-    variation_key = "variation"
-    diff_key = "diff_%"
-
-    sentinel_value = nan
-    if base_variation is not None and len(df[df[variation_key] == base_variation]):
-        # this may still return nan
-        sentinel_value = df.loc[df[variation_key] == base_variation][target_metric_key].item()
-    if math.isnan(sentinel_value):
-        # as a fallback, use the minimal value as the sentinel
-        sentinel_value = df.loc[df[target_metric_key] != nan][target_metric_key].min()
-
-    # create diff column if possible
-    if not math.isnan(sentinel_value):
-        df[diff_key] = df.apply(
-            lambda r: round(100 * (r[target_metric_key] - sentinel_value) / sentinel_value)
-            if not math.isnan(r[target_metric_key])
-            else 0,
-            axis="columns",
-        )
-
-    # re-order columns
-    cols = [variation_key, target_metric_key, diff_key, *report_metric_keys]
-    df = df.reindex(cols, axis="columns")  # reorder cols
-
-    # capitalize
-    df = df.rename(str.capitalize, axis="columns")
-
-    # make the cols as narrow as possible
-    df_github = df.rename(lambda c: c.replace("_", "<br>"), axis="columns")
-    df_console = df.rename(lambda c: c.replace("_", "\n"), axis="columns")
-
-    report = ["", "Copy between the cut-here-lines and paste as is to github or a forum"]
-    report += ["----------8<-----------------8<--------"]
-    report += ["*** Results:", df_github.to_markdown(index=False, floatfmt=".2f")]
-    report += ["```"]
-    report += ["*** Setup:", get_versions()]
-    report += ["*** The benchmark command line was:", get_original_command()]
-    report += ["```"]
-    report += ["----------8<-----------------8<--------"]
-    report += ["*** Results (console):", df_console.to_markdown(index=False, floatfmt=".2f")]
-
-    print("\n\n".join(report))
-
-
-def main():
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--base-cmd",
-        default=None,
-        type=str,
-        required=True,
-        help="Base cmd",
-    )
-    parser.add_argument(
-        "--variations",
-        default=None,
-        type=str,
-        nargs="+",
-        required=True,
-        help="Multi-dimensional variations, example: '|--fp16|--bf16' '|--tf32'",
-    )
-    parser.add_argument(
-        "--base-variation",
-        default=None,
-        type=str,
-        help="Baseline variation to compare to. if None the minimal target value will be used to compare against",
-    )
-    parser.add_argument(
-        "--target-metric-key",
-        default=None,
-        type=str,
-        required=True,
-        help="Target metric key in output_dir/all_results.json, e.g., train_samples_per_second",
-    )
-    parser.add_argument(
-        "--report-metric-keys",
-        default="",
-        type=str,
-        help="Report metric keys - other metric keys from output_dir/all_results.json to report, e.g., train_loss. Use a single argument e.g., 'train_loss train_samples",
-    )
-    parser.add_argument(
-        "--repeat-times",
-        default=1,
-        type=int,
-        help="How many times to re-run each variation - an average will be reported",
-    )
-    parser.add_argument(
-        "--output_dir",
-        default="output_benchmark",
-        type=str,
-        help="The output directory where all the benchmark reports will go to and additionally this directory will be used to override --output_dir in the script that is being benchmarked",
-    )
-    parser.add_argument(
-        "--verbose",
-        default=False,
-        action="store_true",
-        help="Whether to show the outputs of each run or just the benchmark progress",
-    )
-    args = parser.parse_args()
-
-    output_dir = args.output_dir
-    Path(output_dir).mkdir(exist_ok=True)
-    base_cmd = get_base_command(args, output_dir)
-
-    # split each dimension into its --foo variations
-    dims = [list(map(str.strip, re.split(r"\|", x))) for x in args.variations]
-    # build a cartesian product of dimensions and convert those back into cmd-line arg strings,
-    # while stripping white space for inputs that were empty
-    variations = list(map(str.strip, map(" ".join, itertools.product(*dims))))
-    longest_variation_len = max(len(x) for x in variations)
-
-    # split wanted keys
-    report_metric_keys = args.report_metric_keys.split()
-
-    # capture prints into a log file for convenience
-    report_fn = f"benchmark-report-{datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S')}.txt"
-    print(f"\nNote: each run's output is also logged under {output_dir}/log.*.std*.txt")
-    print(f"and this script's output is also piped into {report_fn}")
-
-    sys.stdout = Tee(report_fn)
-
-    print(f"\n*** Running {len(variations)} benchmarks:")
-    print(f"Base command: {' '.join(base_cmd)}")
-
-    variation_key = "variation"
-    results = []
-    for id, variation in enumerate(tqdm(variations, desc="Total completion: ", leave=False)):
-        cmd = base_cmd + variation.split()
-        results.append(
-            process_run(
-                id + 1,
-                cmd,
-                variation_key,
-                variation,
-                longest_variation_len,
-                args.target_metric_key,
-                report_metric_keys,
-                args.repeat_times,
-                output_dir,
-                args.verbose,
-            )
-        )
-
-    process_results(results, args.target_metric_key, report_metric_keys, args.base_variation, output_dir)
-
-
-if __name__ == "__main__":
-    main()

scripts/deberta_scrtipt.py

@@ -1,85 +0,0 @@
-import time
-
-import torch
-
-from transformers import AutoModel, AutoTokenizer, pipeline
-
-
-test_sentence = 'Do you [MASK] the muffin man?'
-
-# for comparison
-bert = pipeline('fill-mask', model = 'bert-base-uncased')
-print('\n'.join([d['sequence'] for d in bert(test_sentence)]))
-
-
-deberta = pipeline('fill-mask', model = 'microsoft/deberta-v3-base', model_kwargs={"legacy": False})
-print('\n'.join([d['sequence'] for d in deberta(test_sentence)]))
-
-
-tokenizer = AutoTokenizer.from_pretrained("microsoft/deberta-v3-base")
-
-tokenized_dict = tokenizer(
-    ["Is this working",], ["Not yet",],
-    return_tensors="pt"
-)
-
-deberta.model.forward = torch.compile(deberta.model.forward)
-start=time.time()
-deberta.model(**tokenized_dict)
-end=time.time()
-print(end-start)
-
-
-start=time.time()
-deberta.model(**tokenized_dict)
-end=time.time()
-print(end-start)
-
-
-start=time.time()
-deberta.model(**tokenized_dict)
-end=time.time()
-print(end-start)
-
-
-model = AutoModel.from_pretrained('microsoft/deberta-base')
-model.config.return_dict = False
-model.config.output_hidden_states=False
-input_tuple = (tokenized_dict['input_ids'], tokenized_dict['attention_mask'])
-
-
-start=time.time()
-traced_model = torch.jit.trace(model, input_tuple)
-end=time.time()
-print(end-start)
-
-
-start=time.time()
-traced_model(tokenized_dict['input_ids'], tokenized_dict['attention_mask'])
-end=time.time()
-print(end-start)
-
-
-start=time.time()
-traced_model(tokenized_dict['input_ids'], tokenized_dict['attention_mask'])
-end=time.time()
-print(end-start)
-
-
-start=time.time()
-traced_model(tokenized_dict['input_ids'], tokenized_dict['attention_mask'])
-end=time.time()
-print(end-start)
-
-
-start=time.time()
-traced_model(tokenized_dict['input_ids'], tokenized_dict['attention_mask'])
-end=time.time()
-print(end-start)
-
-
-torch.jit.save(traced_model, "compiled_deberta.pt")
-
-
-
-# my_script_module = torch.jit.script(model)

scripts/fsmt/convert-allenai-wmt16.sh

@@ -1,71 +0,0 @@
-#!/usr/bin/env bash
-# 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.
-
-# this script acquires data and converts it to fsmt model
-# it covers:
-# - allenai/wmt16-en-de-dist-12-1
-# - allenai/wmt16-en-de-dist-6-1
-# - allenai/wmt16-en-de-12-1
-
-# this script needs to be run from the top level of the transformers repo
-if [ ! -d "src/transformers" ]; then
-    echo "Error: This script needs to be run from the top of the transformers repo"
-    exit 1
-fi
-
-mkdir data
-
-# get data (run once)
-
-cd data
-gdown 'https://drive.google.com/uc?id=1x_G2cjvM1nW5hjAB8-vWxRqtQTlmIaQU'
-gdown 'https://drive.google.com/uc?id=1oA2aqZlVNj5FarxBlNXEHpBS4lRetTzU'
-gdown 'https://drive.google.com/uc?id=1Wup2D318QYBFPW_NKI1mfP_hXOfmUI9r'
-tar -xvzf trans_ende_12-1_0.2.tar.gz
-tar -xvzf trans_ende-dist_12-1_0.2.tar.gz
-tar -xvzf trans_ende-dist_6-1_0.2.tar.gz
-gdown 'https://drive.google.com/uc?id=1mNufoynJ9-Zy1kJh2TA_lHm2squji0i9'
-gdown 'https://drive.google.com/uc?id=1iO7um-HWoNoRKDtw27YUSgyeubn9uXqj'
-tar -xvzf wmt16.en-de.deep-shallow.dist.tar.gz
-tar -xvzf wmt16.en-de.deep-shallow.tar.gz
-cp wmt16.en-de.deep-shallow/data-bin/dict.*.txt trans_ende_12-1_0.2
-cp wmt16.en-de.deep-shallow.dist/data-bin/dict.*.txt trans_ende-dist_12-1_0.2
-cp wmt16.en-de.deep-shallow.dist/data-bin/dict.*.txt trans_ende-dist_6-1_0.2
-cp wmt16.en-de.deep-shallow/bpecodes trans_ende_12-1_0.2
-cp wmt16.en-de.deep-shallow.dist/bpecodes trans_ende-dist_12-1_0.2
-cp wmt16.en-de.deep-shallow.dist/bpecodes trans_ende-dist_6-1_0.2
-cd -
-
-# run conversions and uploads
-
-PYTHONPATH="src" python src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py --fsmt_checkpoint_path data/trans_ende-dist_12-1_0.2/checkpoint_top5_average.pt --pytorch_dump_folder_path data/wmt16-en-de-dist-12-1
-
-PYTHONPATH="src" python src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py --fsmt_checkpoint_path data/trans_ende-dist_6-1_0.2/checkpoint_top5_average.pt --pytorch_dump_folder_path data/wmt16-en-de-dist-6-1
-
-PYTHONPATH="src" python src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py --fsmt_checkpoint_path data/trans_ende_12-1_0.2/checkpoint_top5_average.pt --pytorch_dump_folder_path data/wmt16-en-de-12-1
-
-
-# upload
-cd data
-transformers-cli upload -y wmt16-en-de-dist-12-1
-transformers-cli upload -y wmt16-en-de-dist-6-1
-transformers-cli upload -y wmt16-en-de-12-1
-cd -
-
-
-# if updating just small files and not the large models, here is a script to generate the right commands:
-perl -le 'for $f (@ARGV) { print qq[transformers-cli upload -y $_/$f --filename $_/$f] for ("wmt16-en-de-dist-12-1", "wmt16-en-de-dist-6-1", "wmt16-en-de-12-1")}' vocab-src.json vocab-tgt.json tokenizer_config.json config.json
-# add/remove files as needed
-

scripts/fsmt/convert-allenai-wmt19.sh

@@ -1,59 +0,0 @@
-#!/usr/bin/env bash
-# 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.
-
-# this script acquires data and converts it to fsmt model
-# it covers:
-# - allenai/wmt19-de-en-6-6-base
-# - allenai/wmt19-de-en-6-6-big
-
-# this script needs to be run from the top level of the transformers repo
-if [ ! -d "src/transformers" ]; then
-    echo "Error: This script needs to be run from the top of the transformers repo"
-    exit 1
-fi
-
-mkdir data
-
-# get data (run once)
-
-cd data
-gdown 'https://drive.google.com/uc?id=1j6z9fYdlUyOYsh7KJoumRlr1yHczxR5T'
-gdown 'https://drive.google.com/uc?id=1yT7ZjqfvUYOBXvMjeY8uGRHQFWoSo8Q5'
-gdown 'https://drive.google.com/uc?id=15gAzHeRUCs-QV8vHeTReMPEh1j8excNE'
-tar -xvzf wmt19.de-en.tar.gz
-tar -xvzf wmt19_deen_base_dr0.1_1.tar.gz
-tar -xvzf wmt19_deen_big_dr0.1_2.tar.gz
-cp wmt19.de-en/data-bin/dict.*.txt wmt19_deen_base_dr0.1_1
-cp wmt19.de-en/data-bin/dict.*.txt wmt19_deen_big_dr0.1_2
-cd -
-
-# run conversions and uploads
-
-PYTHONPATH="src" python src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py --fsmt_checkpoint_path data/wmt19_deen_base_dr0.1_1/checkpoint_last3_avg.pt --pytorch_dump_folder_path data/wmt19-de-en-6-6-base
-
-PYTHONPATH="src" python src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py --fsmt_checkpoint_path data/wmt19_deen_big_dr0.1_2/checkpoint_last3_avg.pt --pytorch_dump_folder_path data/wmt19-de-en-6-6-big
-
-
-# upload
-cd data
-transformers-cli upload -y wmt19-de-en-6-6-base
-transformers-cli upload -y wmt19-de-en-6-6-big
-cd -
-
-
-# if updating just small files and not the large models, here is a script to generate the right commands:
-perl -le 'for $f (@ARGV) { print qq[transformers-cli upload -y $_/$f --filename $_/$f] for ("wmt19-de-en-6-6-base", "wmt19-de-en-6-6-big")}' vocab-src.json vocab-tgt.json tokenizer_config.json config.json
-# add/remove files as needed
-

scripts/fsmt/convert-facebook-wmt19.sh

@@ -1,70 +0,0 @@
-#!/usr/bin/env bash
-# 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.
-
-# this script acquires data and converts it to fsmt model
-# it covers:
-# - facebook/wmt19-ru-en
-# - facebook/wmt19-en-ru
-# - facebook/wmt19-de-en
-# - facebook/wmt19-en-de
-
-# this script needs to be run from the top level of the transformers repo
-if [ ! -d "src/transformers" ]; then
-    echo "Error: This script needs to be run from the top of the transformers repo"
-    exit 1
-fi
-
-mkdir data
-
-# get data (run once)
-
-cd data
-wget https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-de.joined-dict.ensemble.tar.gz
-wget https://dl.fbaipublicfiles.com/fairseq/models/wmt19.de-en.joined-dict.ensemble.tar.gz
-wget https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-ru.ensemble.tar.gz
-wget https://dl.fbaipublicfiles.com/fairseq/models/wmt19.ru-en.ensemble.tar.gz
-tar -xvzf wmt19.en-de.joined-dict.ensemble.tar.gz
-tar -xvzf wmt19.de-en.joined-dict.ensemble.tar.gz
-tar -xvzf wmt19.en-ru.ensemble.tar.gz
-tar -xvzf wmt19.ru-en.ensemble.tar.gz
-cd -
-
-# run conversions and uploads
-
-export PAIR=ru-en
-PYTHONPATH="src" python src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py --fsmt_checkpoint_path data/wmt19.$PAIR.ensemble/model4.pt --pytorch_dump_folder_path data/wmt19-$PAIR
-
-export PAIR=en-ru
-PYTHONPATH="src" python src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py --fsmt_checkpoint_path data/wmt19.$PAIR.ensemble/model4.pt --pytorch_dump_folder_path data/wmt19-$PAIR
-
-export PAIR=de-en
-PYTHONPATH="src" python src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py --fsmt_checkpoint_path data/wmt19.$PAIR.joined-dict.ensemble/model4.pt --pytorch_dump_folder_path data/wmt19-$PAIR
-
-export PAIR=en-de
-PYTHONPATH="src" python src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py --fsmt_checkpoint_path data/wmt19.$PAIR.joined-dict.ensemble/model4.pt --pytorch_dump_folder_path data/wmt19-$PAIR
-
-
-# upload
-cd data
-transformers-cli upload -y wmt19-ru-en
-transformers-cli upload -y wmt19-en-ru
-transformers-cli upload -y wmt19-de-en
-transformers-cli upload -y wmt19-en-de
-cd -
-
-# if updating just small files and not the large models, here is a script to generate the right commands:
-perl -le 'for $f (@ARGV) { print qq[transformers-cli upload -y $_/$f --filename $_/$f] for map { "wmt19-$_" } ("en-ru", "ru-en", "de-en", "en-de")}' vocab-src.json vocab-tgt.json tokenizer_config.json config.json
-# add/remove files as needed
-

scripts/fsmt/eval-allenai-wmt16.sh

@@ -1,79 +0,0 @@
-#!/usr/bin/env bash
-# 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.
-
-# this script evals the following fsmt models
-# it covers:
-# - allenai/wmt16-en-de-dist-12-1
-# - allenai/wmt16-en-de-dist-6-1
-# - allenai/wmt16-en-de-12-1
-
-# this script needs to be run from the top level of the transformers repo
-if [ ! -d "src/transformers" ]; then
-    echo "Error: This script needs to be run from the top of the transformers repo"
-    exit 1
-fi
-
-# In these scripts you may have to lower BS if you get CUDA OOM (or increase it if you have a large GPU)
-
-### Normal eval ###
-
-export PAIR=en-de
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=64
-export NUM_BEAMS=5
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-
-MODEL_PATH=allenai/wmt16-en-de-dist-12-1
-echo $PAIR $MODEL_PATH
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py $MODEL_PATH $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-
-MODEL_PATH=allenai/wmt16-en-de-dist-6-1
-echo $PAIR $MODEL_PATH
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py $MODEL_PATH $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-
-MODEL_PATH=allenai/wmt16-en-de-12-1
-echo $PAIR $MODEL_PATH
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py $MODEL_PATH $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-
-
-
-### Searching hparams eval ###
-
-
-export PAIR=en-de
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=32
-export NUM_BEAMS=5
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-
-MODEL_PATH=allenai/wmt16-en-de-dist-12-1
-echo $PAIR $MODEL_PATH
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval_search.py $MODEL_PATH $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --search="num_beams=5:10:15 length_penalty=0.6:0.7:0.8:0.9:1.0:1.1"
-
-
-MODEL_PATH=allenai/wmt16-en-de-dist-6-1
-echo $PAIR $MODEL_PATH
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval_search.py $MODEL_PATH $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --search="num_beams=5:10:15 length_penalty=0.6:0.7:0.8:0.9:1.0:1.1"
-
-
-MODEL_PATH=allenai/wmt16-en-de-12-1
-echo $PAIR $MODEL_PATH
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval_search.py $MODEL_PATH $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --search="num_beams=5:10:15 length_penalty=0.6:0.7:0.8:0.9:1.0:1.1"

scripts/fsmt/eval-allenai-wmt19.sh

@@ -1,67 +0,0 @@
-#!/usr/bin/env bash
-# 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.
-
-# this script evals the following fsmt models
-# it covers:
-# - allenai/wmt19-de-en-6-6-base
-# - allenai/wmt19-de-en-6-6-big
-
-# this script needs to be run from the top level of the transformers repo
-if [ ! -d "src/transformers" ]; then
-    echo "Error: This script needs to be run from the top of the transformers repo"
-    exit 1
-fi
-
-# In these scripts you may have to lower BS if you get CUDA OOM (or increase it if you have a large GPU)
-
-### Normal eval ###
-
-export PAIR=de-en
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=64
-export NUM_BEAMS=5
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-
-MODEL_PATH=allenai/wmt19-de-en-6-6-base
-echo $PAIR $MODEL_PATH
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py $MODEL_PATH $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-
-MODEL_PATH=allenai/wmt19-de-en-6-6-big
-echo $PAIR $MODEL_PATH
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py $MODEL_PATH $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-
-
-
-### Searching hparams eval ###
-
-export PAIR=de-en
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=16
-export NUM_BEAMS=5
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-
-MODEL_PATH=allenai/wmt19-de-en-6-6-base
-echo $PAIR $MODEL_PATH
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval_search.py $MODEL_PATH $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --search="num_beams=5:10:15 length_penalty=0.6:0.7:0.8:0.9:1.0:1.1"
-
-MODEL_PATH=allenai/wmt19-de-en-6-6-big
-echo $PAIR $MODEL_PATH
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval_search.py $MODEL_PATH $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --search="num_beams=5:10:15 length_penalty=0.6:0.7:0.8:0.9:1.0:1.1"

scripts/fsmt/eval-facebook-wmt19.sh

@@ -1,161 +0,0 @@
-#!/usr/bin/env bash
-# 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.
-
-# this script evals the following fsmt models
-# it covers:
-# - facebook/wmt19-ru-en
-# - facebook/wmt19-en-ru
-# - facebook/wmt19-de-en
-# - facebook/wmt19-en-de
-
-
-# this script needs to be run from the top level of the transformers repo
-if [ ! -d "src/transformers" ]; then
-    echo "Error: This script needs to be run from the top of the transformers repo"
-    exit 1
-fi
-
-
-# In these scripts you may have to lower BS if you get CUDA OOM (or increase it if you have a large GPU)
-
-### a short estimate version for quick testing ###
-
-export PAIR=en-ru
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=8
-export NUM_BEAMS=8
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src | head -10 > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref | head -10 > $DATA_DIR/val.target
-echo $PAIR
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-
-
-
-### Normal eval ###
-
-# ru-en
-
-export PAIR=ru-en
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=8
-export NUM_BEAMS=50
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-
-
-# (target BLEU: 41.3 http://matrix.statmt.org/matrix/output/1907?run_id=6937)
-
-
-# en-ru
-
-export PAIR=en-ru
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=8
-export NUM_BEAMS=50
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-echo $PAIR
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-
-# (target BLEU: 36.4 http://matrix.statmt.org/matrix/output/1914?score_id=37605)
-
-
-
-# en-de
-
-export PAIR=en-de
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=8
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-echo $PAIR
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-
-# (target BLEU: 43.1 http://matrix.statmt.org/matrix/output/1909?run_id=6862)
-
-
-# de-en
-
-export PAIR=de-en
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=8
-export NUM_BEAMS=50
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-echo $PAIR
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-
-# (target BLEU: 42.3 http://matrix.statmt.org/matrix/output/1902?run_id=6750)
-
-
-### Searching hparams eval ###
-
-# en-ru
-
-export PAIR=ru-en
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=32
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-CUDA_VISIBLE_DEVICES="0" PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval_search.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --search="num_beams=5 length_penalty=0.6:0.7:0.8:0.9:1.0:1.1"
-
-
-# en-ru
-
-export PAIR=en-ru
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=16
-mkdir -p $DATA_DIR
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-CUDA_VISIBLE_DEVICES="0" PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval_search.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --search="num_beams=5:8:11:15 length_penalty=0.6:0.7:0.8:0.9:1.0:1.1 early_stopping=true:false"
-
-# en-de
-
-export PAIR=en-de
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=16
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-CUDA_VISIBLE_DEVICES="1" PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval_search.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --search="num_beams=5:8:11:15 length_penalty=0.6:0.7:0.8:0.9:1.0:1.1 early_stopping=true:false"
-
-# de-en
-
-export PAIR=de-en
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=16
-mkdir -p $DATA_DIR
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-CUDA_VISIBLE_DEVICES="1" PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval_search.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --search="num_beams=5:8:11:15 length_penalty=0.6:0.7:0.8:0.9:1.0:1.1 early_stopping=true:false"

scripts/fsmt/fsmt-make-super-tiny-model.py

@@ -1,88 +0,0 @@
-#!/usr/bin/env python
-# coding: utf-8
-# Copyright 2020 The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# This script creates a super tiny model that is useful inside tests, when we just want to test that
-# the machinery works, without needing to the check the quality of the outcomes.
-#
-# This version creates a tiny vocab first, and then a tiny model - so the outcome is truly tiny -
-# all files ~60KB. As compared to taking a full-size model, reducing to the minimum its layers and
-# emb dimensions, but keeping the full vocab + merges files, leading to ~3MB in total for all files.
-# The latter is done by `fsmt-make-super-tiny-model.py`.
-#
-# It will be used then as "stas/tiny-wmt19-en-ru"
-
-import json
-import tempfile
-from pathlib import Path
-
-from transformers import FSMTConfig, FSMTForConditionalGeneration, FSMTTokenizer
-from transformers.models.fsmt.tokenization_fsmt import VOCAB_FILES_NAMES
-
-
-mname_tiny = "tiny-wmt19-en-ru"
-
-# Build
-
-# borrowed from a test
-vocab = [ "l", "o", "w", "e", "r", "s", "t", "i", "d", "n", "w</w>", "r</w>", "t</w>", "lo", "low", "er</w>", "low</w>", "lowest</w>", "newer</w>", "wider</w>", "<unk>", ]
-vocab_tokens = dict(zip(vocab, range(len(vocab))))
-merges = ["l o 123", "lo w 1456", "e r</w> 1789", ""]
-
-with tempfile.TemporaryDirectory() as tmpdirname:
-    build_dir = Path(tmpdirname)
-    src_vocab_file = build_dir / VOCAB_FILES_NAMES["src_vocab_file"]
-    tgt_vocab_file = build_dir / VOCAB_FILES_NAMES["tgt_vocab_file"]
-    merges_file = build_dir / VOCAB_FILES_NAMES["merges_file"]
-    with open(src_vocab_file, "w") as fp: fp.write(json.dumps(vocab_tokens))
-    with open(tgt_vocab_file, "w") as fp: fp.write(json.dumps(vocab_tokens))
-    with open(merges_file, "w") as fp   : fp.write("\n".join(merges))
-
-    tokenizer = FSMTTokenizer(
-        langs=["en", "ru"],
-        src_vocab_size = len(vocab),
-        tgt_vocab_size = len(vocab),
-        src_vocab_file=src_vocab_file,
-        tgt_vocab_file=tgt_vocab_file,
-        merges_file=merges_file,
-    )
-
-config = FSMTConfig(
-    langs=['ru', 'en'],
-    src_vocab_size=1000, tgt_vocab_size=1000,
-    d_model=4,
-    encoder_layers=1, decoder_layers=1,
-    encoder_ffn_dim=4, decoder_ffn_dim=4,
-    encoder_attention_heads=1, decoder_attention_heads=1,
-)
-
-tiny_model = FSMTForConditionalGeneration(config)
-print(f"num of params {tiny_model.num_parameters()}")
-
-# Test
-batch = tokenizer(["Making tiny model"], return_tensors="pt")
-outputs = tiny_model(**batch)
-
-print("test output:", len(outputs.logits[0]))
-
-# Save
-tiny_model.half() # makes it smaller
-tiny_model.save_pretrained(mname_tiny)
-tokenizer.save_pretrained(mname_tiny)
-
-print(f"Generated {mname_tiny}")
-
-# Upload
-# transformers-cli upload tiny-wmt19-en-ru

scripts/fsmt/fsmt-make-tiny-model.py

@@ -1,61 +0,0 @@
-#!/usr/bin/env python
-# coding: utf-8
-# Copyright 2020 The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# This script creates a super tiny model that is useful inside tests, when we just want to test that
-# the machinery works, without needing to the check the quality of the outcomes.
-#
-# This version creates a tiny model through reduction of a normal pre-trained model, but keeping the
-# full vocab, merges file, and thus also resulting in a larger model due to a large vocab size.
-# This gives ~3MB in total for all files.
-#
-# If you want a 50 times smaller than this see `fsmt-make-super-tiny-model.py`, which is slightly more complicated
-#
-#
-# It will be used then as "stas/tiny-wmt19-en-de"
-
-# Build
-from transformers import FSMTConfig, FSMTForConditionalGeneration, FSMTTokenizer
-
-
-mname = "facebook/wmt19-en-de"
-tokenizer = FSMTTokenizer.from_pretrained(mname)
-# get the correct vocab sizes, etc. from the master model
-config = FSMTConfig.from_pretrained(mname)
-config.update({
-    "d_model": 4,
-    "encoder_layers": 1, "decoder_layers": 1,
-    "encoder_ffn_dim": 4, "decoder_ffn_dim": 4,
-    "encoder_attention_heads": 1, "decoder_attention_heads": 1})
-
-tiny_model = FSMTForConditionalGeneration(config)
-print(f"num of params {tiny_model.num_parameters()}")
-
-# Test
-batch = tokenizer(["Making tiny model"], return_tensors="pt")
-outputs = tiny_model(**batch)
-
-print("test output:", len(outputs.logits[0]))
-
-# Save
-mname_tiny = "tiny-wmt19-en-de"
-tiny_model.half() # makes it smaller
-tiny_model.save_pretrained(mname_tiny)
-tokenizer.save_pretrained(mname_tiny)
-
-print(f"Generated {mname_tiny}")
-
-# Upload
-# transformers-cli upload tiny-wmt19-en-de

scripts/fsmt/gen-card-allenai-wmt16.py

@@ -1,156 +0,0 @@
-#!/usr/bin/env python
-# 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.
-
-# Usage:
-# ./gen-card-allenai-wmt16.py
-
-import os
-from pathlib import Path
-
-
-def write_model_card(model_card_dir, src_lang, tgt_lang, model_name):
-
-    texts = {
-        "en": "Machine learning is great, isn't it?",
-        "ru": "Машинное обучение - это здорово, не так ли?",
-        "de": "Maschinelles Lernen ist großartig, nicht wahr?",
-    }
-
-    # BLUE scores as follows:
-    # "pair": [fairseq, transformers]
-    scores = {
-        "wmt16-en-de-dist-12-1": [28.3, 27.52],
-        "wmt16-en-de-dist-6-1": [27.4, 27.11],
-        "wmt16-en-de-12-1": [26.9, 25.75],
-    }
-    pair = f"{src_lang}-{tgt_lang}"
-
-    readme = f"""
----
-language:
-- {src_lang}
-- {tgt_lang}
-thumbnail:
-tags:
-- translation
-- wmt16
-- allenai
-license: apache-2.0
-datasets:
-- wmt16
-metrics:
-- bleu
----
-
-# FSMT
-
-## Model description
-
-This is a ported version of fairseq-based [wmt16 transformer](https://github.com/jungokasai/deep-shallow/) for {src_lang}-{tgt_lang}.
-
-For more details, please, see [Deep Encoder, Shallow Decoder: Reevaluating the Speed-Quality Tradeoff in Machine Translation](https://arxiv.org/abs/2006.10369).
-
-All 3 models are available:
-
-* [wmt16-en-de-dist-12-1](https://huggingface.co/allenai/wmt16-en-de-dist-12-1)
-* [wmt16-en-de-dist-6-1](https://huggingface.co/allenai/wmt16-en-de-dist-6-1)
-* [wmt16-en-de-12-1](https://huggingface.co/allenai/wmt16-en-de-12-1)
-
-
-## Intended uses & limitations
-
-#### How to use
-
-```python
-from transformers import FSMTForConditionalGeneration, FSMTTokenizer
-mname = "allenai/{model_name}"
-tokenizer = FSMTTokenizer.from_pretrained(mname)
-model = FSMTForConditionalGeneration.from_pretrained(mname)
-
-input = "{texts[src_lang]}"
-input_ids = tokenizer.encode(input, return_tensors="pt")
-outputs = model.generate(input_ids)
-decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
-print(decoded) # {texts[tgt_lang]}
-
-```
-
-#### Limitations and bias
-
-
-## Training data
-
-Pretrained weights were left identical to the original model released by allenai. For more details, please, see the [paper](https://arxiv.org/abs/2006.10369).
-
-## Eval results
-
-Here are the BLEU scores:
-
-model   | fairseq | transformers
--------|---------|----------
-{model_name}  | {scores[model_name][0]} | {scores[model_name][1]}
-
-The score is slightly below the score reported in the paper, as the researchers don't use `sacrebleu` and measure the score on tokenized outputs. `transformers` score was measured using `sacrebleu` on detokenized outputs.
-
-The score was calculated using this code:
-
-```bash
-git clone https://github.com/huggingface/transformers
-cd transformers
-export PAIR={pair}
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=8
-export NUM_BEAMS=5
-mkdir -p $DATA_DIR
-sacrebleu -t wmt16 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt16 -l $PAIR --echo ref > $DATA_DIR/val.target
-echo $PAIR
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py allenai/{model_name} $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-```
-
-## Data Sources
-
-- [training, etc.](http://www.statmt.org/wmt16/)
-- [test set](http://matrix.statmt.org/test_sets/newstest2016.tgz?1504722372)
-
-
-### BibTeX entry and citation info
-
-```
-@misc{{kasai2020deep,
-    title={{Deep Encoder, Shallow Decoder: Reevaluating the Speed-Quality Tradeoff in Machine Translation}},
-    author={{Jungo Kasai and Nikolaos Pappas and Hao Peng and James Cross and Noah A. Smith}},
-    year={{2020}},
-    eprint={{2006.10369}},
-    archivePrefix={{arXiv}},
-    primaryClass={{cs.CL}}
-}}
-```
-
-"""
-    model_card_dir.mkdir(parents=True, exist_ok=True)
-    path = os.path.join(model_card_dir, "README.md")
-    print(f"Generating {path}")
-    with open(path, "w", encoding="utf-8") as f:
-        f.write(readme)
-
-# make sure we are under the root of the project
-repo_dir = Path(__file__).resolve().parent.parent.parent
-model_cards_dir = repo_dir / "model_cards"
-
-for model_name in ["wmt16-en-de-dist-12-1", "wmt16-en-de-dist-6-1", "wmt16-en-de-12-1"]:
-    model_card_dir = model_cards_dir / "allenai" / model_name
-    write_model_card(model_card_dir, src_lang="en", tgt_lang="de", model_name=model_name)

scripts/fsmt/gen-card-allenai-wmt19.py

@@ -1,153 +0,0 @@
-#!/usr/bin/env python
-# 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.
-
-# Usage:
-# ./gen-card-allenai-wmt19.py
-
-import os
-from pathlib import Path
-
-
-def write_model_card(model_card_dir, src_lang, tgt_lang, model_name):
-
-    texts = {
-        "en": "Machine learning is great, isn't it?",
-        "ru": "Машинное обучение - это здорово, не так ли?",
-        "de": "Maschinelles Lernen ist großartig, nicht wahr?",
-    }
-
-    # BLUE scores as follows:
-    # "pair": [fairseq, transformers]
-    scores = {
-        "wmt19-de-en-6-6-base": [0, 38.37],
-        "wmt19-de-en-6-6-big": [0, 39.90],
-    }
-    pair = f"{src_lang}-{tgt_lang}"
-
-    readme = f"""
----
-
-language:
-- {src_lang}
-- {tgt_lang}
-thumbnail:
-tags:
-- translation
-- wmt19
-- allenai
-license: apache-2.0
-datasets:
-- wmt19
-metrics:
-- bleu
----
-
-# FSMT
-
-## Model description
-
-This is a ported version of fairseq-based [wmt19 transformer](https://github.com/jungokasai/deep-shallow/) for {src_lang}-{tgt_lang}.
-
-For more details, please, see [Deep Encoder, Shallow Decoder: Reevaluating the Speed-Quality Tradeoff in Machine Translation](https://arxiv.org/abs/2006.10369).
-
-2 models are available:
-
-* [wmt19-de-en-6-6-big](https://huggingface.co/allenai/wmt19-de-en-6-6-big)
-* [wmt19-de-en-6-6-base](https://huggingface.co/allenai/wmt19-de-en-6-6-base)
-
-
-## Intended uses & limitations
-
-#### How to use
-
-```python
-from transformers import FSMTForConditionalGeneration, FSMTTokenizer
-mname = "allenai/{model_name}"
-tokenizer = FSMTTokenizer.from_pretrained(mname)
-model = FSMTForConditionalGeneration.from_pretrained(mname)
-
-input = "{texts[src_lang]}"
-input_ids = tokenizer.encode(input, return_tensors="pt")
-outputs = model.generate(input_ids)
-decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
-print(decoded) # {texts[tgt_lang]}
-
-```
-
-#### Limitations and bias
-
-
-## Training data
-
-Pretrained weights were left identical to the original model released by allenai. For more details, please, see the [paper](https://arxiv.org/abs/2006.10369).
-
-## Eval results
-
-Here are the BLEU scores:
-
-model   |  transformers
--------|---------
-{model_name}  |  {scores[model_name][1]}
-
-The score was calculated using this code:
-
-```bash
-git clone https://github.com/huggingface/transformers
-cd transformers
-export PAIR={pair}
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=8
-export NUM_BEAMS=5
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-echo $PAIR
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py allenai/{model_name} $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-```
-
-## Data Sources
-
-- [training, etc.](http://www.statmt.org/wmt19/)
-- [test set](http://matrix.statmt.org/test_sets/newstest2019.tgz?1556572561)
-
-
-### BibTeX entry and citation info
-
-```
-@misc{{kasai2020deep,
-    title={{Deep Encoder, Shallow Decoder: Reevaluating the Speed-Quality Tradeoff in Machine Translation}},
-    author={{Jungo Kasai and Nikolaos Pappas and Hao Peng and James Cross and Noah A. Smith}},
-    year={{2020}},
-    eprint={{2006.10369}},
-    archivePrefix={{arXiv}},
-    primaryClass={{cs.CL}}
-}}
-```
-
-"""
-    model_card_dir.mkdir(parents=True, exist_ok=True)
-    path = os.path.join(model_card_dir, "README.md")
-    print(f"Generating {path}")
-    with open(path, "w", encoding="utf-8") as f:
-        f.write(readme)
-
-# make sure we are under the root of the project
-repo_dir = Path(__file__).resolve().parent.parent.parent
-model_cards_dir = repo_dir / "model_cards"
-
-for model_name in ["wmt19-de-en-6-6-base", "wmt19-de-en-6-6-big"]:
-    model_card_dir = model_cards_dir / "allenai" / model_name
-    write_model_card(model_card_dir, src_lang="de", tgt_lang="en", model_name=model_name)

scripts/fsmt/gen-card-facebook-wmt19.py

@@ -1,165 +0,0 @@
-#!/usr/bin/env python
-# 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.
-
-# Usage:
-# ./gen-card-facebook-wmt19.py
-
-import os
-from pathlib import Path
-
-
-def write_model_card(model_card_dir, src_lang, tgt_lang):
-
-    texts = {
-        "en": "Machine learning is great, isn't it?",
-        "ru": "Машинное обучение - это здорово, не так ли?",
-        "de": "Maschinelles Lernen ist großartig, oder?",
-    }
-
-    # BLUE scores as follows:
-    # "pair": [fairseq, transformers]
-    scores = {
-        "ru-en": ["[41.3](http://matrix.statmt.org/matrix/output/1907?run_id=6937)", "39.20"],
-        "en-ru": ["[36.4](http://matrix.statmt.org/matrix/output/1914?run_id=6724)", "33.47"],
-        "en-de": ["[43.1](http://matrix.statmt.org/matrix/output/1909?run_id=6862)", "42.83"],
-        "de-en": ["[42.3](http://matrix.statmt.org/matrix/output/1902?run_id=6750)", "41.35"],
-    }
-    pair = f"{src_lang}-{tgt_lang}"
-
-    readme = f"""
----
-language:
-- {src_lang}
-- {tgt_lang}
-thumbnail:
-tags:
-- translation
-- wmt19
-- facebook
-license: apache-2.0
-datasets:
-- wmt19
-metrics:
-- bleu
----
-
-# FSMT
-
-## Model description
-
-This is a ported version of [fairseq wmt19 transformer](https://github.com/pytorch/fairseq/blob/master/examples/wmt19/README.md) for {src_lang}-{tgt_lang}.
-
-For more details, please see, [Facebook FAIR's WMT19 News Translation Task Submission](https://arxiv.org/abs/1907.06616).
-
-The abbreviation FSMT stands for FairSeqMachineTranslation
-
-All four models are available:
-
-* [wmt19-en-ru](https://huggingface.co/facebook/wmt19-en-ru)
-* [wmt19-ru-en](https://huggingface.co/facebook/wmt19-ru-en)
-* [wmt19-en-de](https://huggingface.co/facebook/wmt19-en-de)
-* [wmt19-de-en](https://huggingface.co/facebook/wmt19-de-en)
-
-## Intended uses & limitations
-
-#### How to use
-
-```python
-from transformers import FSMTForConditionalGeneration, FSMTTokenizer
-mname = "facebook/wmt19-{src_lang}-{tgt_lang}"
-tokenizer = FSMTTokenizer.from_pretrained(mname)
-model = FSMTForConditionalGeneration.from_pretrained(mname)
-
-input = "{texts[src_lang]}"
-input_ids = tokenizer.encode(input, return_tensors="pt")
-outputs = model.generate(input_ids)
-decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
-print(decoded) # {texts[tgt_lang]}
-
-```
-
-#### Limitations and bias
-
-- The original (and this ported model) doesn't seem to handle well inputs with repeated sub-phrases, [content gets truncated](https://discuss.huggingface.co/t/issues-with-translating-inputs-containing-repeated-phrases/981)
-
-## Training data
-
-Pretrained weights were left identical to the original model released by fairseq. For more details, please, see the [paper](https://arxiv.org/abs/1907.06616).
-
-## Eval results
-
-pair   | fairseq | transformers
--------|---------|----------
-{pair}  | {scores[pair][0]} | {scores[pair][1]}
-
-The score is slightly below the score reported by `fairseq`, since `transformers`` currently doesn't support:
-- model ensemble, therefore the best performing checkpoint was ported (``model4.pt``).
-- re-ranking
-
-The score was calculated using this code:
-
-```bash
-git clone https://github.com/huggingface/transformers
-cd transformers
-export PAIR={pair}
-export DATA_DIR=data/$PAIR
-export SAVE_DIR=data/$PAIR
-export BS=8
-export NUM_BEAMS=15
-mkdir -p $DATA_DIR
-sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
-sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
-echo $PAIR
-PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
-```
-note: fairseq reports using a beam of 50, so you should get a slightly higher score if re-run with `--num_beams 50`.
-
-## Data Sources
-
-- [training, etc.](http://www.statmt.org/wmt19/)
-- [test set](http://matrix.statmt.org/test_sets/newstest2019.tgz?1556572561)
-
-
-### BibTeX entry and citation info
-
-```bibtex
-@inproceedings{{...,
-  year={{2020}},
-  title={{Facebook FAIR's WMT19 News Translation Task Submission}},
-  author={{Ng, Nathan and Yee, Kyra and Baevski, Alexei and Ott, Myle and Auli, Michael and Edunov, Sergey}},
-  booktitle={{Proc. of WMT}},
-}}
-```
-
-
-## TODO
-
-- port model ensemble (fairseq uses 4 model checkpoints)
-
-"""
-    os.makedirs(model_card_dir, exist_ok=True)
-    path = os.path.join(model_card_dir, "README.md")
-    print(f"Generating {path}")
-    with open(path, "w", encoding="utf-8") as f:
-        f.write(readme)
-
-# make sure we are under the root of the project
-repo_dir = Path(__file__).resolve().parent.parent.parent
-model_cards_dir = repo_dir / "model_cards"
-
-for model_name in ["wmt19-ru-en", "wmt19-en-ru", "wmt19-en-de", "wmt19-de-en"]:
-    base, src_lang, tgt_lang = model_name.split("-")
-    model_card_dir = model_cards_dir / "facebook" / model_name
-    write_model_card(model_card_dir, src_lang=src_lang, tgt_lang=tgt_lang)

scripts/fsmt/s3-move.sh

@@ -1,116 +0,0 @@
-
-# this is the process of uploading the updated models to s3. As I can't upload them directly to the correct orgs, this script shows how this is done
-# 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.
-
-1. upload updated models to my account
-
-transformers-cli upload -y wmt19-ru-en
-transformers-cli upload -y wmt19-en-ru
-transformers-cli upload -y wmt19-de-en
-transformers-cli upload -y wmt19-en-de
-transformers-cli upload -y wmt19-de-en-6-6-base
-transformers-cli upload -y wmt19-de-en-6-6-big
-transformers-cli upload -y wmt16-en-de-dist-12-1
-transformers-cli upload -y wmt16-en-de-dist-6-1
-transformers-cli upload -y wmt16-en-de-12-1
-
-
-2. ask someone to move them to:
-
-* to facebook: "wmt19-ru-en", "wmt19-en-ru", "wmt19-en-de", "wmt19-de-en"
-* to allenai: "wmt16-en-de-dist-12-1", "wmt16-en-de-dist-6-1", "wmt16-en-de-12-1", "wmt19-de-en-6-6-base", "wmt19-de-en-6-6-big"
-
-export b="s3://models.huggingface.co/bert"
-stas_to_fb () {
-	src=$1
-	shift
-	aws s3 sync $b/stas/$src $b/facebook/$src $@
-}
-
-stas_to_allenai () {
-	src=$1
-	shift
-	aws s3 sync $b/stas/$src $b/allenai/$src $@
-}
-
-stas_to_fb wmt19-en-ru
-stas_to_fb wmt19-ru-en
-stas_to_fb wmt19-en-de
-stas_to_fb wmt19-de-en
-
-stas_to_allenai wmt16-en-de-dist-12-1
-stas_to_allenai wmt16-en-de-dist-6-1
-stas_to_allenai wmt16-en-de-6-1
-stas_to_allenai wmt16-en-de-12-1
-stas_to_allenai wmt19-de-en-6-6-base
-stas_to_allenai wmt19-de-en-6-6-big
-
-
-3. and then remove all these model files from my account
-
-transformers-cli s3 rm wmt16-en-de-12-1/config.json
-transformers-cli s3 rm wmt16-en-de-12-1/merges.txt
-transformers-cli s3 rm wmt16-en-de-12-1/pytorch_model.bin
-transformers-cli s3 rm wmt16-en-de-12-1/tokenizer_config.json
-transformers-cli s3 rm wmt16-en-de-12-1/vocab-src.json
-transformers-cli s3 rm wmt16-en-de-12-1/vocab-tgt.json
-transformers-cli s3 rm wmt16-en-de-dist-12-1/config.json
-transformers-cli s3 rm wmt16-en-de-dist-12-1/merges.txt
-transformers-cli s3 rm wmt16-en-de-dist-12-1/pytorch_model.bin
-transformers-cli s3 rm wmt16-en-de-dist-12-1/tokenizer_config.json
-transformers-cli s3 rm wmt16-en-de-dist-12-1/vocab-src.json
-transformers-cli s3 rm wmt16-en-de-dist-12-1/vocab-tgt.json
-transformers-cli s3 rm wmt16-en-de-dist-6-1/config.json
-transformers-cli s3 rm wmt16-en-de-dist-6-1/merges.txt
-transformers-cli s3 rm wmt16-en-de-dist-6-1/pytorch_model.bin
-transformers-cli s3 rm wmt16-en-de-dist-6-1/tokenizer_config.json
-transformers-cli s3 rm wmt16-en-de-dist-6-1/vocab-src.json
-transformers-cli s3 rm wmt16-en-de-dist-6-1/vocab-tgt.json
-transformers-cli s3 rm wmt19-de-en-6-6-base/config.json
-transformers-cli s3 rm wmt19-de-en-6-6-base/merges.txt
-transformers-cli s3 rm wmt19-de-en-6-6-base/pytorch_model.bin
-transformers-cli s3 rm wmt19-de-en-6-6-base/tokenizer_config.json
-transformers-cli s3 rm wmt19-de-en-6-6-base/vocab-src.json
-transformers-cli s3 rm wmt19-de-en-6-6-base/vocab-tgt.json
-transformers-cli s3 rm wmt19-de-en-6-6-big/config.json
-transformers-cli s3 rm wmt19-de-en-6-6-big/merges.txt
-transformers-cli s3 rm wmt19-de-en-6-6-big/pytorch_model.bin
-transformers-cli s3 rm wmt19-de-en-6-6-big/tokenizer_config.json
-transformers-cli s3 rm wmt19-de-en-6-6-big/vocab-src.json
-transformers-cli s3 rm wmt19-de-en-6-6-big/vocab-tgt.json
-transformers-cli s3 rm wmt19-de-en/config.json
-transformers-cli s3 rm wmt19-de-en/merges.txt
-transformers-cli s3 rm wmt19-de-en/pytorch_model.bin
-transformers-cli s3 rm wmt19-de-en/tokenizer_config.json
-transformers-cli s3 rm wmt19-de-en/vocab-src.json
-transformers-cli s3 rm wmt19-de-en/vocab-tgt.json
-transformers-cli s3 rm wmt19-en-de/config.json
-transformers-cli s3 rm wmt19-en-de/merges.txt
-transformers-cli s3 rm wmt19-en-de/pytorch_model.bin
-transformers-cli s3 rm wmt19-en-de/tokenizer_config.json
-transformers-cli s3 rm wmt19-en-de/vocab-src.json
-transformers-cli s3 rm wmt19-en-de/vocab-tgt.json
-transformers-cli s3 rm wmt19-en-ru/config.json
-transformers-cli s3 rm wmt19-en-ru/merges.txt
-transformers-cli s3 rm wmt19-en-ru/pytorch_model.bin
-transformers-cli s3 rm wmt19-en-ru/tokenizer_config.json
-transformers-cli s3 rm wmt19-en-ru/vocab-src.json
-transformers-cli s3 rm wmt19-en-ru/vocab-tgt.json
-transformers-cli s3 rm wmt19-ru-en/config.json
-transformers-cli s3 rm wmt19-ru-en/merges.txt
-transformers-cli s3 rm wmt19-ru-en/pytorch_model.bin
-transformers-cli s3 rm wmt19-ru-en/tokenizer_config.json
-transformers-cli s3 rm wmt19-ru-en/vocab-src.json
-transformers-cli s3 rm wmt19-ru-en/vocab-tgt.json

scripts/pegasus/build_test_sample_spm_no_bos.py

@@ -1,34 +0,0 @@
-#!/usr/bin/env python
-# 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.
-
-# this script builds a small sample spm file tests/fixtures/test_sentencepiece_no_bos.model, with features needed by pegasus
-
-# 1. pip install sentencepiece
-#
-# 2. wget https://raw.githubusercontent.com/google/sentencepiece/master/data/botchan.txt
-
-# 3. build
-import sentencepiece as spm
-
-
-# pegasus:
-# 1. no bos
-# 2. eos_id is 1
-# 3. unk_id is 2
-# build a sample spm file accordingly
-spm.SentencePieceTrainer.train('--input=botchan.txt --model_prefix=test_sentencepiece_no_bos --bos_id=-1 --unk_id=2  --eos_id=1  --vocab_size=1000')
-
-# 4. now update the fixture
-# mv test_sentencepiece_no_bos.model ../../tests/fixtures/

scripts/tatoeba/README.md

@@ -1,72 +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.
--->
-
-Setup transformers following instructions in README.md, (I would fork first).
-```bash
-git clone git@github.com:huggingface/transformers.git
-cd transformers
-pip install -e .
-pip install pandas GitPython wget
-```
-
-Get required metadata
-```bash
-curl https://cdn-datasets.huggingface.co/language_codes/language-codes-3b2.csv  > language-codes-3b2.csv
-curl https://cdn-datasets.huggingface.co/language_codes/iso-639-3.csv > iso-639-3.csv
-```
-
-Install Tatoeba-Challenge repo inside transformers
-```bash
-git clone git@github.com:Helsinki-NLP/Tatoeba-Challenge.git
-```
-
-To convert a few models, call the conversion script from command line:
-```bash
-python src/transformers/models/marian/convert_marian_tatoeba_to_pytorch.py --models heb-eng eng-heb --save_dir converted
-```
-
-To convert lots of models you can pass your list of Tatoeba model names to `resolver.convert_models` in a python client or script.
-
-```python
-from transformers.convert_marian_tatoeba_to_pytorch import TatoebaConverter
-resolver = TatoebaConverter(save_dir='converted')
-resolver.convert_models(['heb-eng', 'eng-heb'])
-```
-
-
-### Upload converted models
-Since version v3.5.0, the model sharing workflow is switched to git-based system . Refer to [model sharing doc](https://huggingface.co/transformers/main/model_sharing.html#model-sharing-and-uploading) for more details.
-
-To upload all converted models, 
-
-1. Install [git-lfs](https://git-lfs.github.com/).
-
-2. Login to `huggingface-cli`
-
-```bash
-huggingface-cli login
-```
-
-3. Run the `upload_models` script
-
-```bash
-./scripts/tatoeba/upload_models.sh
-```
-
-
-### Modifications
-- To change naming logic, change the code near `os.rename`. The model card creation code may also need to change.
-- To change model card content, you must modify `TatoebaCodeResolver.write_model_card`

scripts/tatoeba/upload_models.sh

@@ -1,12 +0,0 @@
-#!/bin/bash
-
-for FILE in converted/*; do 
-  model_name=`basename $FILE`
-  huggingface-cli repo create $model_name -y
-  git clone https://huggingface.co/Helsinki-NLP/$model_name
-  mv $FILE/* $model_name/
-  cd $model_name
-  git add . && git commit -m "initial commit" 
-  git push
-  cd ..
-done