Upload 23 files

.gitattributes

@@ -1,35 +1,36 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.json filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,73 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+
+# Virtual Environment
+venv/
+env/
+ENV/
+.env
+.venv
+env.bak/
+venv.bak/
+
+# IDE
+.idea/
+.vscode/
+*.swp
+*.swo
+.project
+.pydevproject
+.settings/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+*.ipynb
+
+# Distribution / packaging
+.Python
+*.manifest
+*.spec
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+.hypothesis/
+
+# Logs and databases
+*.log
+*.sqlite
+*.db
+
+# OS generated files
+.DS_Store
+.DS_Store?
+._*
+.Spotlight-V100
+.Trashes
+ehthumbs.db
+Thumbs.db

LICENSE

@@ -0,0 +1,52 @@
+Qwen RESEARCH LICENSE AGREEMENT
+
+Qwen RESEARCH LICENSE AGREEMENT Release Date: September 19, 2024
+
+By clicking to agree or by using or distributing any portion or element of the Qwen Materials, you will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately.
+
+1. Definitions
+    a. This Qwen RESEARCH LICENSE AGREEMENT (this "Agreement") shall mean the terms and conditions for use, reproduction, distribution and modification of the Materials as defined by this Agreement.
+    b. "We" (or "Us") shall mean Alibaba Cloud.
+    c. "You" (or "Your") shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Materials for any purpose and in any field of use.
+    d. "Third Parties" shall mean individuals or legal entities that are not under common control with us or you.
+    e. "Qwen" shall mean the large language models, and software and algorithms, consisting of trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing distributed by us.
+    f. "Materials" shall mean, collectively, Alibaba Cloud's proprietary Qwen and Documentation (and any portion thereof) made available under this Agreement.
+    g. "Source" form shall mean the preferred form for making modifications, including but not limited to model source code, documentation source, and configuration files.
+    h. "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types.
+    i. "Non-Commercial" shall mean for research or evaluation purposes only.
+
+2. Grant of Rights
+    a. You are granted a non-exclusive, worldwide, non-transferable and royalty-free limited license under Alibaba Cloud's intellectual property or other rights owned by us embodied in the Materials to use, reproduce, distribute, copy, create derivative works of, and make modifications to the Materials FOR NON-COMMERCIAL PURPOSES ONLY. 
+    b. If you are commercially using the Materials, you shall request a license from us.
+
+3. Redistribution
+You may distribute copies or make the Materials, or derivative works thereof, available as part of a product or service that contains any of them, with or without modifications, and in Source or Object form, provided that you meet the following conditions:
+    a. You shall give any other recipients of the Materials or derivative works a copy of this Agreement;
+    b. You shall cause any modified files to carry prominent notices stating that you changed the files;
+    c. You shall retain in all copies of the Materials that you distribute the following attribution notices within a "Notice" text file distributed as a part of such copies: "Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) Alibaba Cloud. All Rights Reserved."; and
+    d. You may add your own copyright statement to your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of your modifications, or for any such derivative works as a whole, provided your use, reproduction, and distribution of the work otherwise complies with the terms and conditions of this Agreement.
+
+4. Rules of use
+    a. The Materials may be subject to export controls or restrictions in China, the United States or other countries or regions. You shall comply with applicable laws and regulations in your use of the Materials.
+    b. If you use the Materials or any outputs or results therefrom to create, train, fine-tune, or improve an AI model that is distributed or made available, you shall prominently display “Built with Qwen” or “Improved using Qwen” in the related product documentation.
+
+5. Intellectual Property
+    a. We retain ownership of all intellectual property rights in and to the Materials and derivatives made by or for us. Conditioned upon compliance with the terms and conditions of this Agreement, with respect to any derivative works and modifications of the Materials that are made by you, you are and will be the owner of such derivative works and modifications.
+    b. No trademark license is granted to use the trade names, trademarks, service marks, or product names of us, except as required to fulfill notice requirements under this Agreement or as required for reasonable and customary use in describing and redistributing the Materials.
+    c. If you commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against us or any entity alleging that the Materials or any output therefrom, or any part of the foregoing, infringe any intellectual property or other right owned or licensable by you, then all licenses granted to you under this Agreement shall terminate as of the date such lawsuit or other proceeding is commenced or brought.
+6. Disclaimer of Warranty and Limitation of Liability
+    a. We are not obligated to support, update, provide training for, or develop any further version of the Qwen Materials or to grant any license thereto.
+    b. THE MATERIALS ARE PROVIDED "AS IS" WITHOUT ANY EXPRESS OR IMPLIED WARRANTY OF ANY KIND INCLUDING WARRANTIES OF MERCHANTABILITY, NONINFRINGEMENT, OR FITNESS FOR A PARTICULAR PURPOSE. WE MAKE NO WARRANTY AND ASSUME NO RESPONSIBILITY FOR THE SAFETY OR STABILITY OF THE MATERIALS AND ANY OUTPUT THEREFROM.
+    c. IN NO EVENT SHALL WE BE LIABLE TO YOU FOR ANY DAMAGES, INCLUDING, BUT NOT LIMITED TO ANY DIRECT, OR INDIRECT, SPECIAL OR CONSEQUENTIAL DAMAGES ARISING FROM YOUR USE OR INABILITY TO USE THE MATERIALS OR ANY OUTPUT OF IT, NO MATTER HOW IT’S CAUSED.
+    d. You will defend, indemnify and hold harmless us from and against any claim by any third party arising out of or related to your use or distribution of the Materials.
+
+7. Survival and Termination.
+    a. The term of this Agreement shall commence upon your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein.
+    b. We may terminate this Agreement if you breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, you must delete and cease use of the Materials. Sections 6 and 8 shall survive the termination of this Agreement.
+
+8. Governing Law and Jurisdiction.
+    a. This Agreement and any dispute arising out of or relating to it will be governed by the laws of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement.
+    b. The People's Courts in Hangzhou City shall have exclusive jurisdiction over any dispute arising out of this Agreement.
+9. Other Terms and Conditions.
+    a. Any arrangements, understandings, or agreements regarding the Material not stated herein are separate from and independent of the terms and conditions of this Agreement. You shall request a separate license from us, if you use the Materials in ways not expressly agreed to in this Agreement. 
+    b. We shall not be bound by any additional or different terms or conditions communicated by you unless expressly agreed.

README.md

@@ -0,0 +1,365 @@
+---
+tags:
+- vidore
+- colpali
+- multimodal-embedding
+- multilingual-embedding
+- Text-to-Visual Document (T→VD) retrieval
+- feature-extraction
+- sentence-similarity
+- mteb
+- sentence-transformers
+language:
+- multilingual
+inference: false
+library_name: transformers
+pipeline_tag: visual-document-retrieval
+---
+<br><br>
+
+<p align="center">
+<img src="https://huggingface.co/datasets/jinaai/documentation-images/resolve/main/logo.webp" alt="Jina AI: Your Search Foundation, Supercharged!" width="150px">
+</p>
+
+
+<p align="center">
+<b>The embedding model trained by <a href="https://jina.ai/"><b>Jina AI</b></a>.</b>
+</p>
+
+# Jina Embeddings v4: Universal Embeddings for Multimodal Multilingual Retrieval
+
+
+[GGUF](https://github.com/jina-ai/jina-embeddings-v4-gguf) | [Blog](https://jina.ai/news/jina-embeddings-v4-universal-embeddings-for-multimodal-multilingual-retrieval) | [Technical Report](https://arxiv.org/abs/2506.18902) | [API](https://jina.ai/embeddings)
+
+
+## Intended Usage & Model Info
+`jina-embeddings-v4` is a universal embedding model for multimodal and multilingual retrieval. 
+The model is specially designed for complex document retrieval, including visually rich documents with charts, tables, and illustrations. 
+
+
+Built on [Qwen/Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct), `jina-embeddings-v4` features:
+
+- **Unified embeddings** for text, images, and visual documents, supporting both dense (single-vector) and late-interaction (multi-vector) retrieval.
+- **Multilingual support** (30+ languages) and compatibility with a wide range of domains, including technical and visually complex documents.
+- **Task-specific adapters** for retrieval, text matching, and code-related tasks, which can be selected at inference time.
+- **Flexible embedding size**: dense embeddings are 2048 dimensions by default but can be truncated to as low as 128 with minimal performance loss.
+
+
+Summary of features:
+
+| Feature   | Jina Embeddings V4   |
+|------------|------------|
+| Base Model | Qwen2.5-VL-3B-Instruct |
+| Supported Tasks | `retrieval`, `text-matching`, `code` |
+| Model DType | BFloat 16 |
+| Max Sequence Length | 32768 |
+| Single-Vector Dimension | 2048 |
+| Multi-Vector Dimension | 128 |
+| Matryoshka dimensions | 128, 256, 512, 1024, 2048 |
+| Pooling Strategy | Mean pooling |
+| Attention Mechanism | FlashAttention2 |
+
+
+
+## Training & Evaluation
+
+Please refer to our [technical report of jina-embeddings-v4](https://arxiv.org/abs/2506.18902) for training details and benchmarks.
+
+
+## Usage
+
+<details>
+  <summary>Requirements</a></summary>
+  
+The following Python packages are required:
+
+- `transformers>=4.52.0`
+- `torch>=2.6.0`
+- `peft>=0.15.2`
+- `torchvision`
+- `pillow`
+  
+### Optional / Recommended
+- **flash-attention**: Installing [flash-attention](https://github.com/Dao-AILab/flash-attention) is recommended for improved inference speed and efficiency, but not mandatory.
+- **sentence-transformers**: If you want to use the model via the `sentence-transformers` interface, install this package as well.
+
+</details>
+
+
+<details>
+  <summary>via <a href="https://jina.ai/embeddings/">Jina AI Embeddings API</a></summary>
+
+
+```bash
+curl https://api.jina.ai/v1/embeddings \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer $JINA_AI_API_TOKEN" \
+  -d @- <<EOFEOF
+  {
+    "model": "jina-embeddings-v4",
+    "task": "text-matching",
+    "input": [
+        {
+            "text": "غروب جميل على الشاطئ"
+        },
+        {
+            "text": "海滩上美丽的日落"
+        },
+        {
+            "text": "A beautiful sunset over the beach"
+        },
+        {
+            "text": "Un beau coucher de soleil sur la plage"
+        },
+        {
+            "text": "Ein wunderschöner Sonnenuntergang am Strand"
+        },
+        {
+            "text": "Ένα όμορφο ηλιοβασίλεμα πάνω από την παραλία"
+        },
+        {
+            "text": "समुद्र तट पर एक खूबसूरत सूर्यास्त"
+        },
+        {
+            "text": "Un bellissimo tramonto sulla spiaggia"
+        },
+        {
+            "text": "浜辺に沈む美しい夕日"
+        },
+        {
+            "text": "해변 위로 아름다운 일몰"
+        },
+        {
+            "image": "https://i.ibb.co/nQNGqL0/beach1.jpg"
+        },
+        {
+            "image": "https://i.ibb.co/r5w8hG8/beach2.jpg"
+        }
+    ]
+  }
+EOFEOF
+```
+
+</details>
+
+<details>
+  <summary>via <a href="https://huggingface.co/docs/transformers/en/index">transformers</a></summary>
+
+```python
+# !pip install transformers>=4.52.0 torch>=2.6.0 peft>=0.15.2 torchvision pillow
+# !pip install
+from transformers import AutoModel
+import torch
+
+# Initialize the model
+model = AutoModel.from_pretrained("jinaai/jina-embeddings-v4", trust_remote_code=True, torch_dtype=torch.float16)
+
+model.to("cuda")
+
+# ========================
+# 1. Retrieval Task
+# ========================
+# Configure truncate_dim, max_length (for texts), max_pixels (for images), vector_type, batch_size in the encode function if needed
+
+# Encode query
+query_embeddings = model.encode_text(
+    texts=["Overview of climate change impacts on coastal cities"],
+    task="retrieval",
+    prompt_name="query",
+)
+
+# Encode passage (text)
+passage_embeddings = model.encode_text(
+    texts=[
+        "Climate change has led to rising sea levels, increased frequency of extreme weather events..."
+    ],
+    task="retrieval",
+    prompt_name="passage",
+)
+
+# Encode image/document
+image_embeddings = model.encode_image(
+    images=["https://i.ibb.co/nQNGqL0/beach1.jpg"],
+    task="retrieval",
+)
+
+# ========================
+# 2. Text Matching Task
+# ========================
+texts = [
+    "غروب جميل على الشاطئ",  # Arabic
+    "海滩上美丽的日落",  # Chinese
+    "Un beau coucher de soleil sur la plage",  # French
+    "Ein wunderschöner Sonnenuntergang am Strand",  # German
+    "Ένα όμορφο ηλιοβασίλεμα πάνω από την παραλία",  # Greek
+    "समुद्र तट पर एक खूबसूरत सूर्यास्त",  # Hindi
+    "Un bellissimo tramonto sulla spiaggia",  # Italian
+    "浜辺に沈む美しい夕日",  # Japanese
+    "해변 위로 아름다운 일몰",  # Korean
+]
+
+text_embeddings = model.encode_text(texts=texts, task="text-matching")
+
+# ========================
+# 3. Code Understanding Task
+# ========================
+
+# Encode query
+query_embedding = model.encode_text(
+    texts=["Find a function that prints a greeting message to the console"],
+    task="code",
+    prompt_name="query",
+)
+
+# Encode code
+code_embeddings = model.encode_text(
+    texts=["def hello_world():\n    print('Hello, World!')"],
+    task="code",
+    prompt_name="passage",
+)
+
+# ========================
+# 4. Use multivectors
+# ========================
+
+multivector_embeddings = model.encode_text(
+    texts=texts,
+    task="retrieval",
+    prompt_name="query",
+    return_multivector=True,
+)
+
+images = ["https://i.ibb.co/nQNGqL0/beach1.jpg", "https://i.ibb.co/r5w8hG8/beach2.jpg"]
+multivector_image_embeddings = model.encode_image(
+    images=images,
+    task="retrieval",
+    return_multivector=True,
+)
+```
+</details>
+
+<details>
+  <summary>via <a href="https://sbert.net/">sentence-transformers</a></summary>
+  
+```python
+from sentence_transformers import SentenceTransformer
+
+# Initialize the model
+model = SentenceTransformer("jinaai/jina-embeddings-v4", trust_remote_code=True)
+# ========================
+# 1. Retrieval Task
+# ========================
+# Encode query
+query_embeddings = model.encode(
+    sentences=["Overview of climate change impacts on coastal cities"],
+    task="retrieval",
+    prompt_name="query",
+)
+
+print(f"query_embeddings.shape = {query_embeddings.shape}")
+
+# Encode passage (text)
+passage_embeddings = model.encode(
+    sentences=[
+        "Climate change has led to rising sea levels, increased frequency of extreme weather events..."
+    ],
+    task="retrieval",
+    prompt_name="passage",
+)
+
+print(f"passage_embeddings.shape = {passage_embeddings.shape}")
+
+# Encode image/document
+image_embeddings = model.encode(
+    sentences=["https://i.ibb.co/nQNGqL0/beach1.jpg"],
+    task="retrieval",
+)
+
+print(f"image_embeddings.shape = {image_embeddings.shape}")
+
+# ========================
+# 2. Text Matching Task
+# ========================
+texts = [
+    "غروب جميل على الشاطئ",  # Arabic
+    "海滩上美丽的日落",  # Chinese
+    "Un beau coucher de soleil sur la plage",  # French
+    "Ein wunderschöner Sonnenuntergang am Strand",  # German
+    "Ένα όμορφο ηλιοβασίλεμα πάνω από την παραλία",  # Greek
+    "समुद्र तट पर एक खूबसूरत सूर्यास्त",  # Hindi
+    "Un bellissimo tramonto sulla spiaggia",  # Italian
+    "浜辺に沈む美しい夕日",  # Japanese
+    "해변 위로 아름다운 일몰",  # Korean
+]
+
+text_embeddings = model.encode(sentences=texts, task="text-matching")
+
+# ========================
+# 3. Code Understanding Task
+# ========================
+
+# Encode query
+query_embeddings = model.encode(
+    sentences=["Find a function that prints a greeting message to the console"],
+    task="code",
+    prompt_name="query",
+)
+
+# Encode code
+code_embeddings = model.encode(
+    sentences=["def hello_world():\n    print('Hello, World!')"],
+    task="code",
+    prompt_name="passage",
+)
+
+# ========================
+# 4. Use multivectors
+# ========================
+# If you want to use multi-vector embeddings, please use the Hugging Face model directly.
+```
+</details>
+
+<details>
+  <summary>via <a href="https://github.com/vllm-project/vllm">vLLM</a></summary>
+
+We provide separate model versions for each task (`retrieval`, `text-matching`, `code`) where specific adapter is merged into the base `Qwen2.5-VL` weights. 
+This modification enables native compatibility with vLLM.
+
+Instructions and usage examples for each task are available in their respective directories:
+- [jina-embeddings-v4-vllm-retrieval](https://huggingface.co/jinaai/jina-embeddings-v4-vllm-retrieval)
+- [jina-embeddings-v4-vllm-text-matching](https://huggingface.co/jinaai/jina-embeddings-v4-vllm-text-matching)
+- [jina-embeddings-v4-vllm-code](https://huggingface.co/jinaai/jina-embeddings-v4-vllm-code)
+
+Please refer to the directory that matches your task for more details.
+
+</details>
+
+
+## Jina-VDR
+Alongside `jina-embeddings-v4`, we’re releasing [Jina VDR](https://github.com/jina-ai/jina-vdr), a multilingual, multi-domain benchmark for visual document retrieval. The task collection can be viewed [here](https://huggingface.co/collections/jinaai/jinavdr-visual-document-retrieval-684831c022c53b21c313b449), and evaluation instructions can be found [here](https://github.com/jina-ai/jina-vdr).
+
+
+## License
+
+This model was initially released under cc-by-nc-4.0 due to an error.
+The correct license is the Qwen Research License, as this model is derived from Qwen-2.5-VL-3B which is governed by that license.
+
+## Contact
+
+Join our [Discord community](https://discord.jina.ai) and chat with other community members about ideas.
+
+
+## Citation
+
+If you find `jina-embeddings-v4` useful in your research, please cite the following paper:
+```
+@misc{günther2025jinaembeddingsv4universalembeddingsmultimodal,
+      title={jina-embeddings-v4: Universal Embeddings for Multimodal Multilingual Retrieval}, 
+      author={Michael Günther and Saba Sturua and Mohammad Kalim Akram and Isabelle Mohr and Andrei Ungureanu and Sedigheh Eslami and Scott Martens and Bo Wang and Nan Wang and Han Xiao},
+      year={2025},
+      eprint={2506.18902},
+      archivePrefix={arXiv},
+      primaryClass={cs.AI},
+      url={https://arxiv.org/abs/2506.18902}, 
+}
+```

added_tokens.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58b54bbe36fc752f79a24a271ef66a0a0830054b4dfad94bde757d851968060b
+size 605

chat_template.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94174d7176c52a7192f96fc34eb2cf23c7c2059d63cdbfadca1586ba89731fb7
+size 1049

config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bd26fbef1d930d8f968729480fd8ce1dbdd468d5f32079ef3eb0e3e2c59ba24
+size 2751

config_sentence_transformers.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1eee316c1ced66356d6472a0f3e2ff28084e8a693cbb2bb758ed98cc3f20ba22
+size 274

configuration_jina_embeddings_v4.py

@@ -0,0 +1,23 @@
+from transformers.models.qwen2_5_vl import Qwen2_5_VLConfig
+
+from typing import Optional
+
+
+class JinaEmbeddingsV4Config(Qwen2_5_VLConfig):
+    """
+    Configuration for the JinaEmbeddingsV4 model.
+    """
+
+    def __init__(
+        self,
+        single_vector_pool_strategy: str = "mean",
+        multi_vector_projector_dim: int = 128,
+        pretrained_peft_model_name_or_path: Optional[str] = None,
+        verbosity: int = 1,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.single_vector_pool_strategy = single_vector_pool_strategy
+        self.multi_vector_projector_dim = multi_vector_projector_dim
+        self.pretrained_peft_model_name_or_path = pretrained_peft_model_name_or_path
+        self.verbosity = verbosity

custom_lora_module.py

@@ -0,0 +1,193 @@
+from __future__ import annotations
+
+import math
+import warnings
+from typing import Any, Optional, Union, List
+
+import torch
+import torch.nn as nn
+
+from peft.tuners.lora import LoraLayer
+
+class MultiAdapterLinear(nn.Module, LoraLayer):
+    """
+    Custom LoRA module supporting multiple adapters for a linear layer.
+    
+    This module extends the standard LoRA implementation to support multiple task-specific
+    adapters that can be dynamically selected during the forward pass. The task_label
+    parameter passed to the forward function determines which LoRA adapter(s) to use:
+    - If task_label is a string, all examples in the batch use the same adapter
+    - If task_label is a list of strings, each example can use a different adapter
+    
+    This enables efficient multi-task inference where all task-specific LoRA adapters
+    are loaded in memory simultaneously and dynamically selected per example, eliminating
+    the need to switch adapter states between tasks and allowing optimal throughput
+    for mixed-task batches.
+    
+    Derived from peft.tuners.lora.Linear.
+    """
+    def __init__(
+        self,
+        base_layer,
+        adapter_name: str,
+        task_names: List[str],
+        r: int = 0,
+        lora_alpha: int = 1,
+        lora_dropout: float = 0.0,
+        fan_in_fan_out: bool = False,  # Set this to True if the layer to replace stores weight like (fan_in, fan_out)
+        is_target_conv_1d_layer: bool = False,
+        init_lora_weights: Union[bool, str] = True,
+        use_rslora: bool = False,
+        use_dora: bool = False,
+        lora_bias: bool = False,
+        **kwargs,
+    ) -> None:
+        super().__init__()
+        LoraLayer.__init__(self, base_layer, **kwargs)
+
+        self.fan_in_fan_out = fan_in_fan_out
+        self.task_names = task_names
+        self._active_adapter = adapter_name
+        self.update_layer(
+            adapter_name,
+            r,
+            lora_alpha=lora_alpha,
+            lora_dropout=lora_dropout,
+            init_lora_weights=init_lora_weights,
+            use_rslora=use_rslora,
+            use_dora=use_dora,
+            lora_bias=lora_bias,
+        )
+        self.is_target_conv_1d_layer = is_target_conv_1d_layer
+
+
+    def forward(self, x: torch.Tensor, task_label: Union[str, List[str]], *args: Any, **kwargs: Any) -> torch.Tensor:
+        self._check_forward_args(x, *args, **kwargs)
+
+        if self.disable_adapters:
+            if self.merged:
+                self.unmerge()
+            result = self.base_layer(x, *args, **kwargs)
+        elif self.merged:
+            result = self.base_layer(x, *args, **kwargs)
+        else:
+            result = self.base_layer(x, *args, **kwargs)
+            torch_result_dtype = result.dtype
+
+            lora_A_keys = self.lora_A.keys()
+            for active_adapter in self.active_adapters:
+                if active_adapter not in lora_A_keys:
+                    continue
+                
+                if isinstance(task_label, str):
+                    lora_A = self.lora_A[active_adapter][task_label]
+                    lora_B = self.lora_B[active_adapter][task_label]
+                    dropout = self.lora_dropout[active_adapter]
+                    scaling = self.scaling[active_adapter]
+                    x = self._cast_input_dtype(x, lora_A.weight.dtype)
+                    result = result + lora_B(lora_A(dropout(x))) * scaling
+                else:
+                    unique_tasks = list(set(task_label))
+                    lora_output = torch.zeros_like(result)
+                    
+                    for task in unique_tasks:
+                        task_indices = [i for i, t in enumerate(task_label) if t == task]
+                        task_x = x[task_indices]
+                        
+                        lora_A = self.lora_A[active_adapter][task]
+                        lora_B = self.lora_B[active_adapter][task]
+                        dropout = self.lora_dropout[active_adapter]
+                        scaling = self.scaling[active_adapter]
+                        
+                        task_x = self._cast_input_dtype(task_x, lora_A.weight.dtype)
+                        task_lora_value = lora_B(lora_A(dropout(task_x))) * scaling
+                        
+                        for i, idx in enumerate(task_indices):
+                            lora_output[idx] = task_lora_value[i]
+                    
+                    result = result + lora_output
+
+            result = result.to(torch_result_dtype)
+
+        return result
+
+    def __repr__(self) -> str:
+        rep = super().__repr__()
+        return "lora." + rep
+
+
+    def update_layer(
+        self,
+        adapter_name,
+        r,
+        lora_alpha,
+        lora_dropout,
+        init_lora_weights,
+        use_rslora,
+        use_dora: bool = False,
+        lora_bias: bool = False,
+    ):
+        # This code works for linear layers, override for other layer types
+        if r <= 0:
+            raise ValueError(f"`r` should be a positive integer value but the value passed is {r}")
+
+        self.r[adapter_name] = r
+        self.lora_alpha[adapter_name] = lora_alpha
+        if lora_dropout > 0.0:
+            lora_dropout_layer = nn.Dropout(p=lora_dropout)
+        else:
+            lora_dropout_layer = nn.Identity()
+
+        self.lora_dropout.update(nn.ModuleDict({adapter_name: lora_dropout_layer}))
+        # Actual trainable parameters
+        self.lora_A[adapter_name] = nn.ModuleDict({
+            task_name: nn.Linear(self.in_features, r, bias=False)
+            for task_name in self.task_names
+        })
+        self.lora_B[adapter_name] = nn.ModuleDict({
+            task_name: nn.Linear(r, self.out_features, bias=lora_bias)
+            for task_name in self.task_names
+        })
+        self.lora_bias[adapter_name] = lora_bias
+
+        if use_rslora:
+            self.scaling[adapter_name] = lora_alpha / math.sqrt(r)
+        else:
+            self.scaling[adapter_name] = lora_alpha / r
+
+        self.reset_lora_parameters(adapter_name, init_lora_weights)
+        self._move_adapter_to_device_of_base_layer(adapter_name)
+        self.use_dora[adapter_name] = False
+        self.set_adapter(self.active_adapters)
+
+    def reset_lora_parameters(self, adapter_name, init_lora_weights):
+        if init_lora_weights is False:
+            return
+        if init_lora_weights is True:
+            # initialize A the same way as the default for nn.Linear and B to zero
+            # https://github.com/microsoft/LoRA/blob/a0a92e0f26c067cf94747bdbf1ce73793fa44d19/loralib/layers.py#L124
+            for task_name in self.task_names:
+                nn.init.kaiming_uniform_(self.lora_A[adapter_name][task_name].weight, a=math.sqrt(5))
+        elif init_lora_weights.lower() == "gaussian":
+            for task_name in self.task_names:
+                nn.init.normal_(self.lora_A[adapter_name][task_name].weight, std=1 / self.r[adapter_name])
+        else:
+            raise ValueError(f"Unknown initialization {init_lora_weights=}")
+        for task_name in self.task_names:
+            nn.init.zeros_(self.lora_B[adapter_name][task_name].weight)
+        if self.lora_bias[adapter_name]:
+            for task_name in self.task_names:
+                nn.init.zeros_(self.lora_B[adapter_name][task_name].bias)
+    
+
+    def merge(self, safe_merge: bool = False, adapter_names: Optional[list[str]] = None) -> None:
+        """
+        Merge the active adapter weights into the base weights
+        """
+        raise NotImplementedError("Merge operation is not supported")
+
+    def unmerge(self) -> None:
+        """
+        This method unmerges all merged adapter layers from the base weights.
+        """
+        raise NotImplementedError("Unmerge operation is not supported")

custom_st.py

@@ -0,0 +1,185 @@
+import json
+import os
+from io import BytesIO
+from pathlib import Path
+from typing import Any, Dict, List, Literal, Optional, Union
+
+import requests
+import torch
+from PIL import Image
+from torch import nn
+from transformers import AutoConfig, AutoModel, AutoProcessor
+
+
+class Transformer(nn.Module):
+
+    save_in_root: bool = True
+
+    def __init__(
+        self,
+        model_name_or_path: str = "jinaai/jina-embeddings-v4",
+        max_seq_length: Optional[int] = None,
+        config_args: Optional[Dict[str, Any]] = None,
+        model_args: Optional[Dict[str, Any]] = None,
+        tokenizer_args: Optional[Dict[str, Any]] = None,
+        cache_dir: Optional[str] = None,
+        backend: Literal["torch", "onnx", "openvino"] = "torch",
+        **kwargs,
+    ) -> None:
+        super(Transformer, self).__init__()
+        if backend != "torch":
+            raise ValueError(
+                f"Backend '{backend}' is not supported, please use 'torch' instead"
+            )
+        config_kwargs = config_args or {}
+        model_kwargs = model_args or {}
+        tokenizer_kwargs = tokenizer_args or {}
+
+        self.config = AutoConfig.from_pretrained(
+            model_name_or_path, cache_dir=cache_dir, **config_kwargs
+        )
+        self.default_task = model_args.pop("default_task", None)
+        if self.default_task and self.default_task not in self.config.task_names:
+            raise ValueError(
+                f"Invalid task: {self.default_task}. Must be one of {self.config.task_names}."
+            )
+
+        self.model = AutoModel.from_pretrained(
+            model_name_or_path, config=self.config, cache_dir=cache_dir, **model_kwargs
+        )
+        self.processor = AutoProcessor.from_pretrained(
+            model_name_or_path,
+            cache_dir=cache_dir,
+            use_fast=True,
+            **tokenizer_kwargs,
+        )
+        self.max_seq_length = max_seq_length or 8192
+
+    def tokenize(
+        self, texts: List[Union[str, Image.Image]], padding: Union[str, bool] = True
+    ) -> Dict[str, torch.Tensor]:
+        encoding = {}
+        text_indices = []
+        image_indices = []
+        for i, text in enumerate(texts):
+            if isinstance(text, str):
+                # Remove Query: or Passage: prefixes when checking for URLs or file paths
+                clean_text = text
+                if text.startswith("Query: "):
+                    clean_text = text[len("Query: ") :]
+                elif text.startswith("Passage: "):
+                    clean_text = text[len("Passage: ") :]
+
+                if clean_text.startswith("http"):
+                    response = requests.get(clean_text)
+                    texts[i] = Image.open(BytesIO(response.content)).convert("RGB")
+                    image_indices.append(i)
+                else:
+                    try:
+                        if Path(clean_text).is_file():
+                            texts[i] = Image.open(clean_text).convert("RGB")
+                            image_indices.append(i)
+                        else:
+                            text_indices.append(i)
+                    except Exception as e:
+                        text_indices.append(i)
+            elif isinstance(text, Image.Image):
+                image_indices.append(i)
+            else:
+                raise ValueError(f"Invalid input type: {type(text)}")
+        if text_indices:
+            _texts = [texts[i] for i in text_indices]
+            text_features = self.processor.process_texts(
+                _texts, max_length=self.max_seq_length
+            )
+            for key, value in text_features.items():
+                encoding[f"text_{key}"] = value
+            encoding["text_indices"] = text_indices
+
+        if image_indices:
+            _images = [texts[i] for i in image_indices]
+            img_features = self.processor.process_images(_images)
+            for key, value in img_features.items():
+                encoding[f"image_{key}"] = value
+            encoding["image_indices"] = image_indices
+
+        return encoding
+
+    def forward(
+        self,
+        features: Dict[str, torch.Tensor],
+        task: Optional[str] = None,
+        truncate_dim: Optional[int] = None,
+    ) -> Dict[str, torch.Tensor]:
+        self.model.eval()
+
+        if task is None:
+            if self.default_task is None:
+                raise ValueError(
+                    "Task must be specified before encoding data. You can set it either during "
+                    "loading the model (e.g., model_kwargs={'default_task': 'retrieval'}) or "
+                    "pass it as an argument to the encode method (e.g., model.encode(texts, task='retrieval'))."
+                )
+            task = self.default_task
+        else:
+            if task not in self.config.task_names:
+                raise ValueError(
+                    f"Invalid task: {task}. Must be one of {self.config.task_names}."
+                )
+
+        device = self.model.device.type
+        all_embeddings = []
+
+        with torch.no_grad():
+            if any(k.startswith("text_") for k in features.keys()):
+                text_batch = {
+                    k[len("text_") :]: v.to(device)
+                    for k, v in features.items()
+                    if k.startswith("text_") and k != "text_indices"
+                }
+                text_indices = features.get("text_indices", [])
+                with torch.autocast(device_type=device, dtype=torch.bfloat16):
+                    text_embeddings = self.model(
+                        **text_batch, task_label=task
+                    ).single_vec_emb
+                    if truncate_dim:
+                        text_embeddings = text_embeddings[:, :truncate_dim]
+                        text_embeddings = torch.nn.functional.normalize(
+                            text_embeddings, p=2, dim=-1
+                        )
+                for i, embedding in enumerate(text_embeddings):
+                    all_embeddings.append((text_indices[i], embedding))
+
+            if any(k.startswith("image_") for k in features.keys()):
+                image_batch = {
+                    k[len("image_") :]: v.to(device)
+                    for k, v in features.items()
+                    if k.startswith("image_") and k != "image_indices"
+                }
+                image_indices = features.get("image_indices", [])
+
+                with torch.autocast(device_type=device, dtype=torch.bfloat16):
+                    img_embeddings = self.model(
+                        **image_batch, task_label=task
+                    ).single_vec_emb
+                    if truncate_dim:
+                        img_embeddings = img_embeddings[:, :truncate_dim]
+                        img_embeddings = torch.nn.functional.normalize(
+                            img_embeddings, p=2, dim=-1
+                        )
+
+                for i, embedding in enumerate(img_embeddings):
+                    all_embeddings.append((image_indices[i], embedding))
+
+        if not all_embeddings:
+            raise RuntimeError("No embeddings were generated")
+
+        all_embeddings.sort(key=lambda x: x[0])  # sort by original index
+        combined_embeddings = torch.stack([emb for _, emb in all_embeddings])
+        features["sentence_embedding"] = combined_embeddings
+
+        return features
+
+    @classmethod
+    def load(cls, input_path: str) -> "Transformer":
+        return cls(model_name_or_path=input_path)

generation_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4be8e7b43a811255a415e39b11c9b78a3a267120e20dd198774b1a14dcc5ea86
+size 126

model.safetensors.index.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:360eb251d21193fe4a424dcc2f65ee1e9ca66dcbcb497b8141722f9f6ed56f8a
+size 65592

modeling_jina_embeddings_v4.py

@@ -0,0 +1,609 @@
+# Jina Embeddings V4 Model implementation was inspired by the ColPali codebase:
+# https://github.com/illuin-tech/colpali
+
+import os
+from dataclasses import dataclass
+from enum import Enum
+from functools import partial
+from io import BytesIO
+from typing import Any, Callable, ClassVar, Dict, List, Optional, Union, cast
+
+import numpy as np
+import requests
+import torch
+from huggingface_hub import snapshot_download
+from peft import LoraConfig, PeftModel
+from PIL import Image
+from torch import nn
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+from transformers import BatchFeature
+from transformers.utils import is_flash_attn_2_available
+
+from .configuration_jina_embeddings_v4 import JinaEmbeddingsV4Config
+from .custom_lora_module import MultiAdapterLinear
+from .qwen2_5_vl import Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLProcessor
+
+
+class PromptType(str, Enum):
+    query = "query"
+    passage = "passage"
+
+
+PREFIX_DICT = {"query": "Query", "passage": "Passage"}
+
+
+class JinaEmbeddingsV4Processor(Qwen2_5_VLProcessor):
+    def __init__(self, *args, **kwargs) -> None:
+        Qwen2_5_VLProcessor.__init__(self, *args, **kwargs)
+        self.assistant_prefix_len = 58
+        self.text_max_length = 32768
+
+    def process_images(
+        self,
+        images: Union[List[Image.Image], List[List[Image.Image]]],
+    ) -> BatchFeature:
+
+        if isinstance(images[0], list):
+            images = cast(List[List[Image.Image]], images)
+            text_doc = []
+            for i in range(len(images)):
+                conversation = [
+                    {"role": "user", "content": [{"type": "image"}] * len(images[i])}
+                ]
+                template = self.apply_chat_template(
+                    conversation, add_generation_prompt=False
+                )
+                text_doc.append(template[self.assistant_prefix_len :])
+
+        else:
+            images = cast(List[Image.Image], images)
+            text_doc = [
+                "<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Describe the image.<|im_end|>\n"
+            ] * len(images)
+
+        # The following code is a hack to make sure the scatter in DDP is done correctly when training on multiple GPUs
+        batch_doc = self(text=text_doc, images=images, padding="longest", return_tensors="pt")  # type: ignore
+        # Separate pixel_values for each image
+        offsets = batch_doc["image_grid_thw"][:, 1] * batch_doc["image_grid_thw"][:, 2]
+        # Pad pixel_values to the same length to be able to make it into a tensor
+        pixel_values = torch.split(batch_doc["pixel_values"], offsets.tolist())
+
+        max_length = max([len(pv) for pv in pixel_values])
+
+        pixel_values = [
+            torch.cat(
+                [
+                    pv,
+                    torch.zeros(
+                        (max_length - len(pv), pv.shape[1]),
+                        dtype=pv.dtype,
+                        device=pv.device,
+                    ),
+                ]
+            )
+            for pv in pixel_values
+        ]
+
+        batch_doc["pixel_values"] = torch.stack(pixel_values)
+        return batch_doc
+
+    def process_texts(
+        self,
+        texts: List[str],
+        max_length: Optional[int] = None,
+        prefix: Optional[str] = None,
+        padding: Optional[str] = None,
+    ) -> BatchFeature:
+
+        max_length = (
+            self.text_max_length
+            if max_length is None
+            else min(max_length, self.text_max_length)
+        )
+        padded_texts: List[str] = []
+
+        for text in texts:
+            if prefix:
+                text = f"{prefix}: {text}"
+            padded_texts.append(text)
+
+        text_batch = self(
+            text=padded_texts,
+            return_tensors="pt",
+            padding=padding or "longest",
+            max_length=max_length,
+            truncation=True,
+        )
+
+        return text_batch
+
+
+@dataclass
+class JinaEmbeddingsV4ModelOutput:
+    """
+    Base class for the Hybrid Model outputs.
+    Args:
+        vlm_last_hidden_states (torch.Tensor, optional): Last hidden states of the VLM.
+        single_vec_emb (torch.Tensor, optional): Single-vector embeddings.
+        multi_vec_emb (torch.Tensor, optional): Multi-vector embeddings.
+    """
+
+    vlm_last_hidden_states: Optional[torch.Tensor] = None
+    single_vec_emb: Optional[torch.Tensor] = None
+    multi_vec_emb: Optional[torch.Tensor] = None
+
+
+class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
+    config_class = JinaEmbeddingsV4Config
+    main_input_name: ClassVar[str] = "doc_input_ids"
+
+    def __init__(self, config: JinaEmbeddingsV4Config):
+        Qwen2_5_VLForConditionalGeneration.__init__(self, config)
+        self._init_projection_layer(config)
+        self.post_init()
+        self.processor = JinaEmbeddingsV4Processor.from_pretrained(
+            self.name_or_path, trust_remote_code=True, use_fast=True
+        )
+        self.multi_vector_projector_dim = config.multi_vector_projector_dim
+        self.verbosity = config.verbosity
+        self._task = None
+
+    @property
+    def task(self) -> Optional[str]:
+        """Get the current task set for the model."""
+        return self._task
+
+    @task.setter
+    def task(self, task: str):
+        """
+        Set the task for the model.
+
+        Args:
+            task (str): The task name. Must be one of ['retrieval', 'text-matching', 'code']
+        """
+        if task not in self.config.task_names:
+            raise ValueError(
+                f"Invalid task: {task}. Must be one of {self.config.task_names}."
+            )
+        self._task = task
+
+    def get_last_hidden_states(
+        self,
+        task_label: Union[str, List[str]],
+        input_ids: torch.LongTensor,
+        attention_mask: torch.Tensor,
+        **kwargs,
+    ) -> torch.Tensor:
+        if "pixel_values" in kwargs:
+            offsets = kwargs["image_grid_thw"][:, 1] * kwargs["image_grid_thw"][:, 2]
+            kwargs["pixel_values"] = torch.cat(
+                [pv[:o] for pv, o in zip(kwargs["pixel_values"], offsets)], dim=0
+            )
+        position_ids, rope_deltas = self.model.get_rope_index(
+            input_ids=input_ids,
+            image_grid_thw=kwargs.get("image_grid_thw", None),
+            attention_mask=attention_mask,
+        )
+
+        kwargs["output_hidden_states"] = True
+        outputs = super().forward(
+            task_label=task_label,
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            **kwargs,
+            position_ids=position_ids,
+            rope_deltas=rope_deltas,
+            use_cache=False,
+        )
+
+        hidden_states = outputs.hidden_states
+        if not hidden_states:
+            raise ValueError("Hidden states not found in model output")
+
+        return hidden_states[-1]
+
+    def _init_projection_layer(self, config) -> None:
+        """
+        Initializes projection layers.
+        """
+        self.config.multi_vector_projector_dim = config.multi_vector_projector_dim
+
+        self.multi_vector_projector = nn.Linear(
+            in_features=self.config.text_config.hidden_size,
+            out_features=self.config.multi_vector_projector_dim,
+        )
+
+    def get_single_vector_embeddings(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        input_ids: Optional[torch.LongTensor] = None,
+    ) -> torch.Tensor:
+        """
+        Get the single-vector embeddings from the hidden states.
+        """
+        if self._input_has_image(input_ids[0]):  # got document image
+            img_start_positions = torch.where(
+                input_ids == self.config.vision_start_token_id
+            )[1]
+            img_end_positions = torch.where(
+                input_ids == self.config.vision_end_token_id
+            )[1]
+
+            batch_size, seq_len = input_ids.shape
+            position_indices = torch.arange(seq_len, device=input_ids.device).expand(
+                batch_size, -1
+            )
+            image_mask = (position_indices >= img_start_positions.unsqueeze(1)) & (
+                position_indices <= img_end_positions.unsqueeze(1)
+            )
+
+            masked_hidden_states = hidden_states * image_mask.unsqueeze(-1)
+            pooled_output = masked_hidden_states.sum(dim=1) / image_mask.sum(
+                dim=1, keepdim=True
+            )
+        else:  # got query text
+            pooled_output = torch.sum(
+                hidden_states * attention_mask.unsqueeze(-1), dim=1
+            ) / torch.sum(attention_mask, dim=1, keepdim=True)
+
+        return torch.nn.functional.normalize(pooled_output, dim=-1)
+
+    def get_multi_vector_embeddings(
+        self,
+        task_label: Union[str, List[str]],
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Project the hidden states to multi-vector embeddings.
+        """
+        multi_vec_emb = self.multi_vector_projector(
+            hidden_states, task_label=task_label
+        )
+        multi_vec_emb = torch.nn.functional.normalize(multi_vec_emb, dim=-1)
+        return multi_vec_emb * attention_mask.unsqueeze(-1)
+
+    def _input_has_image(self, input_ids):
+        return self.config.vision_start_token_id in input_ids
+
+    def forward(
+        self,
+        task_label: Union[str, List[str]],
+        input_ids: torch.LongTensor,
+        attention_mask: torch.Tensor,
+        output_vlm_last_hidden_states: bool = False,
+        **kwargs,
+    ) -> JinaEmbeddingsV4ModelOutput:
+        """
+        Forward pass through the model. Returns both single-vector and multi-vector embeddings.
+        Args:
+            input_ids (torch.Tensor): The input tokens tensor.
+            attention_mask (torch.Tensor): The attention mask tensor.
+        Returns:
+            JinaEmbeddingsV4ModelOutput:
+                vlm_last_hidden_states (torch.Tensor, optional): Last hidden states of the VLM.
+                single_vec_emb (torch.Tensor, optional): Single-vector embeddings.
+                multi_vec_emb (torch.Tensor, optional): Multi-vector embeddings.
+        """
+        # Forward pass through the VLM
+        hidden_states = self.get_last_hidden_states(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            task_label=task_label,
+            **kwargs,
+        )  # (batch_size, seq_length, hidden_size)
+        # Compute the embeddings
+        single_vec_emb = self.get_single_vector_embeddings(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            input_ids=input_ids,
+        )
+        multi_vec_emb = self.get_multi_vector_embeddings(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            task_label=task_label,
+        )
+
+        return JinaEmbeddingsV4ModelOutput(
+            vlm_last_hidden_states=(
+                hidden_states if output_vlm_last_hidden_states else None
+            ),
+            single_vec_emb=single_vec_emb,
+            multi_vec_emb=multi_vec_emb,
+        )
+
+    def _process_batches(
+        self,
+        data: List[Union[str, Image.Image]],
+        task_label: Union[str, List[str]],
+        processor_fn: Callable,
+        desc: str,
+        return_multivector: bool = False,
+        return_numpy: bool = False,
+        batch_size: int = 32,
+        truncate_dim: Optional[int] = None,
+    ) -> Union[np.ndarray, List[torch.Tensor]]:
+        dataloader = DataLoader(
+            dataset=data,
+            batch_size=batch_size,
+            shuffle=False,
+            collate_fn=processor_fn,
+        )
+        if return_multivector and len(data) > 1:
+            assert (
+                not return_numpy
+            ), "`return_numpy` is not supported when `return_multivector=True` and more than one data is encoded"
+        results = []
+        self.eval()
+        for batch in tqdm(dataloader, desc=desc, disable=self.verbosity == 0):
+            with torch.no_grad():
+                batch = {k: v.to(self.device) for k, v in batch.items()}
+                with torch.autocast(
+                    device_type=torch.device(self.device).type, dtype=torch.bfloat16
+                ):
+                    embeddings = self(**batch, task_label=task_label)
+                    if not return_multivector:
+                        embeddings = embeddings.single_vec_emb
+                        if truncate_dim is not None:
+                            embeddings = embeddings[:, :truncate_dim]
+                            embeddings = torch.nn.functional.normalize(
+                                embeddings, p=2, dim=-1
+                            )
+                    else:
+                        embeddings = embeddings.multi_vec_emb
+
+                    if return_multivector and not return_numpy:
+                        valid_tokens = batch["attention_mask"].bool()
+                        embeddings = [
+                            emb[mask] for emb, mask in zip(embeddings, valid_tokens)
+                        ]
+                        results.append(embeddings)
+                    else:
+                        results.append(
+                            embeddings.cpu()
+                            if return_numpy
+                            else list(torch.unbind(embeddings))
+                        )
+        if return_numpy:
+            return np.concatenate([result.numpy() for result in results], axis=0)
+        return [item for sublist in results for item in sublist]
+
+    def _validate_encoding_params(
+        self,
+        truncate_dim: Optional[int] = None,
+        prompt_name: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        encode_kwargs = {}
+        if prompt_name is not None:
+            if prompt_name not in PREFIX_DICT:
+                raise ValueError(
+                    f"Invalid prompt_name: {prompt_name}. Must be one of {list(PREFIX_DICT.keys())}."
+                )
+            else:
+                encode_kwargs["prefix"] = (
+                    PREFIX_DICT[prompt_name]
+                    if self.task != "text-matching"
+                    else PREFIX_DICT["query"]
+                )
+
+        truncate_dim = truncate_dim or self.config.truncate_dim
+        if truncate_dim is not None and truncate_dim not in self.config.matryoshka_dims:
+            raise ValueError(
+                f"Invalid truncate_dim: {truncate_dim}. Must be one of {self.config.matryoshka_dims}."
+            )
+        else:
+            encode_kwargs["truncate_dim"] = truncate_dim
+
+        return encode_kwargs
+
+    def _validate_task(self, task: Optional[str] = None) -> str:
+        if task is None:
+            if self.task is None:
+                raise ValueError(
+                    "Task must be specified before encoding data. You can set it either as a model property "
+                    "(e.g., model.task = 'retrieval') or pass it as an argument to the encode method."
+                )
+            task = self.task
+        else:
+            if task not in self.config.task_names:
+                raise ValueError(
+                    f"Invalid task: {task}. Must be one of {self.config.task_names}."
+                )
+        return task
+
+    def encode_text(
+        self,
+        texts: Union[str, List[str]],
+        task: Optional[str] = None,
+        max_length: int = 32768,
+        batch_size: int = 8,
+        return_multivector: bool = False,
+        return_numpy: bool = False,
+        truncate_dim: Optional[int] = None,
+        prompt_name: Optional[str] = None,
+    ) -> Union[List[torch.Tensor], torch.Tensor]:
+        """
+        Encodes a list of texts into embeddings.
+
+        Args:
+            texts: text or list of text strings to encode
+            max_length: Maximum token length for text processing
+            batch_size: Number of texts to process at once
+            return_multivector: Whether to return multi-vector embeddings instead of single-vector embeddings
+            return_numpy: Whether to return numpy arrays instead of torch tensors
+            truncate_dim: Dimension to truncate embeddings to (128, 256, 512, or 1024)
+            prompt_name: Type of text being encoded ('query' or 'passage')
+
+        Returns:
+            List of text embeddings as tensors or numpy arrays when encoding multiple texts, or single text embedding as tensor when encoding a single text
+        """
+        prompt_name = prompt_name or "query"
+        encode_kwargs = self._validate_encoding_params(
+            truncate_dim=truncate_dim, prompt_name=prompt_name
+        )
+
+        task = self._validate_task(task)
+
+        processor_fn = partial(
+            self.processor.process_texts,
+            max_length=max_length,
+            prefix=encode_kwargs.pop("prefix"),
+        )
+
+        return_list = isinstance(texts, list)
+
+        # If return_multivector is True and encoding multiple texts, ignore return_numpy
+        if return_multivector and return_list and len(texts) > 1:
+            if return_numpy:
+                print(
+                    "Warning: `return_numpy` is ignored when `return_multivector=True` and `len(texts) > 1`"
+                )
+            return_numpy = False
+
+        if isinstance(texts, str):
+            texts = [texts]
+
+        embeddings = self._process_batches(
+            data=texts,
+            processor_fn=processor_fn,
+            desc="Encoding texts...",
+            task_label=task,
+            return_multivector=return_multivector,
+            return_numpy=return_numpy,
+            batch_size=batch_size,
+            **encode_kwargs,
+        )
+
+        return embeddings if return_list else embeddings[0]
+
+    def _load_images_if_needed(
+        self, images: List[Union[str, Image.Image]]
+    ) -> List[Image.Image]:
+        loaded_images = []
+        for image in images:
+            if isinstance(image, str):
+                if image.startswith("http"):
+                    response = requests.get(image)
+                    image = Image.open(BytesIO(response.content)).convert("RGB")
+                else:
+                    image = Image.open(image).convert("RGB")
+            loaded_images.append(image)
+        return loaded_images
+
+    def encode_image(
+        self,
+        images: Union[str, Image.Image, List[Union[str, Image.Image]]],
+        task: Optional[str] = None,
+        batch_size: int = 8,
+        return_multivector: bool = False,
+        return_numpy: bool = False,
+        truncate_dim: Optional[int] = None,
+        max_pixels: Optional[int] = None,
+    ) -> Union[List[torch.Tensor], torch.Tensor]:
+        """
+        Encodes a list of images or a single image into embedding(s).
+
+        Args:
+            images: image(s) to encode, can be PIL Image(s), URL(s), or local file path(s)
+            batch_size: Number of images to process at once
+            return_multivector: Whether to return multi-vector embeddings instead of single-vector embeddings
+            return_numpy: Whether to return numpy arrays instead of torch tensors. If `return_multivector` is `True` and more than one image is encoded, this parameter is ignored.
+            truncate_dim: Dimension to truncate embeddings to (128, 256, 512, or 1024)
+            max_pixels: Maximum number of pixels to process per image
+
+        Returns:
+            List of image embeddings as tensors or numpy arrays when encoding multiple images, or single image embedding as tensor when encoding a single image
+        """
+        if max_pixels:
+            default_max_pixels = self.processor.image_processor.max_pixels
+            self.processor.image_processor.max_pixels = (
+                max_pixels  # change during encoding
+            )
+        encode_kwargs = self._validate_encoding_params(truncate_dim=truncate_dim)
+        task = self._validate_task(task)
+
+        return_list = isinstance(images, list)
+
+        # If return_multivector is True and encoding multiple images, ignore return_numpy
+        if return_multivector and return_list and len(images) > 1:
+            if return_numpy:
+                print(
+                    "Warning: `return_numpy` is ignored when `return_multivector=True` and `len(images) > 1`"
+                )
+            return_numpy = False
+
+        # Convert single image to list
+        if isinstance(images, (str, Image.Image)):
+            images = [images]
+
+        images = self._load_images_if_needed(images)
+        embeddings = self._process_batches(
+            data=images,
+            processor_fn=self.processor.process_images,
+            desc="Encoding images...",
+            task_label=task,
+            batch_size=batch_size,
+            return_multivector=return_multivector,
+            return_numpy=return_numpy,
+            **encode_kwargs,
+        )
+
+        if max_pixels:
+            self.processor.image_processor.max_pixels = default_max_pixels
+
+        return embeddings if return_list else embeddings[0]
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path,
+        *args,
+        **kwargs,
+    ):
+        """
+        Loads a pretrained model and configures it with the appropriate task adapter (`retrieval` by default).
+        """
+        if "torch_dtype" not in kwargs:
+            kwargs["torch_dtype"] = "auto"
+
+        kwargs["key_mapping"] = super()._checkpoint_conversion_mapping
+        if not is_flash_attn_2_available():
+            kwargs["attn_implementation"] = "sdpa"
+
+        base_model = super().from_pretrained(
+            pretrained_model_name_or_path, *args, **kwargs
+        )
+
+        # Configure adapter directory
+        if os.path.isdir(base_model.name_or_path):
+            adapter_dir = os.path.join(base_model.name_or_path, "adapters")
+        else:
+            adapter_cache_path = snapshot_download(
+                repo_id=base_model.name_or_path, allow_patterns=["adapters/*"]
+            )
+            adapter_dir = os.path.join(adapter_cache_path, "adapters")
+
+        lora_config = LoraConfig.from_pretrained(adapter_dir)
+        lora_config._custom_modules = {
+            torch.nn.modules.linear.Linear: partial(
+                MultiAdapterLinear,
+                task_names=base_model.config.task_names,
+            )
+        }
+        peft_model = PeftModel.from_pretrained(
+            model=base_model,
+            model_id=adapter_dir,
+            config=lora_config,
+        )
+
+        def task_getter(self):
+            return self.model.task
+
+        def task_setter(self, value):
+            self.model.task = value
+
+        peft_model.__class__.task = property(task_getter, task_setter)
+
+        return peft_model

modules.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b54acf92eab134d664abbbc5e42fd92f27535d3605666962c15dc3c32b2d9744
+size 168

preprocessor_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45e4720bcf1eac214b977199a08aae9e50b5797033987c389b46f601cb359a68
+size 726

results.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:304bc4d3d620c30be991b788fc0e5c3176495a404a6119d11dbc3c298d5ad575
+size 20325

special_tokens_map.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd
+size 613

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c5ae00e602b8860cbd784ba82a8aa14e8feecec692e7076590d014d7b7fdafa
+size 11421896

tokenizer_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13d28527663126ad9ab8a34aa6a4028b3f0b25f100defec89ee90b442d368dde
+size 7306

vocab.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910
+size 2776833