Merge remote-tracking branch 'upstream/main' into add-Dinov2-with-reg…

…isters-attempt-2

docs/source/en/_toctree.yml

@@ -659,6 +659,8 @@
         title: GLPN
       - local: model_doc/hiera
         title: Hiera
+      - local: model_doc/ijepa
+        title: I-JEPA
       - local: model_doc/imagegpt
         title: ImageGPT
       - local: model_doc/levit

docs/source/en/index.md

@@ -169,6 +169,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                         [Hiera](model_doc/hiera)                         |       ✅        |         ❌         |      ❌      |
 |                        [Hubert](model_doc/hubert)                        |       ✅        |         ✅         |      ❌      |
 |                        [I-BERT](model_doc/ibert)                         |       ✅        |         ❌         |      ❌      |
+|                        [I-JEPA](model_doc/ijepa)                         |       ✅        |         ❌         |      ❌      |
 |                       [IDEFICS](model_doc/idefics)                       |       ✅        |         ✅         |      ❌      |
 |                      [Idefics2](model_doc/idefics2)                      |       ✅        |         ❌         |      ❌      |
 |                      [Idefics3](model_doc/idefics3)                      |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/ijepa.md

@@ -0,0 +1,78 @@
+<!--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.
+
+-->
+
+# I-JEPA
+
+## Overview
+
+The I-JEPA model was proposed in [Image-based Joint-Embedding Predictive Architecture](https://arxiv.org/pdf/2301.08243.pdf) by Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, Nicolas Ballas.
+I-JEPA is a self-supervised learning method that predicts the representations of one part of an image based on other parts of the same image. This approach focuses on learning semantic features without relying on pre-defined invariances from hand-crafted data transformations, which can bias specific tasks, or on filling in pixel-level details, which often leads to less meaningful representations.
+
+The abstract from the paper is the following:
+
+This paper demonstrates an approach for learning highly semantic image representations without relying on hand-crafted data-augmentations. We introduce the Image- based Joint-Embedding Predictive Architecture (I-JEPA), a non-generative approach for self-supervised learning from images. The idea behind I-JEPA is simple: from a single context block, predict the representations of various target blocks in the same image. A core design choice to guide I-JEPA towards producing semantic representations is the masking strategy; specifically, it is crucial to (a) sample tar- get blocks with sufficiently large scale (semantic), and to (b) use a sufficiently informative (spatially distributed) context block. Empirically, when combined with Vision Transform- ers, we find I-JEPA to be highly scalable. For instance, we train a ViT-Huge/14 on ImageNet using 16 A100 GPUs in under 72 hours to achieve strong downstream performance across a wide range of tasks, from linear classification to object counting and depth prediction.
+
+This model was contributed by [jmtzt](https://huggingface.co/jmtzt).
+The original code can be found [here](https://github.com/facebookresearch/ijepa).
+
+## How to use
+
+Here is how to use this model for image feature extraction:
+
+```python
+import requests
+import torch
+from PIL import Image
+from torch.nn.functional import cosine_similarity
+
+from transformers import AutoModel, AutoProcessor
+
+url_1 = "http://images.cocodataset.org/val2017/000000039769.jpg"
+url_2 = "http://images.cocodataset.org/val2017/000000219578.jpg"
+image_1 = Image.open(requests.get(url_1, stream=True).raw)
+image_2 = Image.open(requests.get(url_2, stream=True).raw)
+
+model_id = "jmtzt/ijepa_vith14_1k"
+processor = AutoProcessor.from_pretrained(model_id)
+model = AutoModel.from_pretrained(model_id)
+
+@torch.no_grad()
+def infer(image):
+    inputs = processor(image, return_tensors="pt")
+    outputs = model(**inputs)
+    return outputs.last_hidden_state.mean(dim=1)
+
+
+embed_1 = infer(image_1)
+embed_2 = infer(image_2)
+
+similarity = cosine_similarity(embed_1, embed_2)
+print(similarity)
+```
+
+## IJepaConfig
+
+[[autodoc]] IJepaConfig
+
+## IJepaModel
+
+[[autodoc]] IJepaModel
+    - forward
+
+## IJepaForImageClassification
+
+[[autodoc]] IJepaForImageClassification
+    - forward

docs/source/en/perf_infer_gpu_one.md

@@ -235,14 +235,15 @@ For now, Transformers supports SDPA inference and training for the following arc
 * [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
 * [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel)
 * [Gemma2](https://huggingface.co/docs/transformers/model_doc/gemma2#transformers.Gemma2Model)
+* [Granite](https://huggingface.co/docs/transformers/model_doc/granite#transformers.GraniteModel)
 * [GPT2](https://huggingface.co/docs/transformers/model_doc/gpt2)
 * [GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode#transformers.GPTBigCodeModel)
 * [GPTNeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox#transformers.GPTNeoXModel)
 * [Hubert](https://huggingface.co/docs/transformers/model_doc/hubert#transformers.HubertModel)
 * [Idefics](https://huggingface.co/docs/transformers/model_doc/idefics#transformers.IdeficsModel)
 * [Idefics2](https://huggingface.co/docs/transformers/model_doc/idefics2#transformers.Idefics2Model)
 * [Idefics3](https://huggingface.co/docs/transformers/model_doc/idefics3#transformers.Idefics3Model)
-* [Granite](https://huggingface.co/docs/transformers/model_doc/granite#transformers.GraniteModel)
+* [I-JEPA](https://huggingface.co/docs/transformers/model_doc/ijepa#transformers.IJepaModel)
 * [GraniteMoe](https://huggingface.co/docs/transformers/model_doc/granitemoe#transformers.GraniteMoeModel)
 * [JetMoe](https://huggingface.co/docs/transformers/model_doc/jetmoe#transformers.JetMoeModel)
 * [Jamba](https://huggingface.co/docs/transformers/model_doc/jamba#transformers.JambaModel)

setup.py

@@ -180,7 +180,7 @@
     "timeout-decorator",
     "tiktoken",
     "timm<=1.0.11",
-    "tokenizers>=0.20,<0.21",
+    "tokenizers>=0.21,<0.22",
     "torch",
     "torchaudio",
     "torchvision",

src/transformers/__init__.py

@@ -486,6 +486,7 @@
     "models.idefics": ["IdeficsConfig"],
     "models.idefics2": ["Idefics2Config"],
     "models.idefics3": ["Idefics3Config"],
+    "models.ijepa": ["IJepaConfig"],
     "models.imagegpt": ["ImageGPTConfig"],
     "models.informer": ["InformerConfig"],
     "models.instructblip": [
@@ -2471,6 +2472,13 @@
             "Idefics3Processor",
         ]
     )
+    _import_structure["models.ijepa"].extend(
+        [
+            "IJepaForImageClassification",
+            "IJepaModel",
+            "IJepaPreTrainedModel",
+        ]
+    )
     _import_structure["models.imagegpt"].extend(
         [
             "ImageGPTForCausalImageModeling",
@@ -5378,6 +5386,7 @@
     )
     from .models.idefics2 import Idefics2Config
     from .models.idefics3 import Idefics3Config
+    from .models.ijepa import IJepaConfig
     from .models.imagegpt import ImageGPTConfig
     from .models.informer import InformerConfig
     from .models.instructblip import (
@@ -7197,6 +7206,11 @@
             Idefics3PreTrainedModel,
             Idefics3Processor,
         )
+        from .models.ijepa import (
+            IJepaForImageClassification,
+            IJepaModel,
+            IJepaPreTrainedModel,
+        )
         from .models.imagegpt import (
             ImageGPTForCausalImageModeling,
             ImageGPTForImageClassification,

src/transformers/cache_utils.py

@@ -12,7 +12,6 @@
 from .utils import (
     is_hqq_available,
     is_optimum_quanto_available,
-    is_quanto_available,
     is_torchdynamo_compiling,
     logging,
 )
@@ -790,17 +789,6 @@ def __init__(self, cache_config: CacheConfig) -> None:
                     f"You need optimum-quanto package version to be greater or equal than 0.2.5 to use `QuantoQuantizedCache`. Detected version {optimum_quanto_version}."
                 )
             from optimum.quanto import MaxOptimizer, qint2, qint4
-        elif is_quanto_available():
-            logger.warning_once(
-                "Importing from quanto will be deprecated in v4.47. Please install optimum-quanto instead `pip install optimum-quanto`"
-            )
-            quanto_version = version.parse(importlib.metadata.version("quanto"))
-            if quanto_version < version.parse("0.2.0"):
-                raise ImportError(
-                    f"You need quanto package version to be greater or equal than 0.2.0 to use `QuantoQuantizedCache`. Detected version {quanto_version}. "
-                    f"Since quanto will be deprecated, please install optimum-quanto instead with `pip install -U optimum-quanto`"
-                )
-            from quanto import MaxOptimizer, qint2, qint4
 
         if self.nbits not in [2, 4]:
             raise ValueError(f"`nbits` for `quanto` backend has to be one of [`2`, `4`] but got {self.nbits}")
@@ -824,16 +812,6 @@ def _quantize(self, tensor, axis):
             scale, zeropoint = self.optimizer(tensor, self.qtype, axis, self.q_group_size)
             qtensor = quantize_weight(tensor, self.qtype, axis, scale, zeropoint, self.q_group_size)
             return qtensor
-        elif is_quanto_available():
-            logger.warning_once(
-                "Importing from quanto will be deprecated in v4.47. Please install optimum-quanto instead `pip install optimum-quanto`"
-            )
-            from quanto import AffineQuantizer
-
-            scale, zeropoint = self.optimizer(tensor, self.qtype.bits, axis, self.q_group_size)
-            qtensor = AffineQuantizer.apply(tensor, self.qtype, axis, self.q_group_size, scale, zeropoint)
-
-        return qtensor
 
     def _dequantize(self, qtensor):
         return qtensor.dequantize()

src/transformers/dependency_versions_table.py

@@ -85,7 +85,7 @@
     "timeout-decorator": "timeout-decorator",
     "tiktoken": "tiktoken",
     "timm": "timm<=1.0.11",
-    "tokenizers": "tokenizers>=0.20,<0.21",
+    "tokenizers": "tokenizers>=0.21,<0.22",
     "torch": "torch",
     "torchaudio": "torchaudio",
     "torchvision": "torchvision",

src/transformers/generation/utils.py

@@ -45,7 +45,6 @@
     is_accelerate_available,
     is_hqq_available,
     is_optimum_quanto_available,
-    is_quanto_available,
     is_torchdynamo_compiling,
     logging,
 )
@@ -1787,7 +1786,7 @@ def _prepare_cache_for_generation(
                 )
                 cache_class = QUANT_BACKEND_CLASSES_MAPPING[cache_config.backend]
 
-                if cache_config.backend == "quanto" and not (is_optimum_quanto_available() or is_quanto_available()):
+                if cache_config.backend == "quanto" and not is_optimum_quanto_available():
                     raise ImportError(
                         "You need to install optimum-quanto in order to use KV cache quantization with optimum-quanto backend. "
                         "Please install it via  with `pip install optimum-quanto`"

src/transformers/integrations/quanto.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ..utils import is_optimum_quanto_available, is_quanto_available, is_torch_available, logging
+from ..utils import is_optimum_quanto_available, is_torch_available, logging
 
 
 if is_torch_available():
@@ -50,11 +50,6 @@ def replace_with_quanto_layers(
 
     if is_optimum_quanto_available():
         from optimum.quanto import QLayerNorm, QLinear, qfloat8, qint2, qint4, qint8
-    elif is_quanto_available():
-        logger.warning_once(
-            "Importing from quanto will be deprecated in v4.47. Please install optimum-quanto instead `pip install optimum-quanto`"
-        )
-        from quanto import QLayerNorm, QLinear, qfloat8, qint2, qint4, qint8
 
     w_mapping = {"float8": qfloat8, "int8": qint8, "int4": qint4, "int2": qint2}
     a_mapping = {None: None, "float8": qfloat8, "int8": qint8}

src/transformers/modeling_gguf_pytorch_utils.py

@@ -291,7 +291,6 @@ def load_gguf_checkpoint(gguf_checkpoint_path, return_tensors=False):
     # FIXME: Currnetly this implementation is only for flan-t5 architecture.
     # It needs to be developed for supporting legacy t5.
     elif "t5" in architecture or "t5encoder" in architecture:
-        parsed_parameters["config"]["tie_word_embeddings"] = False
         parsed_parameters["config"]["is_gated_act"] = True
         updated_architecture = "t5"
     else:
@@ -326,6 +325,12 @@ def load_gguf_checkpoint(gguf_checkpoint_path, return_tensors=False):
     if architecture + model_size not in GGUF_SUPPORTED_ARCHITECTURES:
         raise ValueError(f"Architecture {architecture + model_size} not supported")
 
+    # Handle tie_word_embeddings, if lm_head.weight is not present in tensors,
+    # tie_word_embeddings is true otherwise false
+    parsed_parameters["config"]["tie_word_embeddings"] = all(
+        "output.weight" != tensor.name for tensor in reader.tensors
+    )
+
     # List all key-value pairs in a columnized format
     for gguf_key, field in reader.fields.items():
         gguf_key = gguf_key.replace(architecture, updated_architecture)

src/transformers/models/__init__.py

@@ -118,6 +118,7 @@
     idefics,
     idefics2,
     idefics3,
+    ijepa,
     imagegpt,
     informer,
     instructblip,

src/transformers/models/auto/configuration_auto.py

@@ -136,6 +136,7 @@
         ("idefics", "IdeficsConfig"),
         ("idefics2", "Idefics2Config"),
         ("idefics3", "Idefics3Config"),
+        ("ijepa", "IJepaConfig"),
         ("imagegpt", "ImageGPTConfig"),
         ("informer", "InformerConfig"),
         ("instructblip", "InstructBlipConfig"),
@@ -442,6 +443,7 @@
         ("idefics", "IDEFICS"),
         ("idefics2", "Idefics2"),
         ("idefics3", "Idefics3"),
+        ("ijepa", "I-JEPA"),
         ("imagegpt", "ImageGPT"),
         ("informer", "Informer"),
         ("instructblip", "InstructBLIP"),

src/transformers/models/auto/image_processing_auto.py

@@ -90,6 +90,7 @@
             ("idefics", ("IdeficsImageProcessor",)),
             ("idefics2", ("Idefics2ImageProcessor",)),
             ("idefics3", ("Idefics3ImageProcessor",)),
+            ("ijepa", ("ViTImageProcessor", "ViTImageProcessorFast")),
             ("imagegpt", ("ImageGPTImageProcessor",)),
             ("instructblip", ("BlipImageProcessor",)),
             ("instructblipvideo", ("InstructBlipVideoImageProcessor",)),
@@ -433,7 +434,9 @@ def from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs):
         if image_processor_class is None and image_processor_auto_map is None:
             if not isinstance(config, PretrainedConfig):
                 config = AutoConfig.from_pretrained(
-                    pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
+                    pretrained_model_name_or_path,
+                    trust_remote_code=trust_remote_code,
+                    **kwargs,
                 )
             # It could be in `config.image_processor_type``
             image_processor_class = getattr(config, "image_processor_type", None)

src/transformers/models/auto/modeling_auto.py

@@ -133,6 +133,7 @@
         ("idefics", "IdeficsModel"),
         ("idefics2", "Idefics2Model"),
         ("idefics3", "Idefics3Model"),
+        ("ijepa", "IJepaModel"),
         ("imagegpt", "ImageGPTModel"),
         ("informer", "InformerModel"),
         ("jamba", "JambaModel"),
@@ -580,6 +581,7 @@
         ("focalnet", "FocalNetModel"),
         ("glpn", "GLPNModel"),
         ("hiera", "HieraModel"),
+        ("ijepa", "IJepaModel"),
         ("imagegpt", "ImageGPTModel"),
         ("levit", "LevitModel"),
         ("mllama", "MllamaVisionModel"),
@@ -658,6 +660,7 @@
         ("efficientnet", "EfficientNetForImageClassification"),
         ("focalnet", "FocalNetForImageClassification"),
         ("hiera", "HieraForImageClassification"),
+        ("ijepa", "IJepaForImageClassification"),
         ("imagegpt", "ImageGPTForImageClassification"),
         (
             "levit",

src/transformers/models/blip_2/modeling_blip_2.py

@@ -2311,7 +2311,7 @@ def generate(
         if input_ids is None:
             start_tokens = [self.config.text_config.bos_token_id]
             if getattr(self.config, "image_token_index", None) is not None:
-                start_tokens += [self.config.image_token_index] * self.config.num_query_tokens
+                start_tokens = [self.config.image_token_index] * self.config.num_query_tokens + start_tokens
             input_ids = torch.tensor([start_tokens], dtype=torch.long, device=image_embeds.device)
             input_ids = input_ids.repeat(batch_size, 1)