Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,8 @@ saved_model/**/* 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
+flash_attn-2.3.6+cu122torch2.1cxx11abiFALSE-cp39-cp39-linux_x86_64.whl filter=lfs diff=lfs merge=lfs -text
+images/demo_cli.gif filter=lfs diff=lfs merge=lfs -text
+masp_094_v2/video_chair/masp_pred_nvv_sample.json filter=lfs diff=lfs merge=lfs -text
+masp_094_v2/video_chair/uniform_pred_result.json filter=lfs diff=lfs merge=lfs -text
+masp_094_v2/video_chair/uniform_pred_result_detail_res.json filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "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.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      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 conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   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.

README.md

@@ -1,3 +1,458 @@
----
-license: mit
----
+# 🌋 LLaVA: Large Language and Vision Assistant
+
+*Visual instruction tuning towards large language and vision models with GPT-4 level capabilities.*
+
+[📢 [LLaVA-NeXT Blog](https://llava-vl.github.io/blog/2024-01-30-llava-next/)] [[Project Page](https://llava-vl.github.io/)] [[Demo](https://llava.hliu.cc/)]  [[Data](https://github.com/haotian-liu/LLaVA/blob/main/docs/Data.md)] [[Model Zoo](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)]
+
+🤝Community Contributions: [[llama.cpp](https://github.com/ggerganov/llama.cpp/pull/3436)] [[Colab](https://github.com/camenduru/LLaVA-colab)] [[🤗Space](https://huggingface.co/spaces/badayvedat/LLaVA)] [[Replicate](https://replicate.com/yorickvp/llava-13b)] [[AutoGen](https://github.com/microsoft/autogen/blob/main/notebook/agentchat_lmm_llava.ipynb)]  [[BakLLaVA](https://github.com/SkunkworksAI/BakLLaVA)]
+
+**Improved Baselines with Visual Instruction Tuning** [[Paper](https://arxiv.org/abs/2310.03744)] [[HF](https://huggingface.co/papers/2310.03744)] <br>
+[Haotian Liu](https://hliu.cc), [Chunyuan Li](https://chunyuan.li/), [Yuheng Li](https://yuheng-li.github.io/), [Yong Jae Lee](https://pages.cs.wisc.edu/~yongjaelee/)
+
+**Visual Instruction Tuning** (NeurIPS 2023, **Oral**) [[Paper](https://arxiv.org/abs/2304.08485)] [[HF](https://huggingface.co/papers/2304.08485)] <br>
+[Haotian Liu*](https://hliu.cc), [Chunyuan Li*](https://chunyuan.li/), [Qingyang Wu](https://scholar.google.ca/citations?user=HDiw-TsAAAAJ&hl=en/), [Yong Jae Lee](https://pages.cs.wisc.edu/~yongjaelee/) (*Equal Contribution)
+
+<!--p align="center">
+    <a href="https://llava.hliu.cc/"><img src="images/llava_logo.png" width="50%"></a> <br>
+    Generated by <a href="https://gligen.github.io/">GLIGEN</a> via "a cute lava llama with glasses" and box prompt
+</p-->
+
+
+## Release
+- [1/30] 🔥 LLaVA-NeXT (LLaVA-1.6) is out! With additional scaling to LLaVA-1.5, LLaVA-NeXT-34B outperforms Gemini Pro on some benchmarks. It can now process 4x more pixels and perform more tasks/applications than before. Check out the [blog post](https://llava-vl.github.io/blog/2024-01-30-llava-next/), and explore the [demo](https://llava.hliu.cc/)! Models are available in [Model Zoo](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md). Training/eval data and scripts coming soon.
+- [11/10] [LLaVA-Plus](https://llava-vl.github.io/llava-plus/) is released: Learning to Use Tools for Creating Multimodal Agents, with LLaVA-Plus (LLaVA that Plug and Learn to Use Skills). [[Project Page](https://llava-vl.github.io/llava-plus/)] [[Demo](https://llavaplus.ngrok.io/)] [[Code](https://github.com/LLaVA-VL/LLaVA-Plus-Codebase)] [[Paper](https://arxiv.org/abs/2311.05437)]
+- [11/2] [LLaVA-Interactive](https://llava-vl.github.io/llava-interactive/) is released: Experience the future of human-AI multimodal interaction with an all-in-one demo for Image Chat, Segmentation, Generation and Editing. [[Project Page](https://llava-vl.github.io/llava-interactive/)] [[Demo](https://llavainteractive.ngrok.io/)] [[Code](https://github.com/LLaVA-VL/LLaVA-Interactive-Demo)] [[Paper](https://arxiv.org/abs/2311.00571)]
+- [10/26] 🔥 LLaVA-1.5 with LoRA achieves comparable performance as full-model finetuning, with a reduced GPU RAM requirement ([ckpts](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md#llava-v15), [script](https://github.com/haotian-liu/LLaVA#train)). We also provide a [doc](https://github.com/haotian-liu/LLaVA/blob/main/docs/Finetune_Custom_Data.md) on how to finetune LLaVA-1.5 on your own dataset with LoRA.
+- [10/12] Check out the Korean LLaVA (Ko-LLaVA), created by ETRI, who has generously supported our research! [[🤗 Demo](https://huggingface.co/spaces/etri-vilab/Ko-LLaVA)]
+- [10/5] 🔥 LLaVA-1.5 is out! Achieving SoTA on 11 benchmarks, with just simple modifications to the original LLaVA, utilizes all public data, completes training in ~1 day on a single 8-A100 node, and surpasses methods like Qwen-VL-Chat that use billion-scale data. Check out the [technical report](https://arxiv.org/abs/2310.03744), and explore the [demo](https://llava.hliu.cc/)! Models are available in [Model Zoo](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md). The training data and scripts of LLaVA-1.5 are released [here](https://github.com/haotian-liu/LLaVA#train), and evaluation scripts are released [here](https://github.com/haotian-liu/LLaVA/blob/main/docs/Evaluation.md)!
+- [9/26] LLaVA is improved with reinforcement learning from human feedback (RLHF) to improve fact grounding and reduce hallucination. Check out the new SFT and RLHF checkpoints at project [[LLavA-RLHF]](https://llava-rlhf.github.io/)
+- [9/22] [LLaVA](https://arxiv.org/abs/2304.08485) is accepted by NeurIPS 2023 as **oral presentation**, and [LLaVA-Med](https://arxiv.org/abs/2306.00890) is accepted by NeurIPS 2023 Datasets and Benchmarks Track as **spotlight presentation**.
+
+<details>
+<summary>More</summary>
+
+- [11/6] Support **Intel** dGPU and CPU platforms. [More details here.](https://github.com/haotian-liu/LLaVA/tree/intel/docs/intel)
+- [10/12] LLaVA is now supported in [llama.cpp](https://github.com/ggerganov/llama.cpp/pull/3436) with 4-bit / 5-bit quantization support!
+- [10/11] The training data and scripts of LLaVA-1.5 are released [here](https://github.com/haotian-liu/LLaVA#train), and evaluation scripts are released [here](https://github.com/haotian-liu/LLaVA/blob/main/docs/Evaluation.md)!
+- [10/10] [Roboflow Deep Dive](https://blog.roboflow.com/first-impressions-with-llava-1-5/): First Impressions with LLaVA-1.5.
+- [9/20] We summarize our empirical study of training 33B and 65B LLaVA models in a [note](https://arxiv.org/abs/2309.09958). Further, if you are interested in the comprehensive review, evolution and trend of multimodal foundation models, please check out our recent survey paper [``Multimodal Foundation Models: From Specialists to General-Purpose Assistants''.](https://arxiv.org/abs/2309.10020)
+<p align="center">
+  <img src="https://github.com/Computer-Vision-in-the-Wild/CVinW_Readings/blob/main/images/mfm_evolution.jpeg?raw=true" width=50%/>
+</p>
+
+- [7/19] 🔥 We release a major upgrade, including support for LLaMA-2, LoRA training, 4-/8-bit inference, higher resolution (336x336), and a lot more. We release [LLaVA Bench](https://github.com/haotian-liu/LLaVA/blob/main/docs/LLaVA_Bench.md) for benchmarking open-ended visual chat with results from Bard and Bing-Chat. We also support and verify training with RTX 3090 and RTX A6000. Check out [LLaVA-from-LLaMA-2](https://github.com/haotian-liu/LLaVA/blob/main/docs/LLaVA_from_LLaMA2.md), and our [model zoo](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)!
+- [6/26] [CVPR 2023 Tutorial](https://vlp-tutorial.github.io/) on **Large Multimodal Models: Towards Building and Surpassing Multimodal GPT-4**!  Please check out [[Slides](https://datarelease.blob.core.windows.net/tutorial/vision_foundation_models_2023/slides/Chunyuan_cvpr2023_tutorial_lmm.pdf)] [[Notes](https://arxiv.org/abs/2306.14895)] [[YouTube](https://youtu.be/mkI7EPD1vp8)] [[Bilibli](https://www.bilibili.com/video/BV1Ng4y1T7v3/)].
+- [6/11] We released the preview for the most requested feature: DeepSpeed and LoRA support!  Please see documentations [here](./docs/LoRA.md).
+- [6/1] We released **LLaVA-Med: Large Language and Vision Assistant for Biomedicine**, a step towards building biomedical domain large language and vision models with GPT-4 level capabilities.  Checkout the [paper](https://arxiv.org/abs/2306.00890) and [page](https://github.com/microsoft/LLaVA-Med).
+- [5/6] We are releasing [LLaVA-Lighting-MPT-7B-preview](https://huggingface.co/liuhaotian/LLaVA-Lightning-MPT-7B-preview), based on MPT-7B-Chat!  See [here](#LLaVA-MPT-7b) for more details.
+- [5/2] 🔥 We are releasing LLaVA-Lighting!  Train a lite, multimodal GPT-4 with just $40 in 3 hours!  See [here](#train-llava-lightning) for more details.
+- [4/27] Thanks to the community effort, LLaVA-13B with 4-bit quantization allows you to run on a GPU with as few as 12GB VRAM!  Try it out [here](https://github.com/oobabooga/text-generation-webui/tree/main/extensions/llava).
+- [4/17] 🔥 We released **LLaVA: Large Language and Vision Assistant**. We propose visual instruction tuning, towards building large language and vision models with GPT-4 level capabilities.  Checkout the [paper](https://arxiv.org/abs/2304.08485) and [demo](https://llava.hliu.cc/).
+
+</details>
+
+<!-- <a href="https://llava.hliu.cc/"><img src="assets/demo.gif" width="70%"></a> -->
+
+[![Code License](https://img.shields.io/badge/Code%20License-Apache_2.0-green.svg)](https://github.com/tatsu-lab/stanford_alpaca/blob/main/LICENSE)
+**Usage and License Notices**: This project utilizes certain datasets and checkpoints that are subject to their respective original licenses. Users must comply with all terms and conditions of these original licenses, including but not limited to the [OpenAI Terms of Use](https://openai.com/policies/terms-of-use) for the dataset and the specific licenses for base language models for checkpoints trained using the dataset (e.g. [Llama community license](https://ai.meta.com/llama/license/) for LLaMA-2 and Vicuna-v1.5). This project does not impose any additional constraints beyond those stipulated in the original licenses. Furthermore, users are reminded to ensure that their use of the dataset and checkpoints is in compliance with all applicable laws and regulations.
+
+
+## Contents
+- [Install](#install)
+- [LLaVA Weights](#llava-weights)
+- [Demo](#Demo)
+- [Model Zoo](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)
+- [Dataset](https://github.com/haotian-liu/LLaVA/blob/main/docs/Data.md)
+- [Train](#train)
+- [Evaluation](#evaluation)
+
+## Install
+
+If you are not using Linux, do *NOT* proceed, see instructions for [macOS](https://github.com/haotian-liu/LLaVA/blob/main/docs/macOS.md) and [Windows](https://github.com/haotian-liu/LLaVA/blob/main/docs/Windows.md).
+
+1. Clone this repository and navigate to LLaVA folder
+```bash
+git clone https://github.com/haotian-liu/LLaVA.git
+cd LLaVA
+```
+
+2. Install Package
+```Shell
+conda create -n llava python=3.10 -y
+conda activate llava
+pip install --upgrade pip  # enable PEP 660 support
+pip install -e .
+```
+
+3. Install additional packages for training cases
+```
+pip install -e ".[train]"
+pip install flash-attn --no-build-isolation
+```
+
+### Upgrade to latest code base
+
+```Shell
+git pull
+pip install -e .
+
+# if you see some import errors when you upgrade, please try running the command below (without #)
+# pip install flash-attn --no-build-isolation --no-cache-dir
+```
+
+### Quick Start With HuggingFace
+
+<details>
+<summary>Example Code</summary>
+
+```Python
+from llava.model.builder import load_pretrained_model
+from llava.mm_utils import get_model_name_from_path
+from llava.eval.run_llava import eval_model
+
+model_path = "liuhaotian/llava-v1.5-7b"
+
+tokenizer, model, image_processor, context_len = load_pretrained_model(
+    model_path=model_path,
+    model_base=None,
+    model_name=get_model_name_from_path(model_path)
+)
+```
+
+Check out the details wth the `load_pretrained_model` function in `llava/model/builder.py`.
+
+You can also use the `eval_model` function in `llava/eval/run_llava.py` to get the output easily. By doing so, you can use this code on Colab directly after downloading this repository.
+
+``` python
+model_path = "liuhaotian/llava-v1.5-7b"
+prompt = "What are the things I should be cautious about when I visit here?"
+image_file = "https://llava-vl.github.io/static/images/view.jpg"
+
+args = type('Args', (), {
+    "model_path": model_path,
+    "model_base": None,
+    "model_name": get_model_name_from_path(model_path),
+    "query": prompt,
+    "conv_mode": None,
+    "image_file": image_file,
+    "sep": ",",
+    "temperature": 0,
+    "top_p": None,
+    "num_beams": 1,
+    "max_new_tokens": 512
+})()
+
+eval_model(args)
+```
+</details>
+
+## LLaVA Weights
+Please check out our [Model Zoo](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md) for all public LLaVA checkpoints, and the instructions of how to use the weights.
+
+## Demo
+
+### Gradio Web UI
+
+To launch a Gradio demo locally, please run the following commands one by one. If you plan to launch multiple model workers to compare between different checkpoints, you only need to launch the controller and the web server *ONCE*.
+
+```mermaid
+flowchart BT
+    %% Declare Nodes
+    gws("Gradio (UI Server)")
+    c("Controller (API Server):<br/>PORT: 10000")
+    mw7b("Model Worker:<br/>llava-v1.5-7b<br/>PORT: 40000")
+    mw13b("Model Worker:<br/>llava-v1.5-13b<br/>PORT: 40001")
+    sglw13b("SGLang Backend:<br/>llava-v1.6-34b<br/>http://localhost:30000")
+    lsglw13b("SGLang Worker:<br/>llava-v1.6-34b<br/>PORT: 40002")
+
+    %% Declare Styles
+    classDef data fill:#3af,stroke:#48a,stroke-width:2px,color:#444
+    classDef success fill:#8f8,stroke:#0a0,stroke-width:2px,color:#444
+    classDef failure fill:#f88,stroke:#f00,stroke-width:2px,color:#444
+
+    %% Assign Styles
+    class id,od data;
+    class cimg,cs_s,scsim_s success;
+    class ncimg,cs_f,scsim_f failure;
+
+    subgraph Demo Connections
+        direction BT
+        c<-->gws
+        
+        mw7b<-->c
+        mw13b<-->c
+        lsglw13b<-->c
+        sglw13b<-->lsglw13b
+    end
+```
+
+#### Launch a controller
+```Shell
+python -m llava.serve.controller --host 0.0.0.0 --port 10000
+```
+
+#### Launch a gradio web server.
+```Shell
+python -m llava.serve.gradio_web_server --controller http://localhost:10000 --model-list-mode reload
+```
+You just launched the Gradio web interface. Now, you can open the web interface with the URL printed on the screen. You may notice that there is no model in the model list. Do not worry, as we have not launched any model worker yet. It will be automatically updated when you launch a model worker.
+
+#### Launch a SGLang worker
+
+This is the recommended way to serve LLaVA model with high throughput, and you need to install SGLang first. Note that currently `4-bit` quantization is not supported yet on SGLang-LLaVA, and if you have limited GPU VRAM, please check out model worker with [quantization](https://github.com/haotian-liu/LLaVA?tab=readme-ov-file#launch-a-model-worker-4-bit-8-bit-inference-quantized).
+
+```Shell
+pip install "sglang[all]"
+```
+
+You'll first launch a SGLang backend worker which will execute the models on GPUs. Remember the `--port` you've set and you'll use that later.
+
+```Shell
+# Single GPU
+CUDA_VISIBLE_DEVICES=0 python3 -m sglang.launch_server --model-path liuhaotian/llava-v1.5-7b --tokenizer-path llava-hf/llava-1.5-7b-hf --port 30000
+
+# Multiple GPUs with tensor parallel
+CUDA_VISIBLE_DEVICES=0,1 python3 -m sglang.launch_server --model-path liuhaotian/llava-v1.5-13b --tokenizer-path llava-hf/llava-1.5-13b-hf --port 30000 --tp 2
+```
+
+Tokenizers (temporary): `llava-hf/llava-1.5-7b-hf`, `llava-hf/llava-1.5-13b-hf`, `liuhaotian/llava-v1.6-34b-tokenizer`.
+
+You'll then launch a LLaVA-SGLang worker that will communicate between LLaVA controller and SGLang backend to route the requests. Set `--sgl-endpoint` to `http://127.0.0.1:port` where `port` is the one you just set (default: 30000).
+
+```Shell
+python -m llava.serve.sglang_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --sgl-endpoint http://127.0.0.1:30000
+```
+
+#### Launch a model worker
+
+This is the actual *worker* that performs the inference on the GPU.  Each worker is responsible for a single model specified in `--model-path`.
+
+```Shell
+python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path liuhaotian/llava-v1.5-13b
+```
+Wait until the process finishes loading the model and you see "Uvicorn running on ...".  Now, refresh your Gradio web UI, and you will see the model you just launched in the model list.
+
+You can launch as many workers as you want, and compare between different model checkpoints in the same Gradio interface. Please keep the `--controller` the same, and modify the `--port` and `--worker` to a different port number for each worker.
+```Shell
+python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port <different from 40000, say 40001> --worker http://localhost:<change accordingly, i.e. 40001> --model-path <ckpt2>
+```
+
+If you are using an Apple device with an M1 or M2 chip, you can specify the mps device by using the `--device` flag: `--device mps`.
+
+#### Launch a model worker (Multiple GPUs, when GPU VRAM <= 24GB)
+
+If the VRAM of your GPU is less than 24GB (e.g., RTX 3090, RTX 4090, etc.), you may try running it with multiple GPUs. Our latest code base will automatically try to use multiple GPUs if you have more than one GPU. You can specify which GPUs to use with `CUDA_VISIBLE_DEVICES`. Below is an example of running with the first two GPUs.
+
+```Shell
+CUDA_VISIBLE_DEVICES=0,1 python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path liuhaotian/llava-v1.5-13b
+```
+
+#### Launch a model worker (4-bit, 8-bit inference, quantized)
+
+You can launch the model worker with quantized bits (4-bit, 8-bit), which allows you to run the inference with reduced GPU memory footprint, potentially allowing you to run on a GPU with as few as 12GB VRAM. Note that inference with quantized bits may not be as accurate as the full-precision model. Simply append `--load-4bit` or `--load-8bit` to the **model worker** command that you are executing. Below is an example of running with 4-bit quantization.
+
+```Shell
+python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path liuhaotian/llava-v1.5-13b --load-4bit
+```
+
+#### Launch a model worker (LoRA weights, unmerged)
+
+You can launch the model worker with LoRA weights, without merging them with the base checkpoint, to save disk space. There will be additional loading time, while the inference speed is the same as the merged checkpoints. Unmerged LoRA checkpoints do not have `lora-merge` in the model name, and are usually much smaller (less than 1GB) than the merged checkpoints (13G for 7B, and 25G for 13B).
+
+To load unmerged LoRA weights, you simply need to pass an additional argument `--model-base`, which is the base LLM that is used to train the LoRA weights. You can check the base LLM of each LoRA weights in the [model zoo](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md).
+
+```Shell
+python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path liuhaotian/llava-v1-0719-336px-lora-vicuna-13b-v1.3 --model-base lmsys/vicuna-13b-v1.3
+```
+
+### CLI Inference
+
+Chat about images using LLaVA without the need of Gradio interface. It also supports multiple GPUs, 4-bit and 8-bit quantized inference. With 4-bit quantization, for our LLaVA-1.5-7B, it uses less than 8GB VRAM on a single GPU.
+
+```Shell
+python -m llava.serve.cli \
+    --model-path liuhaotian/llava-v1.5-7b \
+    --image-file "https://llava-vl.github.io/static/images/view.jpg" \
+    --load-4bit
+```
+
+<img src="images/demo_cli.gif" width="70%">
+
+## Train
+
+*Below is the latest training configuration for LLaVA v1.5. For legacy models, please refer to README of [this](https://github.com/haotian-liu/LLaVA/tree/v1.0.1) version for now. We'll add them in a separate doc later.*
+
+LLaVA training consists of two stages: (1) feature alignment stage: use our 558K subset of the LAION-CC-SBU dataset to connect a *frozen pretrained* vision encoder to a *frozen LLM*; (2) visual instruction tuning stage: use 150K GPT-generated multimodal instruction-following data, plus around 515K VQA data from academic-oriented tasks, to teach the model to follow multimodal instructions.
+
+LLaVA is trained on 8 A100 GPUs with 80GB memory. To train on fewer GPUs, you can reduce the `per_device_train_batch_size` and increase the `gradient_accumulation_steps` accordingly. Always keep the global batch size the same: `per_device_train_batch_size` x `gradient_accumulation_steps` x `num_gpus`.
+
+### Hyperparameters
+We use a similar set of hyperparameters as Vicuna in finetuning.  Both hyperparameters used in pretraining and finetuning are provided below.
+
+1. Pretraining
+
+| Hyperparameter | Global Batch Size | Learning rate | Epochs | Max length | Weight decay |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| LLaVA-v1.5-13B | 256 | 1e-3 | 1 | 2048 | 0 |
+
+2. Finetuning
+
+| Hyperparameter | Global Batch Size | Learning rate | Epochs | Max length | Weight decay |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| LLaVA-v1.5-13B | 128 | 2e-5 | 1 | 2048 | 0 |
+
+### Download Vicuna checkpoints (automatically)
+
+Our base model Vicuna v1.5, which is an instruction-tuned chatbot, will be downloaded automatically when you run our provided training scripts. No action is needed.
+
+### Pretrain (feature alignment)
+
+Please download the 558K subset of the LAION-CC-SBU dataset with BLIP captions we use in the paper [here](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain).
+
+Pretrain takes around 5.5 hours for LLaVA-v1.5-13B on 8x A100 (80G), due to the increased resolution to 336px. It takes around 3.5 hours for LLaVA-v1.5-7B.
+
+Training script with DeepSpeed ZeRO-2: [`pretrain.sh`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/v1_5/pretrain.sh).
+
+- `--mm_projector_type mlp2x_gelu`: the two-layer MLP vision-language connector.
+- `--vision_tower openai/clip-vit-large-patch14-336`: CLIP ViT-L/14 336px.
+
+<details>
+<summary>Pretrain takes around 20 hours for LLaVA-7B on 8x V100 (32G)</summary>
+
+ We provide training script with DeepSpeed [here](https://github.com/haotian-liu/LLaVA/blob/main/scripts/pretrain_xformers.sh).
+Tips:
+- If you are using V100 which is not supported by FlashAttention, you can use the [memory-efficient attention](https://arxiv.org/abs/2112.05682) implemented in [xFormers](https://github.com/facebookresearch/xformers). Install xformers and replace `llava/train/train_mem.py` above with [llava/train/train_xformers.py](llava/train/train_xformers.py).
+</details>
+
+### Visual Instruction Tuning
+
+1. Prepare data
+
+Please download the annotation of the final mixture our instruction tuning data [llava_v1_5_mix665k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/llava_v1_5_mix665k.json), and download the images from constituting datasets:
+
+- COCO: [train2017](http://images.cocodataset.org/zips/train2017.zip)
+- GQA: [images](https://downloads.cs.stanford.edu/nlp/data/gqa/images.zip)
+- OCR-VQA: [download script](https://drive.google.com/drive/folders/1_GYPY5UkUy7HIcR0zq3ZCFgeZN7BAfm_?usp=sharing), **we save all files as `.jpg`**
+- TextVQA: [train_val_images](https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip)
+- VisualGenome: [part1](https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip), [part2](https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip)
+
+After downloading all of them, organize the data as follows in `./playground/data`,
+
+```
+├── coco
+│   └── train2017
+├── gqa
+│   └── images
+├── ocr_vqa
+│   └── images
+├── textvqa
+│   └── train_images
+└── vg
+    ├── VG_100K
+    └── VG_100K_2
+```
+
+2. Start training!
+
+You may download our pretrained projectors in [Model Zoo](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md). It is not recommended to use legacy projectors, as they may be trained with a different version of the codebase, and if any option is off, the model will not function/train as we expected.
+
+Visual instruction tuning takes around 20 hours for LLaVA-v1.5-13B on 8x A100 (80G), due to the increased resolution to 336px. It takes around 10 hours for LLaVA-v1.5-7B on 8x A100 (40G).
+
+Training script with DeepSpeed ZeRO-3: [`finetune.sh`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/v1_5/finetune.sh).
+
+If you are do not have enough GPU memory:
+
+- Use LoRA: [`finetune_lora.sh`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/v1_5/finetune_lora.sh). We are able to fit 13B training in 8-A100-40G/8-A6000, and 7B training in 8-RTX3090. Make sure `per_device_train_batch_size*gradient_accumulation_steps` is the same as the provided script for best reproducibility.
+- Replace `zero3.json` with `zero3_offload.json` which offloads some parameters to CPU RAM. This slows down the training speed.
+
+If you are interested in finetuning LLaVA model to your own task/data, please check out [`Finetune_Custom_Data.md`](https://github.com/haotian-liu/LLaVA/blob/main/docs/Finetune_Custom_Data.md)。
+
+New options to note:
+
+- `--mm_projector_type mlp2x_gelu`: the two-layer MLP vision-language connector.
+- `--vision_tower openai/clip-vit-large-patch14-336`: CLIP ViT-L/14 336px.
+- `--image_aspect_ratio pad`: this pads the non-square images to square, instead of cropping them; it slightly reduces hallucination.
+- `--group_by_modality_length True`: this should only be used when your instruction tuning dataset contains both language (e.g. ShareGPT) and multimodal (e.g. LLaVA-Instruct). It makes the training sampler only sample a single modality (either image or language) during training, which we observe to speed up training by ~25%, and does not affect the final outcome.
+
+## Evaluation
+
+In LLaVA-1.5, we evaluate models on a diverse set of 12 benchmarks. To ensure the reproducibility, we evaluate the models with greedy decoding. We do not evaluate using beam search to make the inference process consistent with the chat demo of real-time outputs.
+
+See [Evaluation.md](https://github.com/haotian-liu/LLaVA/blob/main/docs/Evaluation.md).
+
+### GPT-assisted Evaluation
+
+Our GPT-assisted evaluation pipeline for multimodal modeling is provided for a comprehensive understanding of the capabilities of vision-language models.  Please see our paper for more details.
+
+1. Generate LLaVA responses
+
+```Shell
+python model_vqa.py \
+    --model-path ./checkpoints/LLaVA-13B-v0 \
+    --question-file \
+    playground/data/coco2014_val_qa_eval/qa90_questions.jsonl \
+    --image-folder \
+    /path/to/coco2014_val \
+    --answers-file \
+    /path/to/answer-file-our.jsonl
+```
+
+2. Evaluate the generated responses.  In our case, [`answer-file-ref.jsonl`](./playground/data/coco2014_val_qa_eval/qa90_gpt4_answer.jsonl) is the response generated by text-only GPT-4 (0314), with the context captions/boxes provided.
+
+```Shell
+OPENAI_API_KEY="sk-***********************************" python llava/eval/eval_gpt_review_visual.py \
+    --question playground/data/coco2014_val_qa_eval/qa90_questions.jsonl \
+    --context llava/eval/table/caps_boxes_coco2014_val_80.jsonl \
+    --answer-list \
+    /path/to/answer-file-ref.jsonl \
+    /path/to/answer-file-our.jsonl \
+    --rule llava/eval/table/rule.json \
+    --output /path/to/review.json
+```
+
+3. Summarize the evaluation results
+
+```Shell
+python summarize_gpt_review.py
+```
+
+## Citation
+
+If you find LLaVA useful for your research and applications, please cite using this BibTeX:
+```bibtex
+@misc{liu2024llavanext,
+    title={LLaVA-NeXT: Improved reasoning, OCR, and world knowledge},
+    url={https://llava-vl.github.io/blog/2024-01-30-llava-next/},
+    author={Liu, Haotian and Li, Chunyuan and Li, Yuheng and Li, Bo and Zhang, Yuanhan and Shen, Sheng and Lee, Yong Jae},
+    month={January},
+    year={2024}
+}
+
+@misc{liu2023improvedllava,
+      title={Improved Baselines with Visual Instruction Tuning}, 
+      author={Liu, Haotian and Li, Chunyuan and Li, Yuheng and Lee, Yong Jae},
+      publisher={arXiv:2310.03744},
+      year={2023},
+}
+
+@misc{liu2023llava,
+      title={Visual Instruction Tuning}, 
+      author={Liu, Haotian and Li, Chunyuan and Wu, Qingyang and Lee, Yong Jae},
+      publisher={NeurIPS},
+      year={2023},
+}
+```
+
+## Acknowledgement
+
+- [Vicuna](https://github.com/lm-sys/FastChat): the codebase we built upon, and our base model Vicuna-13B that has the amazing language capabilities!
+
+## Related Projects
+
+- [Instruction Tuning with GPT-4](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)
+- [LLaVA-Med: Training a Large Language-and-Vision Assistant for Biomedicine in One Day](https://github.com/microsoft/LLaVA-Med)
+- [Otter: In-Context Multi-Modal Instruction Tuning](https://github.com/Luodian/Otter)
+
+For future project ideas, please check out:
+- [SEEM: Segment Everything Everywhere All at Once](https://github.com/UX-Decoder/Segment-Everything-Everywhere-All-At-Once)
+- [Grounded-Segment-Anything](https://github.com/IDEA-Research/Grounded-Segment-Anything) to detect, segment, and generate anything by marrying [Grounding DINO](https://github.com/IDEA-Research/GroundingDINO) and [Segment-Anything](https://github.com/facebookresearch/segment-anything).

blip2_pretrained_flant5xxl.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b3839ea6c617f315ead9bf4036bbb0f0cf6bf62695ecfc14968ea626af03a29
+size 433481467

cog.yaml

@@ -0,0 +1,36 @@
+# Configuration for Cog ⚙️
+# Reference: https://github.com/replicate/cog/blob/main/docs/yaml.md
+
+build:
+  gpu: true
+
+  python_version: "3.10"
+
+  python_packages:
+    - "torch==2.1.2"
+    - "accelerate==0.27.2"
+    - "bitsandbytes==0.41.0"
+    - "deepspeed==0.12.2"
+    - "einops-exts==0.0.4"
+    - "einops==0.6.1"
+    - "gradio==3.35.2"
+    - "gradio_client==0.2.9"
+    - "httpx==0.24.0"
+    - "markdown2==2.4.10"
+    - "numpy==1.26.0"
+    - "peft==0.9.0"
+    - "scikit-learn==1.2.2"
+    - "sentencepiece==0.1.99"
+    - "shortuuid==1.0.11"
+    - "timm==0.6.13"
+    - "tokenizers==0.15.2"
+    - "torchvision==0.16.2"
+    - "transformers==4.38.1"
+    - "wandb==0.15.12"
+    - "wavedrom==2.0.3.post3"
+    - "Pygments==2.16.1"
+  run:
+    - curl -o /usr/local/bin/pget -L "https://github.com/replicate/pget/releases/download/v0.0.3/pget" && chmod +x /usr/local/bin/pget
+
+# predict.py defines how predictions are run on your model
+predict: "predict.py:Predictor"

docs/Customize_Component.md

@@ -0,0 +1,20 @@
+# Customize Components in LLaVA
+
+This is an initial guide on how to replace the LLMs, visual encoders, etc. with your choice of components.
+
+## LLM
+
+It is quite simple to swap out LLaMA to any other LLMs.  You can refer to our implementation of [`llava_llama.py`](https://raw.githubusercontent.com/haotian-liu/LLaVA/main/llava/model/language_model/llava_llama.py) for an example of how to replace the LLM.
+
+Although it may seem that it still needs ~100 lines of code, most of them are copied from the original `llama.py` from HF.  The only part that is different is to insert some lines for processing the multimodal inputs.
+
+In `forward` function, you can see that we call `self.prepare_inputs_labels_for_multimodal` to process the multimodal inputs.  This function is defined in `LlavaMetaForCausalLM` and you just need to insert it into the `forward` function of your LLM.
+
+In `prepare_inputs_for_generation` function, you can see that we add `images` to the `model_inputs`.  This is because we need to pass the images to the LLM during generation.
+
+These are basically all the changes you need to make to replace the LLM.
+
+## Visual Encoder
+
+You can check out [`clip_encoder.py`](https://github.com/haotian-liu/LLaVA/blob/main/llava/model/multimodal_encoder/clip_encoder.py) on how we implement the CLIP visual encoder.
+

docs/Data.md

@@ -0,0 +1,29 @@
+## Data
+
+| Data file name | Size |
+| --- | ---: |
+| [llava_instruct_150k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/llava_instruct_150k.json) | 229 MB |
+| [llava_instruct_80k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/llava_instruct_80k.json) | 229 MB |
+| [conversation_58k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/conversation_58k.json) | 126 MB |
+| [detail_23k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/detail_23k.json) | 20.5 MB |
+| [complex_reasoning_77k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/complex_reasoning_77k.json) | 79.6 MB |
+
+### Pretraining Dataset
+The pretraining dataset used in this release is a subset of CC-3M dataset, filtered with a more balanced concept coverage distribution.  Please see [here](https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K) for a detailed description of the dataset structure and how to download the images.
+
+If you already have CC-3M dataset on your disk, the image names follow this format: `GCC_train_000000000.jpg`.  You may edit the `image` field correspondingly if necessary.
+
+| Data | Chat File | Meta Data | Size |
+| --- |  --- |  --- | ---: |
+| CC-3M Concept-balanced 595K | [chat.json](https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K/blob/main/chat.json) | [metadata.json](https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K/blob/main/metadata.json) | 211 MB
+| LAION/CC/SBU BLIP-Caption Concept-balanced 558K | [blip_laion_cc_sbu_558k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain/blob/main/blip_laion_cc_sbu_558k.json) | [metadata.json](#) | 181 MB
+
+**Important notice**: Upon the request from the community, as ~15% images of the original CC-3M dataset are no longer accessible, we upload [`images.zip`](https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K/blob/main/images.zip) for better reproducing our work in research community. It must not be used for any other purposes. The use of these images must comply with the CC-3M license. This may be taken down at any time when requested by the original CC-3M dataset owner or owners of the referenced images.
+
+### GPT-4 Prompts
+
+We provide our prompts and few-shot samples for GPT-4 queries, to better facilitate research in this domain.  Please check out the [`prompts`](https://github.com/haotian-liu/LLaVA/tree/main/playground/data/prompts) folder for three kinds of questions: conversation, detail description, and complex reasoning.
+
+They are organized in a format of `system_message.txt` for system message, pairs of `abc_caps.txt` for few-shot sample user input, and `abc_conv.txt` for few-shot sample reference output.
+
+Note that you may find them in different format. For example, `conversation` is in `jsonl`, and detail description is answer-only.  The selected format in our preliminary experiments works slightly better than a limited set of alternatives that we tried: `jsonl`, more natural format, answer-only.  If interested, you may try other variants or conduct more careful study in this.  Contributions are welcomed!

docs/Evaluation.md

@@ -0,0 +1,167 @@
+# Evaluation
+
+In LLaVA-1.5, we evaluate models on a diverse set of 12 benchmarks. To ensure the reproducibility, we evaluate the models with greedy decoding. We do not evaluate using beam search to make the inference process consistent with the chat demo of real-time outputs.
+
+Currently, we mostly utilize the official toolkit or server for the evaluation.
+
+## Evaluate on Custom Datasets
+
+You can evaluate LLaVA on your custom datasets by converting your dataset to LLaVA's jsonl format, and evaluate using [`model_vqa.py`](https://github.com/haotian-liu/LLaVA/blob/main/llava/eval/model_vqa.py).
+
+Below we provide a general guideline for evaluating datasets with some common formats.
+
+1. Short-answer (e.g. VQAv2, MME).
+
+```
+<question>
+Answer the question using a single word or phrase.
+```
+
+2. Option-only for multiple-choice (e.g. MMBench, SEED-Bench).
+
+```
+<question>
+A. <option_1>
+B. <option_2>
+C. <option_3>
+D. <option_4>
+Answer with the option's letter from the given choices directly.
+```
+
+3. Natural QA (e.g. LLaVA-Bench, MM-Vet).
+
+No postprocessing is needed.
+
+## Scripts
+
+Before preparing task-specific data, **you MUST first download [eval.zip](https://drive.google.com/file/d/1atZSBBrAX54yYpxtVVW33zFvcnaHeFPy/view?usp=sharing)**. It contains custom annotations, scripts, and the prediction files with LLaVA v1.5. Extract to `./playground/data/eval`. This also provides a general structure for all datasets.
+
+### VQAv2
+
+1. Download [`test2015`](http://images.cocodataset.org/zips/test2015.zip) and put it under `./playground/data/eval/vqav2`.
+2. Multi-GPU inference.
+```Shell
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/v1_5/eval/vqav2.sh
+```
+3. Submit the results to the [evaluation server](https://eval.ai/web/challenges/challenge-page/830/my-submission): `./playground/data/eval/vqav2/answers_upload`.
+
+### GQA
+
+1. Download the [data](https://cs.stanford.edu/people/dorarad/gqa/download.html) and [evaluation scripts](https://cs.stanford.edu/people/dorarad/gqa/evaluate.html) following the official instructions and put under `./playground/data/eval/gqa/data`. You may need to modify `eval.py` as [this](https://gist.github.com/haotian-liu/db6eddc2a984b4cbcc8a7f26fd523187) due to the missing assets in the GQA v1.2 release.
+2. Multi-GPU inference.
+```Shell
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/v1_5/eval/gqa.sh
+```
+
+### VisWiz
+
+1. Download [`test.json`](https://vizwiz.cs.colorado.edu/VizWiz_final/vqa_data/Annotations.zip) and extract [`test.zip`](https://vizwiz.cs.colorado.edu/VizWiz_final/images/test.zip) to `test`. Put them under `./playground/data/eval/vizwiz`.
+2. Single-GPU inference.
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/vizwiz.sh
+```
+3. Submit the results to the [evaluation server](https://eval.ai/web/challenges/challenge-page/1911/my-submission): `./playground/data/eval/vizwiz/answers_upload`.
+
+### ScienceQA
+
+1. Under `./playground/data/eval/scienceqa`, download `images`, `pid_splits.json`, `problems.json` from the `data/scienceqa` folder of the ScienceQA [repo](https://github.com/lupantech/ScienceQA).
+2. Single-GPU inference and evaluate.
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/sqa.sh
+```
+
+### TextVQA
+
+1. Download [`TextVQA_0.5.1_val.json`](https://dl.fbaipublicfiles.com/textvqa/data/TextVQA_0.5.1_val.json) and [images](https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip) and extract to `./playground/data/eval/textvqa`.
+2. Single-GPU inference and evaluate.
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/textvqa.sh
+```
+
+### POPE
+
+1. Download `coco` from [POPE](https://github.com/AoiDragon/POPE/tree/e3e39262c85a6a83f26cf5094022a782cb0df58d/output/coco) and put under `./playground/data/eval/pope`.
+2. Single-GPU inference and evaluate.
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/pope.sh
+```
+
+### MME
+
+1. Download the data following the official instructions [here](https://github.com/BradyFU/Awesome-Multimodal-Large-Language-Models/tree/Evaluation).
+2. Downloaded images to `MME_Benchmark_release_version`.
+3. put the official `eval_tool` and `MME_Benchmark_release_version` under `./playground/data/eval/MME`.
+4. Single-GPU inference and evaluate.
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/mme.sh
+```
+
+### MMBench
+
+1. Download [`mmbench_dev_20230712.tsv`](https://download.openmmlab.com/mmclassification/datasets/mmbench/mmbench_dev_20230712.tsv) and put under `./playground/data/eval/mmbench`.
+2. Single-GPU inference.
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/mmbench.sh
+```
+3. Submit the results to the [evaluation server](https://opencompass.org.cn/leaderboard-multimodal): `./playground/data/eval/mmbench/answers_upload/mmbench_dev_20230712`.
+
+### MMBench-CN
+
+1. Download [`mmbench_dev_cn_20231003.tsv`](https://download.openmmlab.com/mmclassification/datasets/mmbench/mmbench_dev_cn_20231003.tsv) and put under `./playground/data/eval/mmbench`.
+2. Single-GPU inference.
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/mmbench_cn.sh
+```
+3. Submit the results to the evaluation server: `./playground/data/eval/mmbench/answers_upload/mmbench_dev_cn_20231003`.
+
+
+### SEED-Bench
+
+1. Following the official [instructions](https://github.com/AILab-CVC/SEED-Bench/blob/main/DATASET.md) to download the images and the videos. Put images under `./playground/data/eval/seed_bench/SEED-Bench-image`.
+2. Extract the video frame in the middle from the downloaded videos, and put them under `./playground/data/eval/seed_bench/SEED-Bench-video-image`. We provide our script `extract_video_frames.py` modified from the official one.
+3. Multiple-GPU inference and evaluate.
+```Shell
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/v1_5/eval/seed.sh
+```
+4. Optionally, submit the results to the leaderboard: `./playground/data/eval/seed_bench/answers_upload` using the official jupyter notebook.
+
+### LLaVA-Bench-in-the-Wild
+
+1. Extract contents of [`llava-bench-in-the-wild`](https://huggingface.co/datasets/liuhaotian/llava-bench-in-the-wild) to `./playground/data/eval/llava-bench-in-the-wild`.
+2. Single-GPU inference and evaluate.
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/llavabench.sh
+```
+
+### MM-Vet
+
+1. Extract [`mm-vet.zip`](https://github.com/yuweihao/MM-Vet/releases/download/v1/mm-vet.zip) to `./playground/data/eval/mmvet`.
+2. Single-GPU inference.
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/mmvet.sh
+```
+3. Evaluate the predictions in `./playground/data/eval/mmvet/results` using the official jupyter notebook.
+
+## More Benchmarks
+
+Below are awesome benchmarks for multimodal understanding from the research community, that are not initially included in the LLaVA-1.5 release.
+
+### Q-Bench
+
+1. Download [`llvisionqa_dev.json`](https://huggingface.co/datasets/nanyangtu/LLVisionQA-QBench/resolve/main/llvisionqa_dev.json) (for `dev`-subset) and [`llvisionqa_test.json`](https://huggingface.co/datasets/nanyangtu/LLVisionQA-QBench/resolve/main/llvisionqa_test.json) (for `test`-subset). Put them under `./playground/data/eval/qbench`. 
+2. Download and extract [images](https://huggingface.co/datasets/nanyangtu/LLVisionQA-QBench/resolve/main/images_llvisionqa.tar) and put all the images directly under `./playground/data/eval/qbench/images_llviqionqa`.
+3. Single-GPU inference (change `dev` to `test` for evaluation on test set).
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/qbench.sh dev
+```
+4. Submit the results by instruction [here](https://github.com/VQAssessment/Q-Bench#option-1-submit-results): `./playground/data/eval/qbench/llvisionqa_dev_answers.jsonl`.
+
+### Chinese-Q-Bench
+
+1. Download [`质衡-问答-验证集.json`](https://huggingface.co/datasets/nanyangtu/LLVisionQA-QBench/resolve/main/%E8%B4%A8%E8%A1%A1-%E9%97%AE%E7%AD%94-%E9%AA%8C%E8%AF%81%E9%9B%86.json) (for `dev`-subset) and [`质衡-问答-测试集.json`](https://huggingface.co/datasets/nanyangtu/LLVisionQA-QBench/resolve/main/%E8%B4%A8%E8%A1%A1-%E9%97%AE%E7%AD%94-%E6%B5%8B%E8%AF%95%E9%9B%86.json) (for `test`-subset). Put them under `./playground/data/eval/qbench`. 
+2. Download and extract [images](https://huggingface.co/datasets/nanyangtu/LLVisionQA-QBench/resolve/main/images_llvisionqa.tar) and put all the images directly under `./playground/data/eval/qbench/images_llviqionqa`.
+3. Single-GPU inference (change `dev` to `test` for evaluation on test set).
+```Shell
+CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/qbench_zh.sh dev
+```
+4. Submit the results by instruction [here](https://github.com/VQAssessment/Q-Bench#option-1-submit-results): `./playground/data/eval/qbench/llvisionqa_zh_dev_answers.jsonl`.

docs/Finetune_Custom_Data.md

@@ -0,0 +1,37 @@
+# Finetune LLaVA on Custom Datasets
+
+## Dataset Format
+
+Convert your data to a JSON file of a List of all samples. Sample metadata should contain `id` (a unique identifier), `image` (the path to the image), and `conversations` (the conversation data between human and AI).
+
+A sample JSON for finetuning LLaVA for generating tag-style captions for Stable Diffusion:
+
+```json
+[
+  {
+    "id": "997bb945-628d-4724-b370-b84de974a19f",
+    "image": "part-000001/997bb945-628d-4724-b370-b84de974a19f.jpg",
+    "conversations": [
+      {
+        "from": "human",
+        "value": "<image>\nWrite a prompt for Stable Diffusion to generate this image."
+      },
+      {
+        "from": "gpt",
+        "value": "a beautiful painting of chernobyl by nekro, pascal blanche, john harris, greg rutkowski, sin jong hun, moebius, simon stalenhag. in style of cg art. ray tracing. cel shading. hyper detailed. realistic. ue 5. maya. octane render. "
+      },
+    ]
+  },
+  ...
+]
+```
+
+## Command
+
+If you have a limited task-specific data, we recommend finetuning from LLaVA checkpoints with LoRA following this [script](https://github.com/haotian-liu/LLaVA/blob/main/scripts/v1_5/finetune_task_lora.sh).
+
+If the amount of the task-specific data is sufficient, you can also finetune from LLaVA checkpoints with full-model finetuning following this [script](https://github.com/haotian-liu/LLaVA/blob/main/scripts/v1_5/finetune_task.sh).
+
+You may need to adjust the hyperparameters to fit each specific dataset and your hardware constraint.
+
+

docs/Intel.md

@@ -0,0 +1,7 @@
+# Intel Platforms 
+
+* Support [Intel GPU Max Series](https://www.intel.com/content/www/us/en/products/details/discrete-gpus/data-center-gpu/max-series.html)    
+* Support [Intel CPU Sapphire Rapides](https://ark.intel.com/content/www/us/en/ark/products/codename/126212/products-formerly-sapphire-rapids.html)    
+* Based on [Intel Extension for Pytorch](https://intel.github.io/intel-extension-for-pytorch)    
+
+More details in  [**intel branch**](https://github.com/haotian-liu/LLaVA/tree/intel/docs/intel)

docs/LLaVA_Bench.md

@@ -0,0 +1,31 @@
+# LLaVA-Bench [[Download](https://huggingface.co/datasets/liuhaotian/llava-bench-in-the-wild)]
+
+**-Introduction-**  Large commercial multimodal chatbots have been released in this week, including 
+- [Multimodal Bing-Chat by Microsoft](https://blogs.bing.com/search/july-2023/Bing-Chat-Enterprise-announced,-multimodal-Visual-Search-rolling-out-to-Bing-Chat) (July 18, 2023) 
+- [Multimodal Bard by Google](https://bard.google.com/). 
+
+These chatbots are presumably supported by proprietary large multimodal models (LMM). Compared with the open-source LMM such as LLaVA, proprietary LMM represent the scaling success upperbound of the current SoTA techniques. They share the goal of developing multimodal chatbots that follow human intents to complete various daily-life visual tasks in the wild. While it remains less explored how to evaluate multimodal chat ability, it provides useful feedback to study open-source LMMs against the commercial multimodal chatbots. In addition to the *LLaVA-Bench (COCO)* dataset we used to develop the early versions of LLaVA, we are releasing  [*LLaVA-Bench (In-the-Wild)*](https://huggingface.co/datasets/liuhaotian/llava-bench-in-the-wild) to the community for the public use.
+
+## LLaVA-Bench (In-the-Wild *[Ongoing work]*)
+
+To evaluate the model's capability in more challenging tasks and generalizability to novel domains, we collect a diverse set of 24 images with 60 questions in total, including indoor and outdoor scenes, memes, paintings, sketches, etc, and associate each image with a highly-detailed and manually-curated description and a proper selection of questions. Such design also assesses the model's robustness to different prompts. In this release, we also categorize questions into three categories: conversation (simple QA), detailed description, and complex reasoning. We continue to expand and improve the diversity of the LLaVA-Bench (In-the-Wild).  We manually query Bing-Chat and Bard to get the responses. 
+
+### Results
+
+The score is measured by comparing against a reference answer generated by text-only GPT-4. It is generated by feeding the question, along with the ground truth image annotations as the context. A text-only GPT-4 evaluator rates both answers. We query GPT-4 by putting the reference answer first, and then the answer generated by the candidate model. We upload images at their original resolution to Bard and Bing-Chat to obtain the results.
+
+| Approach       | Conversation | Detail | Reasoning | Overall |
+|----------------|--------------|--------|-----------|---------|
+| Bard-0718      | 83.7         | 69.7   | 78.7      | 77.8    |
+| Bing-Chat-0629 | 59.6         | 52.2   | 90.1      | 71.5    |
+| LLaVA-13B-v1-336px-0719 (beam=1) | 64.3         | 55.9   | 81.7      | 70.1    |
+| LLaVA-13B-v1-336px-0719 (beam=5) | 68.4         | 59.9   | 84.3      | 73.5    |
+
+Note that Bard sometimes refuses to answer questions about images containing humans, and Bing-Chat blurs the human faces in the images. We also provide the benchmark score for the subset without humans.
+
+| Approach       | Conversation | Detail | Reasoning | Overall |
+|----------------|--------------|--------|-----------|---------|
+| Bard-0718      | 94.9         | 74.3   | 84.3      | 84.6    |
+| Bing-Chat-0629 | 55.8         | 53.6   | 93.5      | 72.6    |
+| LLaVA-13B-v1-336px-0719 (beam=1) | 62.2         | 56.4   | 82.2      | 70.0    |
+| LLaVA-13B-v1-336px-0719 (beam=5) | 65.6         | 61.7   | 85.0      | 73.6    |

docs/LLaVA_from_LLaMA2.md

@@ -0,0 +1,29 @@
+# LLaVA (based on Llama 2 LLM, Preview)
+
+*NOTE: This is a technical preview. We are still running hyperparameter search, and will release the final model soon.  If you'd like to contribute to this, please contact us.*
+
+:llama: **-Introduction-** [Llama 2 is an open-source LLM released by Meta AI](https://about.fb.com/news/2023/07/llama-2/) today (July 18, 2023). Compared with its early version [Llama 1](https://ai.meta.com/blog/large-language-model-llama-meta-ai/), Llama 2 is more favored in ***stronger language performance***, ***longer context window***, and importantly ***commercially usable***! While Llama 2 is changing the LLM market landscape in the language space, its multimodal ability remains unknown. We quickly develop the LLaVA variant based on the latest Llama 2 checkpoints, and release it to the community for the public use.
+
+You need to apply for and download the latest Llama 2 checkpoints to start your own training (apply [here](https://ai.meta.com/resources/models-and-libraries/llama-downloads/))
+
+
+## Training
+
+Please checkout [`pretrain.sh`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/pretrain.sh), [`finetune.sh`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/finetune.sh), [`finetune_lora.sh`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/finetune_lora.sh).
+
+## LLaVA (based on Llama 2), What is different? 
+
+:volcano: How is the new LLaVA based on Llama 2 different from Llama 1? The comparisons of the training process are described:
+- **Pre-training**. The pre-trained base LLM is changed from Llama 1 to Llama 2
+- **Language instruction-tuning**. The previous LLaVA model starts with Vicuna, which is instruct tuned on ShareGPT data from Llama 1; The new LLaVA model starts with Llama 2 Chat, which is an instruct tuned checkpoint on dialogue data from Llama 2.
+- **Multimodal instruction-tuning**. The same LLaVA-Lighting process is applied.
+
+
+### Results
+
+- Llama 2 is better at following the instructions of role playing; Llama 2 fails in following the instructions of translation
+- The quantitative evaluation on [LLaVA-Bench](https://github.com/haotian-liu/LLaVA/blob/main/docs/LLaVA_Bench.md) demonstrates on-par performance between Llama 2 and Llama 1 in LLaVA's multimodal chat ability.
+
+
+<img src="../images/llava_example_cmp.png" width="100%">
+

docs/LoRA.md

@@ -0,0 +1,46 @@
+# LLaVA (LoRA, Preview)
+
+NOTE: This is a technical preview, and is not yet ready for production use. We are still running hyperparameter search for the LoRA model, and will release the final model soon.  If you'd like to contribute to this, please contact us.
+
+You need latest code base for LoRA support (instructions [here](https://github.com/haotian-liu/LLaVA#upgrade-to-latest-code-base))
+
+## Demo (Web UI)
+
+Please execute each of the commands below one by one (after the previous one has finished).  The commands are the same as launching other demos except for an additional `--model-base` flag to specify the base model to use. Please make sure the base model corresponds to the LoRA checkpoint that you are using.  For this technical preview, you need Vicuna v1.1 (7B) checkpoint (if you do not have that already, follow the instructions [here](https://github.com/lm-sys/FastChat#vicuna-weights)).
+
+#### Launch a controller
+```Shell
+python -m llava.serve.controller --host 0.0.0.0 --port 10000
+```
+
+#### Launch a gradio web server.
+```Shell
+python -m llava.serve.gradio_web_server --controller http://localhost:10000 --model-list-mode reload
+```
+You just launched the Gradio web interface. Now, you can open the web interface with the URL printed on the screen. You may notice that there is no model in the model list. Do not worry, as we have not launched any model worker yet. It will be automatically updated when you launch a model worker.
+
+#### Launch a model worker
+```Shell
+python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path liuhaotian/llava-vicuna-7b-v1.1-lcs_558k-instruct_80k_3e-lora-preview-alpha --model-base /path/to/vicuna-v1.1
+```
+Wait until the process finishes loading the model and you see "Uvicorn running on ...".  Now, refresh your Gradio web UI, and you will see the model you just launched in the model list.
+
+You can launch as many workers as you want, and compare between different model checkpoints in the same Gradio interface. Please keep the `--controller` the same, and modify the `--port` and `--worker` to a different port number for each worker.
+
+
+## Training
+
+Please see sample training scripts for [LoRA](https://github.com/haotian-liu/LLaVA/blob/main/scripts/finetune_lora.sh) and [QLoRA](https://github.com/haotian-liu/LLaVA/blob/main/scripts/finetune_qlora.sh).
+
+We provide sample DeepSpeed configs, [`zero3.json`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/zero3.json) is more like PyTorch FSDP, and [`zero3_offload.json`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/zero3_offload.json) can further save memory consumption by offloading parameters to CPU. `zero3.json` is usually faster than `zero3_offload.json` but requires more GPU memory, therefore, we recommend trying `zero3.json` first, and if you run out of GPU memory, try `zero3_offload.json`. You can also tweak the `per_device_train_batch_size` and `gradient_accumulation_steps` in the config to save memory, and just to make sure that `per_device_train_batch_size` and `gradient_accumulation_steps` remains the same.
+
+If you are having issues with ZeRO-3 configs, and there are enough VRAM, you may try [`zero2.json`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/zero2.json). This consumes slightly more memory than ZeRO-3, and behaves more similar to PyTorch FSDP, while still supporting parameter-efficient tuning.
+
+## Create Merged Checkpoints
+
+```Shell
+python scripts/merge_lora_weights.py \
+    --model-path /path/to/lora_model \
+    --model-base /path/to/base_model \
+    --save-model-path /path/to/merge_model
+```

docs/MODEL_ZOO.md

@@ -0,0 +1,150 @@
+# Model Zoo
+
+**To Use LLaVA-1.6 checkpoints, your llava package version must be newer than 1.2.0. [Instructions](https://github.com/haotian-liu/LLaVA#upgrade-to-latest-code-base) on how to upgrade.**
+
+If you are interested in including any other details in Model Zoo, please open an issue :)
+
+The model weights below are *merged* weights. You do not need to apply delta. The usage of LLaVA checkpoints should comply with the base LLM's model license.
+
+## LLaVA-v1.6
+
+| Version | LLM | Schedule | Checkpoint | MMMU | MathVista | VQAv2 | GQA | VizWiz | SQA | TextVQA | POPE | MME | MM-Bench | MM-Bench-CN | SEED-IMG | LLaVA-Bench-Wild | MM-Vet |
+|----------|----------|-----------|-----------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+| LLaVA-1.6 | Vicuna-7B | full_ft-1e | [liuhaotian/llava-v1.6-vicuna-7b](https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b) | 35.8 | 34.6 | 81.8 | 64.2 | 57.6 | 70.1 | 64.9 | 86.5 | 1519/332 | 67.4 | 60.6 | 70.2 | 81.6 | 43.9 |
+| LLaVA-1.6 | Vicuna-13B | full_ft-1e | [liuhaotian/llava-v1.6-vicuna-13b](https://huggingface.co/liuhaotian/llava-v1.6-vicuna-13b) | 36.2 | 35.3 | 82.8 | 65.4 | 60.5 | 73.6 | 67.1 | 86.2 | 1575/326 | 70 | 64.4 | 71.9 | 87.3 | 48.4 |
+| LLaVA-1.6 | Mistral-7B | full_ft-1e | [liuhaotian/llava-v1.6-mistral-7b](https://huggingface.co/liuhaotian/llava-v1.6-mistral-7b) | 35.3 | 37.7 | 82.2 | 64.8 | 60.0 | 72.8 | 65.7 | 86.7 | 1498/321 | 68.7 | 61.2 | 72.2 | 83.2 | 47.3 |
+| LLaVA-1.6 | Hermes-Yi-34B | full_ft-1e | [liuhaotian/llava-v1.6-34b](https://huggingface.co/liuhaotian/llava-v1.6-34b) | 51.1 | 46.5 | 83.7 | 67.1 | 63.8 | 81.8 | 69.5 | 87.7 | 1631/397 | 79.3 | 79 | 75.9 | 89.6 | 57.4 |
+
+*LLaVA-1.6-34B outperforms Gemini Pro on benchmarks like MMMU and MathVista.*
+
+
+## LLaVA-v1.5
+
+| Version | Size | Schedule | Checkpoint | VQAv2 | GQA | VizWiz | SQA | TextVQA | POPE | MME | MM-Bench | MM-Bench-CN | SEED | LLaVA-Bench-Wild | MM-Vet |
+|----------|----------|-----------|-----------|---|---|---|---|---|---|---|---|---|---|---|---|
+| LLaVA-1.5 | 7B | full_ft-1e | [liuhaotian/llava-v1.5-7b](https://huggingface.co/liuhaotian/llava-v1.5-7b) | 78.5 | 62.0 | 50.0 | 66.8 | 58.2 | 85.9 | 1510.7 | 64.3 | 58.3 | 58.6 | 65.4 | 31.1 |
+| LLaVA-1.5 | 13B | full_ft-1e | [liuhaotian/llava-v1.5-13b](https://huggingface.co/liuhaotian/llava-v1.5-13b) | 80.0 | 63.3 | 53.6 | 71.6 | 61.3 | 85.9 | 1531.3 | 67.7 | 63.6 | 61.6 | 72.5 | 36.1 |
+| LLaVA-1.5 | 7B | lora-1e | [liuhaotian/llava-v1.5-7b-lora](https://huggingface.co/liuhaotian/llava-v1.5-7b-lora) | 79.1 | 63.0 | 47.8 | 68.4 | 58.2 | 86.4 | 1476.9 | 66.1 | 58.9 | 60.1 | 67.9 | 30.2 |
+| LLaVA-1.5 | 13B | lora-1e | [liuhaotian/llava-v1.5-13b-lora](https://huggingface.co/liuhaotian/llava-v1.5-13b-lora) | 80.0 | 63.3 | 58.9 | 71.2 | 60.2 | 86.7 | 1541.7 | 68.5 | 61.5 | 61.3 | 69.5 | 38.3 |
+
+Base model: Vicuna v1.5. Training logs: [wandb](https://api.wandb.ai/links/lht/6orh56wc).
+
+<p align="center">
+  <img src="../images/llava_v1_5_radar.jpg" width="500px"> <br>
+  LLaVA-1.5 achieves SoTA performance across 11 benchmarks.
+</p>
+
+
+## LLaVA-v1
+
+*Note: We recommend using the most capable LLaVA-v1.5 series above for the best performance.*
+
+| Base LLM | Vision Encoder | Pretrain Data | Pretraining schedule | Finetuning Data | Finetuning schedule | LLaVA-Bench-Conv | LLaVA-Bench-Detail | LLaVA-Bench-Complex | LLaVA-Bench-Overall | Download |
+|----------|----------------|---------------|----------------------|-----------------|--------------------|------------------|--------------------|---------------------|---------------------|---------------------|
+| Vicuna-13B-v1.3 | CLIP-L-336px | LCS-558K | 1e | LLaVA-Instruct-80K | proj-1e, lora-1e | 64.3 | 55.9 | 81.7 | 70.1 | [LoRA](https://huggingface.co/liuhaotian/llava-v1-0719-336px-lora-vicuna-13b-v1.3) [LoRA-Merged](https://huggingface.co/liuhaotian/llava-v1-0719-336px-lora-merge-vicuna-13b-v1.3) |
+| LLaMA-2-13B-Chat | CLIP-L | LCS-558K | 1e | LLaVA-Instruct-80K | full_ft-1e | 56.7 | 58.6 | 80.0 | 67.9 | [ckpt](https://huggingface.co/liuhaotian/llava-llama-2-13b-chat-lightning-preview) |
+| LLaMA-2-7B-Chat | CLIP-L | LCS-558K | 1e | LLaVA-Instruct-80K | lora-1e | 51.2 | 58.9 | 71.6 | 62.8 | [LoRA](https://huggingface.co/liuhaotian/llava-llama-2-7b-chat-lightning-lora-preview) |
+
+
+## Projector weights
+
+These are projector weights we have pretrained. You can use these projector weights for visual instruction tuning. They are just pretrained on image-text pairs and are NOT instruction-tuned, which means they do NOT follow instructions as well as our official models and can output repetitive, lengthy, and garbled outputs. If you want to have nice conversations with LLaVA, use the checkpoints above (LLaVA v1.5).
+
+NOTE: These projector weights are only compatible with `llava>=1.0.0`. Please check out the latest codebase if your local code version is below v1.0.0.
+
+NOTE: When you use our pretrained projector for visual instruction tuning, it is very important to use the same base LLM and vision encoder as the one we used for pretraining the projector. Otherwise, the performance will be very poor.
+
+When using these projector weights to instruction-tune your LMM, please make sure that these options are correctly set as follows,
+
+```Shell
+--mm_use_start_end False
+--mm_use_patch_token False
+```
+
+| Base LLM | Vision Encoder | Projection | Pretrain Data | Pretraining schedule | Download |
+|----------|----------------|---------------|----------------------|----------|----------|
+| Vicuna-13B-v1.5 | CLIP-L-336px | MLP-2x | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/llava-v1.5-mlp2x-336px-pretrain-vicuna-13b-v1.5) |
+| Vicuna-7B-v1.5 | CLIP-L-336px | MLP-2x | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/llava-v1.5-mlp2x-336px-pretrain-vicuna-7b-v1.5) |
+| LLaMA-2-13B-Chat | CLIP-L-336px | Linear | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/llava-336px-pretrain-llama-2-13b-chat) |
+| LLaMA-2-7B-Chat | CLIP-L-336px | Linear | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/llava-336px-pretrain-llama-2-7b-chat) |
+| LLaMA-2-13B-Chat | CLIP-L | Linear | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/llava-pretrain-llama-2-13b-chat) |
+| LLaMA-2-7B-Chat | CLIP-L | Linear | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/llava-pretrain-llama-2-7b-chat) |
+| Vicuna-13B-v1.3 | CLIP-L-336px | Linear | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/llava-336px-pretrain-vicuna-13b-v1.3) |
+| Vicuna-7B-v1.3 | CLIP-L-336px | Linear | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/llava-336px-pretrain-vicuna-7b-v1.3) |
+| Vicuna-13B-v1.3 | CLIP-L | Linear | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/llava-pretrain-vicuna-13b-v1.3) |
+| Vicuna-7B-v1.3 | CLIP-L | Linear | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/llava-pretrain-vicuna-7b-v1.3) |
+
+
+## Science QA Checkpoints
+
+| Base LLM | Vision Encoder | Pretrain Data | Pretraining schedule | Finetuning Data | Finetuning schedule | Download |
+|----------|----------------|---------------|----------------------|-----------------|--------------------|---------------------|
+| Vicuna-13B-v1.3 | CLIP-L | LCS-558K | 1e | ScienceQA | full_ft-12e | [ckpt](https://huggingface.co/liuhaotian/llava-lcs558k-scienceqa-vicuna-13b-v1.3) |
+
+
+## Legacy Models (merged weights)
+
+The model weights below are *merged* weights. You do not need to apply delta. The usage of LLaVA checkpoints should comply with the base LLM's model license.
+
+| Base LLM | Vision Encoder | Pretrain Data | Pretraining schedule | Finetuning Data | Finetuning schedule | Download |
+|----------|----------------|---------------|----------------------|-----------------|--------------------|------------------|
+| MPT-7B-Chat | CLIP-L | LCS-558K | 1e | LLaVA-Instruct-80K | full_ft-1e | [preview](https://huggingface.co/liuhaotian/LLaVA-Lightning-MPT-7B-preview) |
+
+
+## Legacy Models (delta weights)
+
+The model weights below are *delta* weights. The usage of LLaVA checkpoints should comply with the base LLM's model license: [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md).
+
+You can add our delta to the original LLaMA weights to obtain the LLaVA weights.
+
+Instructions:
+
+1. Get the original LLaMA weights in the huggingface format by following the instructions [here](https://huggingface.co/docs/transformers/main/model_doc/llama).
+2. Use the following scripts to get LLaVA weights by applying our delta. It will automatically download delta weights from our Hugging Face account. In the script below, we use the delta weights of [`liuhaotian/LLaVA-7b-delta-v0`](https://huggingface.co/liuhaotian/LLaVA-7b-delta-v0) as an example. It can be adapted for other delta weights by changing the `--delta` argument (and base/target accordingly).
+
+```bash
+python3 -m llava.model.apply_delta \
+    --base /path/to/llama-7b \
+    --target /output/path/to/LLaVA-7B-v0 \
+    --delta liuhaotian/LLaVA-7b-delta-v0
+```
+
+| Base LLM | Vision Encoder | Pretrain Data | Pretraining schedule | Finetuning Data | Finetuning schedule | Download |
+|----------|----------------|---------------|----------------------|-----------------|--------------------|------------------|
+| Vicuna-13B-v1.1 | CLIP-L | CC-595K | 1e | LLaVA-Instruct-158K | full_ft-3e | [delta-weights](https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1) |
+| Vicuna-7B-v1.1 | CLIP-L | LCS-558K | 1e | LLaVA-Instruct-80K | full_ft-1e | [delta-weights](https://huggingface.co/liuhaotian/LLaVA-Lightning-7B-delta-v1-1) |
+| Vicuna-13B-v0 | CLIP-L | CC-595K | 1e | LLaVA-Instruct-158K | full_ft-3e | [delta-weights](https://huggingface.co/liuhaotian/LLaVA-13b-delta-v0) |
+| Vicuna-13B-v0 | CLIP-L | CC-595K | 1e | ScienceQA | full_ft-12e | [delta-weights](https://huggingface.co/liuhaotian/LLaVA-13b-delta-v0-science_qa) |
+| Vicuna-7B-v0 | CLIP-L | CC-595K | 1e | LLaVA-Instruct-158K | full_ft-3e | [delta-weights](https://huggingface.co/liuhaotian/LLaVA-7b-delta-v0) |
+
+
+
+## Legacy Projector weights
+
+The following projector weights are deprecated, and the support for them may be removed in the future. They do not support zero-shot inference. Please use the projector weights in the [table above](#projector-weights) if possible.
+
+**NOTE**: When you use our pretrained projector for visual instruction tuning, it is very important to **use the same base LLM and vision encoder** as the one we used for pretraining the projector. Otherwise, the performance will be very bad.
+
+When using these projector weights to instruction tune your LMM, please make sure that these options are correctly set as follows,
+
+```Shell
+--mm_use_start_end True
+--mm_use_patch_token False
+```
+
+| Base LLM | Vision Encoder | Pretrain Data | Pretraining schedule | Download |
+|----------|----------------|---------------|----------------------|----------|
+| Vicuna-7B-v1.1 | CLIP-L | LCS-558K | 1e | [projector](https://huggingface.co/liuhaotian/LLaVA-Pretrained-Projectors/blob/main/LLaVA-7b-pretrain-projector-v1-1-LCS-558K-blip_caption.bin) |
+| Vicuna-13B-v0 | CLIP-L | CC-595K | 1e | [projector](https://huggingface.co/liuhaotian/LLaVA-Pretrained-Projectors/blob/main/LLaVA-13b-pretrain-projector-v0-CC3M-595K-original_caption.bin) |
+| Vicuna-7B-v0 | CLIP-L | CC-595K | 1e | [projector](https://huggingface.co/liuhaotian/LLaVA-Pretrained-Projectors/blob/main/LLaVA-7b-pretrain-projector-v0-CC3M-595K-original_caption.bin) |
+
+When using these projector weights to instruction tune your LMM, please make sure that these options are correctly set as follows,
+
+```Shell
+--mm_use_start_end False
+--mm_use_patch_token False
+```
+
+| Base LLM | Vision Encoder | Pretrain Data | Pretraining schedule | Download |
+|----------|----------------|---------------|----------------------|----------|
+| Vicuna-13B-v0 | CLIP-L | CC-595K | 1e | [projector](https://huggingface.co/liuhaotian/LLaVA-Pretrained-Projectors/blob/main/LLaVA-13b-pretrain-projector-v0-CC3M-595K-original_caption-no_im_token.bin) |

docs/ScienceQA.md

@@ -0,0 +1,53 @@
+### ScienceQA
+
+#### Prepare Data
+1. Please see ScienceQA [repo](https://github.com/lupantech/ScienceQA) for setting up the dataset.
+2. Generate ScienceQA dataset for LLaVA conversation-style format.
+
+```Shell
+python scripts/convert_sqa_to_llava.py \
+    convert_to_llava \
+    --base-dir /path/to/ScienceQA/data/scienceqa \
+    --prompt-format "QCM-LEA" \
+    --split {train,val,minival,test,minitest}
+```
+
+#### Training
+
+1. Pretraining
+
+You can download our pretrained projector weights from our [Model Zoo](), or train your own projector weights using [`pretrain.sh`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/pretrain.sh).
+
+2. Finetuning
+
+See [`finetune_sqa.sh`](https://github.com/haotian-liu/LLaVA/blob/main/scripts/finetune_sqa.sh).
+
+#### Evaluation
+
+1. Multiple-GPU inference
+You may evaluate this with multiple GPUs, and concatenate the generated jsonl files.  Please refer to our script for [batch evaluation](https://github.com/haotian-liu/LLaVA/blob/main/scripts/sqa_eval_batch.sh) and [results gathering](https://github.com/haotian-liu/LLaVA/blob/main/scripts/sqa_eval_gather.sh).
+
+2. Single-GPU inference
+
+(a) Generate LLaVA responses on ScienceQA dataset
+
+```Shell
+python -m llava.eval.model_vqa_science \
+    --model-path liuhaotian/llava-lcs558k-scienceqa-vicuna-13b-v1.3 \
+    --question-file /path/to/ScienceQA/data/scienceqa/llava_test_QCM-LEA.json \
+    --image-folder /path/to/ScienceQA/data/scienceqa/images/test \
+    --answers-file vqa/results/ScienceQA/test_llava-13b.jsonl \
+    --conv-mode llava_v1
+```
+
+(b) Evaluate the generated responses
+
+```Shell
+python eval_science_qa.py \
+    --base-dir /path/to/ScienceQA/data/scienceqa \
+    --result-file vqa/results/ScienceQA/test_llava-13b.jsonl \
+    --output-file vqa/results/ScienceQA/test_llava-13b_output.json \
+    --output-result vqa/results/ScienceQA/test_llava-13b_result.json \
+```
+
+For reference, we attach our prediction file [`test_sqa_llava_lcs_558k_sqa_12e_vicuna_v1_3_13b.json`](https://github.com/haotian-liu/LLaVA/blob/main/llava/eval/table/results/test_sqa_llava_lcs_558k_sqa_12e_vicuna_v1_3_13b.json) and [`test_sqa_llava_13b_v0.json`](https://github.com/haotian-liu/LLaVA/blob/main/llava/eval/table/results/test_sqa_llava_13b_v0.json) for comparison when reproducing our results, as well as for further analysis in detail.

docs/Windows.md

@@ -0,0 +1,27 @@
+# Run LLaVA on Windows
+
+*NOTE: LLaVA on Windows is not fully supported. Currently we only support 16-bit inference. For a more complete support, please use [WSL2](https://learn.microsoft.com/en-us/windows/wsl/install) for now. More functionalities on Windows is to be added soon, stay tuned.*
+
+## Installation
+
+1. Clone this repository and navigate to LLaVA folder
+```bash
+git clone https://github.com/haotian-liu/LLaVA.git
+cd LLaVA
+```
+
+2. Install Package
+```Shell
+conda create -n llava python=3.10 -y
+conda activate llava
+python -mpip install --upgrade pip  # enable PEP 660 support
+pip install torch==2.0.1+cu117 torchvision==0.15.2+cu117 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu117
+pip install -e .
+pip uninstall bitsandbytes
+```
+
+## Run demo
+
+See instructions [here](https://github.com/haotian-liu/LLaVA#demo).
+
+Note that quantization (4-bit, 8-bit) is *NOT* supported on Windows. Stay tuned for the 4-bit support on Windows!

docs/macOS.md

@@ -0,0 +1,29 @@
+# Run LLaVA on macOS
+
+*NOTE: LLaVA on macOS is not fully supported. Currently we only support 16-bit inference. More functionalities on macOS is to be added soon, stay tuned.*
+
+## Installation
+
+1. Clone this repository and navigate to LLaVA folder
+```bash
+git clone https://github.com/haotian-liu/LLaVA.git
+cd LLaVA
+```
+
+2. Install Package
+```Shell
+conda create -n llava python=3.10 -y
+conda activate llava
+python -mpip install --upgrade pip  # enable PEP 660 support
+pip install -e .
+pip install torch==2.1.0 torchvision==0.16.0
+pip uninstall bitsandbytes
+```
+
+## Run demo
+
+Specify `--device mps` when launching model worker or CLI.
+
+See instructions [here](https://github.com/haotian-liu/LLaVA#demo).
+
+Note that quantization (4-bit, 8-bit) is *NOT* supported on macOS. Stay tuned for the 4-bit support on macOS!

flash_attn-2.3.6+cu122torch2.1cxx11abiFALSE-cp39-cp39-linux_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6eed754c1d79b6ec4a695a6125107cf8dba714fc7cb32d093bbab98f3d7171f6
+size 56476853

handler.py

@@ -0,0 +1,276 @@
+import sys
+import torch
+import os
+import random
+import base64
+import msgpack
+from io import BytesIO
+import numpy as np
+
+from transformers import AutoTokenizer
+from llava.constants import MM_TOKEN_INDEX, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN, DEFAULT_VIDEO_TOKEN, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava.conversation import conv_templates, SeparatorStyle
+from llava.utils import disable_torch_init
+from llava.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria, process_images_v2
+from llava.model.builder import load_pretrained_model
+from llava.model.multimodal_encoder.processor import Blip2ImageTrainProcessor
+from llava.model import LlavaMistralForCausalLM
+
+
+from transformers import CLIPImageProcessor
+from PIL import Image
+import logging
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+import threading
+
+def select_frames(input_frames, num_segments = 10):
+
+    indices = np.linspace(start=0, stop=len(input_frames)-1, num=num_segments).astype(int)
+
+    frames = [input_frames[ind] for ind in indices]
+
+    return frames
+
+def load_model(model_path, device_map):
+    kwargs = {"device_map": device_map}
+    kwargs['torch_dtype'] = torch.float16 #difference with cpu handler but it needs float16 to ensure no memory issue
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    model = LlavaMistralForCausalLM.from_pretrained(
+        model_path,
+        low_cpu_mem_usage=True,
+        **kwargs
+    )
+    tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN], special_tokens=True)
+    model.resize_token_embeddings(len(tokenizer))
+
+    vision_tower = model.get_vision_tower()
+    if not vision_tower.is_loaded:
+        vision_tower.load_model(device_map=device_map)
+
+    return model, tokenizer
+
+
+class EndpointHandler:
+
+    def __init__(self):
+        model_path = './masp_094_v2'
+        disable_torch_init()
+        model_path = os.path.expanduser(model_path)
+        #print(model_path)
+        model_name = get_model_name_from_path(model_path)
+
+        model, tokenizer = load_model(model_path, device_map={"":0})
+
+        image_processor = Blip2ImageTrainProcessor(
+            image_size=model.config.img_size,
+            is_training=False)
+
+        """
+        import os
+        from PIL import Image
+        input_dir = './v12044gd0000clg1n4fog65p7pag5n6g/video'
+        image_paths = os.listdir(input_dir)
+        images = [Image.open(os.path.join(input_dir, item)) for item in image_paths]
+        num_segments = 10
+        images = images[:num_segments]
+
+        import torch
+        device = torch.device('cuda:0')
+        image_processor = Blip2ImageTrainProcessor(
+            image_size=224,
+            is_training=False)     
+        images_tensor = [image_processor.preprocess(image).cpu().to(device) for image in images]
+        """
+
+        self.tokenizer = tokenizer
+        self.device = torch.device('cuda:0') #another difference here
+        self.model = model.to(self.device)
+
+        self.image_processor = image_processor
+        self.conv_mode = 'v1'
+
+    def inference_frames_batch(self, batch_image_lists, batch_prompts, batch_temperatures):
+        start_time = time.perf_counter()  # Start timer
+
+        batch_size = len(batch_image_lists)
+
+        # Process images and prompts for each item in the batch
+        images_tensors_list = []
+        input_ids_list = []
+        for images, prompt in zip(batch_image_lists, batch_prompts):
+            # Select frames (ensure consistent number of frames)
+            if len(images) > 10:
+                images = select_frames(images)
+            if len(images) < 10:
+                images += [images[-1]] * (10 - len(images))  # Pad to 10 frames
+
+            # Process images
+            images_tensor = process_images_v2(images, self.image_processor, self.model.config)
+            images_tensor = images_tensor.half().to(self.device)  # Ensure correct dtype and device
+            images_tensors_list.append(images_tensor)
+
+            # Prepare the prompt
+            if len(images) == 1:
+                qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + prompt
+            else:
+                qs = DEFAULT_VIDEO_START_TOKEN + DEFAULT_VIDEO_TOKEN + DEFAULT_VIDEO_END_TOKEN + '\n' + prompt
+
+            # Build conversation and tokenize
+            conv = conv_templates[self.conv_mode].copy()
+            conv.append_message(conv.roles[0], qs)
+            conv.append_message(conv.roles[1], None)
+            prompt_text = conv.get_prompt()
+
+            input_ids = tokenizer_image_token(prompt_text, self.tokenizer, MM_TOKEN_INDEX, return_tensors='pt').squeeze(0)
+            input_ids_list.append(input_ids)
+
+        # Pad input IDs to the same length
+        input_ids_padded = torch.nn.utils.rnn.pad_sequence(
+            input_ids_list,
+            batch_first=True,
+            padding_value=self.tokenizer.pad_token_id
+        ).to(self.device)
+
+        # No need to stack images_tensors_list into a tensor
+        # Each item in images_tensors_list is a tensor of shape (num_frames, C, H, W)
+
+        # Prepare stopping criteria
+        conv = conv_templates[self.conv_mode].copy()
+        stop_str = conv.sep if conv.sep2 is None else conv.sep2
+        keywords = [stop_str]
+        stopping_criteria = KeywordsStoppingCriteria(keywords, self.tokenizer, input_ids_padded)
+
+        # Use the first temperature for simplicity
+        temperature = batch_temperatures[0]
+
+        # Perform model inference
+        with torch.inference_mode():
+            output_ids = self.model.generate(
+                input_ids_padded,
+                images=images_tensors_list,
+                temperature=temperature,
+                do_sample=True,
+                top_p=None,
+                num_beams=1,
+                no_repeat_ngram_size=3,
+                max_new_tokens=1024,
+                use_cache=True,
+                stopping_criteria=[stopping_criteria],
+            )
+
+        # Decode outputs
+        outputs = []
+        for output_id in output_ids:
+            output = self.tokenizer.decode(output_id, skip_special_tokens=True).strip()
+            output = output.rstrip(stop_str).strip()
+            outputs.append(output)
+
+        end_time = time.perf_counter()  # End timer
+        latency = end_time - start_time
+        print(f"Latency for this batch inference: {latency:.4f} seconds")
+
+        return outputs
+
+    def __call__(self, request):
+        
+        # Unpack the images and prompts
+        packed_data_list = request['images']  # List of packed image data
+        prompt_list = request.get('prompt', [''.encode()] * len(packed_data_list))
+        temperature_list = request.get('temperature', ['0.01'.encode()] * len(packed_data_list))
+
+        # Initialize lists to collect images, prompts, and temperatures
+        all_image_lists = []  # List of lists of images
+        all_prompts = []
+        all_temperatures = []
+
+        for packed_data, prompt_encoded, temperature_encoded in zip(packed_data_list, prompt_list, temperature_list):
+            # Unpack the images
+            unpacked_data = msgpack.unpackb(packed_data, raw=False)
+            image_list = [Image.open(BytesIO(byte_data)).convert('RGB') for byte_data in unpacked_data]
+            all_image_lists.append(image_list)
+
+            # Decode the prompt
+            prompt = prompt_encoded.decode()
+            if prompt == '':
+                if len(image_list) == 1:
+                    prompt = "Please describe this image in detail."
+                else:
+                    prompt = "Describe the following video in detail."
+            all_prompts.append(prompt)
+
+            # Decode the temperature
+            temperature = float(temperature_encoded.decode())
+            all_temperatures.append(temperature)
+
+        # Now process all_image_lists and all_prompts in batch
+        with torch.no_grad():
+            outputs = self.inference_frames_batch(all_image_lists, all_prompts, all_temperatures)
+
+        return {'output': outputs}
+
+def benchmark_qps_batched(handler, batched_request, num_batches=10):
+    start_time = time.perf_counter()
+    completed_samples = 0
+
+    for _ in range(num_batches):
+        handler(batched_request)
+        completed_samples += len(batched_request['images'])
+
+    end_time = time.perf_counter()
+    total_time = end_time - start_time
+    qps = completed_samples / total_time
+    print(f"Processed {completed_samples} samples in {total_time:.2f} seconds. QPS: {qps:.2f}")
+
+if __name__ == "__main__":
+    # 7347652962333773061
+    video_dir = './v12044gd0000cl5c6rfog65i2eoqcqig'
+    #video_dir = '/mnt/bn/data-tns-algo-masp/kaili.zhao/data/masp_data/train/human_annotation/video_frames_2fps/7347652962333773061'
+    frames = [(int(os.path.splitext(item)[0]), os.path.join(video_dir, item)) for item in os.listdir(video_dir)]
+    frames = [item[1] for item in sorted(frames, key=lambda x: x[0])]
+    out_frames = [Image.open(frame).convert('RGB') for frame in frames]
+
+    # out_frames = select_frames(frames)
+
+    # Number of samples to include in the batch
+    batch_size = 4  # Adjust based on GPU memory
+
+    # Prepare batched data
+    batched_packed_data = []
+    batched_prompts = []
+    batched_temperatures = []
+
+    for _ in range(batch_size):
+        # Convert images to byte format
+        byte_images = []
+        for img in out_frames:
+            byte_io = BytesIO()
+            img.save(byte_io, format='JPEG')
+            byte_images.append(byte_io.getvalue())
+
+        # Pack the byte data with msgpack
+        packed_data = msgpack.packb(byte_images)
+        batched_packed_data.append(packed_data)
+        
+        # Add prompt and temperature for each sample
+        batched_prompts.append(''.encode())  # Or specific prompts
+        batched_temperatures.append('0.01'.encode())
+
+    # Create the batched request
+    batched_request = {
+        'images': batched_packed_data,
+        'prompt': batched_prompts,
+        'temperature': batched_temperatures,
+    }
+
+    handler = EndpointHandler()
+
+    # Measure latency for the batched request
+    #print("\nMeasuring latency for batched request...")
+    response = handler(batched_request)
+    print(response)#['output'])
+
+    # Benchmark QPS with batched requests
+    # print("\nBenchmarking QPS with batched requests...")
+    # num_batches = 10  # Number of batched requests
+    # benchmark_qps_batched(handler, batched_request, num_batches=num_batches)

handler_cpu.py

@@ -0,0 +1,209 @@
+import sys
+import torch
+import os
+import random
+import base64
+import msgpack
+from io import BytesIO
+import numpy as np
+
+from transformers import AutoTokenizer
+from llava.constants import MM_TOKEN_INDEX, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN, DEFAULT_VIDEO_TOKEN, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava.conversation import conv_templates, SeparatorStyle
+from llava.utils import disable_torch_init
+from llava.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria, process_images_v2
+from llava.model.builder import load_pretrained_model
+from llava.model.multimodal_encoder.processor import Blip2ImageTrainProcessor
+from llava.model import LlavaMistralForCausalLM
+
+
+from transformers import CLIPImageProcessor
+from PIL import Image
+import logging
+
+def select_frames(input_frames, num_segments = 10):
+
+    indices = np.linspace(start=0, stop=len(input_frames)-1, num=num_segments).astype(int)
+
+    frames = [input_frames[ind] for ind in indices]
+
+    return frames
+
+def load_model(model_path, device_map):
+    kwargs = {"device_map": device_map}
+    kwargs['torch_dtype'] = torch.float32
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    model = LlavaMistralForCausalLM.from_pretrained(
+        model_path,
+        low_cpu_mem_usage=True,
+        **kwargs
+    )
+    tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN], special_tokens=True)
+    model.resize_token_embeddings(len(tokenizer))
+
+    vision_tower = model.get_vision_tower()
+    if not vision_tower.is_loaded:
+        vision_tower.load_model(device_map=device_map)
+
+    return model, tokenizer
+
+
+class EndpointHandler:
+
+    def __init__(self):
+        model_path = './checkpoint-3000'
+        disable_torch_init()
+        model_path = os.path.expanduser(model_path)
+        #print(model_path)
+        model_name = get_model_name_from_path(model_path)
+
+        model, tokenizer = load_model(model_path, device_map={"":0})
+
+        #tokenizer, model, _, context_len = load_pretrained_model(model_path, None, model_name, device_map={"":0})
+        image_processor = Blip2ImageTrainProcessor(
+            image_size=model.config.img_size,
+            is_training=False)
+
+        """
+        import os
+        from PIL import Image
+        input_dir = './v12044gd0000clg1n4fog65p7pag5n6g/video'
+        image_paths = os.listdir(input_dir)
+        images = [Image.open(os.path.join(input_dir, item)) for item in image_paths]
+        num_segments = 10
+        images = images[:num_segments]
+
+        import torch
+        device = torch.device('cuda:0')
+        image_processor = Blip2ImageTrainProcessor(
+            image_size=224,
+            is_training=False)     
+        images_tensor = [image_processor.preprocess(image).cpu().to(device) for image in images]
+        """
+
+        self.tokenizer = tokenizer
+        self.device = torch.device('cpu')
+        self.model = model.to(self.device)
+
+        self.image_processor = image_processor
+        self.conv_mode = 'v1'
+
+    def inference_frames(self, images, question, temperature):
+        
+        if len(images) > 10:
+            images = select_frames(images)
+
+        conv_mode = self.conv_mode
+        image_processor = self.image_processor
+        # if isinstance(image_processor, CLIPImageProcessor):
+        #     images_tensor = [image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0].cpu().to(self.device) for image in images]
+        # else:
+        #     logging.info(f'length of images:{len(images)}')
+        #images_tensor = [image_processor.preprocess(image).cpu() for image in images]
+        #images_tensor = torch.stack(images_tensor, dim=0).half().to(self.device)
+
+        images_tensor = process_images_v2(images, image_processor, self.model.config)
+        images_tensor = images_tensor.to(self.device)
+        # print(images_tensor.shape)
+
+        qs = question
+
+        if len(images) == 1:
+            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+        else:
+            qs = DEFAULT_VIDEO_START_TOKEN + DEFAULT_VIDEO_TOKEN + DEFAULT_VIDEO_END_TOKEN + '\n' + qs
+
+        conv = conv_templates[conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        input_ids = tokenizer_image_token(prompt, self.tokenizer, MM_TOKEN_INDEX, return_tensors='pt').unsqueeze(
+            0).to(self.device)
+
+        stop_str = conv.sep if conv.sep2 is None else conv.sep2
+        keywords = [stop_str]
+        stopping_criteria = KeywordsStoppingCriteria(keywords, self.tokenizer, input_ids)
+
+        with torch.inference_mode():
+            output_ids = self.model.generate(
+                input_ids,
+                images=[images_tensor],
+                temperature=temperature,
+                do_sample=True,
+                top_p=None,
+                num_beams=1,
+                no_repeat_ngram_size=3,
+                max_new_tokens=1024,
+                use_cache=True,
+                stopping_criteria=[stopping_criteria],
+                )
+
+
+        outputs = self.tokenizer.decode(output_ids[0], skip_special_tokens=True).strip()
+
+        outputs = outputs.strip()
+        if outputs.endswith(conv.sep):
+            outputs = outputs[:-len(stop_str)]
+        outputs = outputs.strip()
+
+        
+        # outputs = outputs[3:-4].strip()
+
+        return outputs
+
+    def __call__(self, request):
+        
+        # Step 5: Unpack the data and convert back to PIL images
+        packed_data= request['images'][0]
+        unpacked_data = msgpack.unpackb(packed_data, raw=False)
+        image_list = [Image.open(BytesIO(byte_data)) for byte_data in unpacked_data]
+        prompt = request.get('prompt', [''.encode()])[0].decode()
+        temperature = request.get('temperature', ['0.01'.encode()])[0].decode()
+        temperature = float(temperature)
+
+        #print(request)
+
+        if prompt=='':
+            if len(image_list) == 1:
+                prompt =  "Please describe this image in detail."
+            else:
+                prompt =  "Please describe this video in detail."            
+                # prompt =  "Describe the following video in detail."            
+            
+        with torch.no_grad():
+            outputs = self.inference_frames(image_list, prompt, temperature)
+
+
+        return {'output': [outputs]}
+
+
+if __name__ == "__main__":
+    video_dir = '/mnt/bn/yukunfeng-nasdrive/xiangchen/masp_data/20231110_ttp/video/v12044gd0000cl5c6rfog65i2eoqcqig'
+    frames = [(int(os.path.splitext(item)[0]), os.path.join(video_dir, item)) for item in os.listdir(video_dir)]
+    frames = [item[1] for item in sorted(frames, key=lambda x: x[0])]
+    out_frames = [Image.open(frame).convert('RGB') for frame in frames]
+
+    # out_frames = select_frames(frames)
+
+    request = {}
+
+    # Step 3: Convert images to byte format
+    byte_images = []
+    for img in out_frames:
+        byte_io = BytesIO()
+        img.save(byte_io, format='JPEG')
+        byte_images.append(byte_io.getvalue())
+
+    # Step 4: Pack the byte data with msgpack
+    packed_data = msgpack.packb(byte_images)
+    request['images'] = [packed_data]
+    # request['temperature'] = ['0.2'.encode()]
+    request['temperature'] = ['0.01'.encode()]
+    # request['prompt'] = ['describe the image in detail'.encode()]
+
+    #new_request = {}
+    #new_request['0'] = request['2']
+    handler = EndpointHandler()
+    print(handler(request))
+

inference_deployment/convert2inf2.py

@@ -0,0 +1,181 @@
+import sys
+import torch
+import os
+import random
+from io import BytesIO
+import numpy as np
+import time
+from llava.constants import MM_TOKEN_INDEX, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN, DEFAULT_VIDEO_TOKEN, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava.conversation import conv_templates, SeparatorStyle
+from llava.utils import disable_torch_init
+from llava.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria, process_images_v2
+from llava.model.builder import load_pretrained_model
+from llava.model.multimodal_encoder.processor import Blip2ImageTrainProcessor
+from llava.model import LlavaMistralForCausalLM
+from llava.model.multimodal_encoder.eva_vit import create_eva_vit_g
+import torch_neuronx
+import torch
+import torch_neuronx
+from llava.model import LlavaMistralForCausalLM
+from transformers import AutoTokenizer
+from llava.constants import MM_TOKEN_INDEX, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN, DEFAULT_VIDEO_TOKEN, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+
+
+from transformers import CLIPImageProcessor
+from PIL import Image
+import logging
+from qformer_tian import BertConfig, BertModel
+
+
+def select_frames(input_frames, num_segments = 10):
+
+    indices = np.linspace(start=0, stop=len(input_frames)-1, num=num_segments).astype(int)
+
+    frames = [input_frames[ind] for ind in indices]
+
+    return frames
+
+    
+def generate_input_ids(tokenizer):
+    conv = conv_templates['v1'].copy()
+    qs = "Describe the following video in detail."
+    qs = DEFAULT_VIDEO_START_TOKEN + DEFAULT_VIDEO_TOKEN + DEFAULT_VIDEO_END_TOKEN + '\n' + qs
+    conv.append_message(conv.roles[0], qs)
+    conv.append_message(conv.roles[1], None)
+    prompt = conv.get_prompt()
+    input_ids = tokenizer_image_token(prompt, tokenizer, MM_TOKEN_INDEX, return_tensors='pt').unsqueeze(0)
+    return input_ids, conv
+
+def uniform_sample(frames, num_segments):
+    indices = np.linspace(start=0, stop=len(frames) - 1, num=num_segments).astype(int)
+    frames = [frames[ind] for ind in indices]
+    return frames
+
+save_root = './inf2_weights'
+if not os.path.isdir(save_root):
+    os.makedirs(save_root)
+
+EVITG_SAVE_PATH = os.path.join(save_root, 'neuron_eva_vit_batch7.pth')
+LAYERNORM_SAVE_PATH = os.path.join(save_root, 'ln_state_dict.pth')
+QUERYTOKEN_SAVE_PATH = os.path.join(save_root, 'query_tokens.pth')
+BERT_SAVE_PATH = os.path.join(save_root, 'neuron_bert.pth')
+POSITION_ENCODING_SAVE_PATH = os.path.join(save_root, 'frame_position_encoding.pth')
+PROJECTOR_SAVE_PATH = os.path.join(save_root, 'projector.pth')
+EMBED_TOKENS_SAVE_PATH = os.path.join(save_root, 'embed_tokens.pth')
+
+
+model_path = './llava-mistral_videollava_ptv12_250k_samep_only_sopv2_mistralv2_scratch/'
+disable_torch_init()
+#print(model_path)
+device_map={"":'cpu'}
+kwargs = {"device_map": device_map}
+kwargs['torch_dtype'] = torch.float32
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+model = LlavaMistralForCausalLM.from_pretrained(
+    model_path,
+    low_cpu_mem_usage=True,
+    **kwargs
+)
+tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN], special_tokens=True)
+model.resize_token_embeddings(len(tokenizer))
+
+model.config.vit_precision == 'fp32'
+vision_tower = model.get_vision_tower()
+vision_tower.is_loaded = False
+vision_tower.load_model(device_map=device_map)
+vision_tower = vision_tower.to(torch.float32)
+
+vision_tower = vision_tower.eval()
+print('vision tower hiidden size')
+print(vision_tower.hidden_size)
+
+batch_size=7
+img_size=224
+input_shape = (batch_size, 3, img_size, img_size)
+input_data=torch.zeros(input_shape, dtype=torch.float32)
+model_neuronx = torch_neuronx.trace(vision_tower, input_data, compiler_args=["--model-type=transformer"])
+model_neuronx.save(EVITG_SAVE_PATH)
+
+image_processor = Blip2ImageTrainProcessor(
+    image_size=model.config.img_size,
+    is_training=False)
+
+input_ids, conv = generate_input_ids(tokenizer)
+device = torch.device('cpu')
+model = model.to(device)
+conv_mode = 'v1'
+NUM_SEGMENTS = 10
+
+video_dir = './v12044gd0000cl5c6rfog65i2eoqcqig'
+frames = [(int(os.path.splitext(item)[0]), os.path.join(video_dir, item)) for item in os.listdir(video_dir)]
+frames = [item[1] for item in sorted(frames, key=lambda x: x[0])]
+images = [Image.open(frame).convert('RGB') for frame in frames]
+images = uniform_sample(images, NUM_SEGMENTS)
+images =  process_images_v2(images, image_processor, model.config)
+
+#save layer norm
+ln_vision = model.get_ln_vision()
+ln_vision = ln_vision.eval()
+ln_state_dict = ln_vision.state_dict()
+torch.save(ln_state_dict, LAYERNORM_SAVE_PATH)
+
+
+query_tokens = model.get_query_tokens()
+#save query tokens
+query_tokens_state_dict = {'query_tokens': query_tokens.data}
+torch.save(query_tokens_state_dict, QUERYTOKEN_SAVE_PATH)
+
+#save qformer
+qformer = model.get_qformer()
+bert_torch = qformer.bert
+bert_torch = bert_torch.eval()
+bert_torch = bert_torch.to(torch.float32)
+
+
+vision_width = 1408
+cross_attention_freq = 2
+num_query_token = 32
+encoder_config = BertConfig.from_pretrained("bert-base-uncased")
+encoder_config.encoder_width = vision_width
+# insert cross-attention layer every other block
+encoder_config.add_cross_attention = True
+encoder_config.cross_attention_freq = cross_attention_freq
+encoder_config.query_length = num_query_token
+bert = BertModel(encoder_config, add_pooling_layer=False)
+bert.embeddings.word_embeddings = None
+bert.embeddings.position_embeddings = None
+
+for layer in bert.encoder.layer:
+    layer.output = None
+    layer.intermediate = None
+
+
+bert.load_state_dict(bert_torch.state_dict())
+bert = bert.eval()
+
+input_example = (
+        torch.zeros(70, 32, 768, dtype=torch.float32),
+        torch.zeros(70, 256, 1408, dtype=torch.float32),
+        torch.zeros(70, 256, dtype=torch.int64)
+)
+neuron_bert = torch_neuronx.trace(bert, input_example)
+neuron_bert.save(BERT_SAVE_PATH)
+
+#save projector and frame position encoding
+frame_position_encoding = model.get_frame_position_encoding()
+projector = model.get_model().mm_projector
+
+frame_position_encoding = frame_position_encoding.eval()
+frame_position_encoding = frame_position_encoding.to(torch.float32)
+
+projector = projector.eval()
+projector = projector.to(torch.float32)
+
+torch.save(frame_position_encoding.state_dict(), POSITION_ENCODING_SAVE_PATH)
+torch.save(projector.state_dict(), PROJECTOR_SAVE_PATH)
+
+#save embed_tokenss
+embed_tokens = model.get_model().embed_tokens
+embed_tokens = embed_tokens.eval()
+embed_tokens = embed_tokens.to(torch.float32)
+torch.save(embed_tokens.state_dict(), EMBED_TOKENS_SAVE_PATH)

inference_deployment/predict.py

@@ -0,0 +1,153 @@
+from llava.model.multimodal_encoder.processor import Blip2ImageTrainProcessor
+from llava.model import LlavaMistralForCausalLM
+from transformers import AutoTokenizer
+from llava.constants import MM_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_VIDEO_PATCH_TOKEN, DEFAULT_VIDEO_TOKEN, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN
+from llava.conversation import conv_templates
+import torch
+from llava.mm_utils import tokenizer_image_token
+import numpy as np
+from PIL import Image
+import os
+
+
+NUM_SEGMENTS = 10
+
+
+def load_model(model_path, device_map):
+    kwargs = {"device_map": device_map}
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    model = LlavaMistralForCausalLM.from_pretrained(
+        model_path,
+        low_cpu_mem_usage=True,
+        **kwargs
+    )
+    tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN], special_tokens=True)
+    model.resize_token_embeddings(len(tokenizer))
+
+    vision_tower = model.get_vision_tower()
+    if not vision_tower.is_loaded:
+        vision_tower.load_model(device_map=device_map)
+
+    image_processor = Blip2ImageTrainProcessor(
+        image_size=model.config.img_size,
+        is_training=False)
+    model.to(torch.float16)
+    return model, tokenizer, image_processor
+
+
+def generate_input_ids(tokenizer):
+    conv = conv_templates['v1'].copy()
+    qs = "Describe the following video in detail."
+    qs = DEFAULT_VIDEO_START_TOKEN + DEFAULT_VIDEO_TOKEN + DEFAULT_VIDEO_END_TOKEN + '\n' + qs
+    conv.append_message(conv.roles[0], qs)
+    conv.append_message(conv.roles[1], None)
+    prompt = conv.get_prompt()
+    input_ids = tokenizer_image_token(prompt, tokenizer, MM_TOKEN_INDEX, return_tensors='pt').unsqueeze(0)
+    return input_ids
+
+
+def generate_images(frame_folder, image_processor):
+    images = load_frames(frame_folder)
+    if len(images) > NUM_SEGMENTS:
+        images = uniform_sample(images, NUM_SEGMENTS)
+    
+    new_images = []
+    for image in images:
+        image = image_processor.preprocess(image)
+        new_images.append(image)
+    if all(x.shape == new_images[0].shape for x in new_images):
+        new_images = torch.stack(new_images, dim=0)
+    return new_images
+
+
+def uniform_sample(frames, num_segments):
+    indices = np.linspace(start=0, stop=len(frames) - 1, num=num_segments).astype(int)
+
+    frames = [frames[ind] for ind in indices]
+
+    return frames
+
+def load_frames(frames_dir):
+    results = []
+    image_files = [(int(os.path.splitext(img)[0]), img) for img in os.listdir(frames_dir) if img.endswith('jpg')]
+    image_files = sorted(image_files, key=lambda img: img[0])
+
+    for frame_name in image_files:
+        image_path = f"{frames_dir}/{frame_name[1]}"
+        image = Image.open(image_path).convert('RGB')
+        results.append(image)
+    return results
+
+
+class MASPVisionWrapper(torch.nn.Module):
+
+    def __init__(self, vision_tower, qformer, projector, query_tokens, frame_position_encoding, ln_vision):
+        super().__init__()
+        self.vision_tower = vision_tower
+        self.qformer = qformer
+        self.projector = projector
+        self.query_tokens = query_tokens
+        self.ln_vision = ln_vision
+        self.frame_position_encoding = frame_position_encoding
+
+    def forward(self, images): 
+        # images: [num_frames, 3, image_size, image_size]
+        image_features = self.vision_tower(images)
+        image_features = self.ln_vision(image_features)
+        attn_mask = torch.ones(image_features.size()[:-1], dtype=torch.long).to(image_features.device) # [num_frames, 256]
+        query_tokens = self.query_tokens.expand(image_features.shape[0], -1, -1) # [num_frames, 32, 768]
+        dtype_ = self.vision_tower.dtype
+        image_features = self.qformer.bert(
+            query_embeds= query_tokens.to(dtype_),
+            encoder_hidden_states=image_features.to(dtype_),
+            encoder_attention_mask=attn_mask,
+            return_dict=True
+        ).last_hidden_state.to(dtype_)
+        frame_ids = torch.arange(image_features.shape[0], dtype=torch.long, device=image_features.device) # [num_frames]
+        image_features += self.frame_position_encoding(frame_ids).unsqueeze(-2) #[num_frames, 1, 768]
+        return self.projector(image_features) 
+
+
+def inference(model_path, frame_folder):
+    # prepare
+    model, tokenizer, image_processor = load_model(model_path, device_map={"":0})
+    input_ids = generate_input_ids(tokenizer)[0].to(model.device) # [token_len]
+    images = generate_images(frame_folder, image_processor).to(model.device) # [num_frames, 3, image_size, image_size]
+    vision_module = MASPVisionWrapper(
+        vision_tower=model.get_vision_tower(),
+        qformer=model.get_qformer(),
+        projector=model.get_model().mm_projector,
+        query_tokens=model.get_query_tokens(),
+        frame_position_encoding=model.get_frame_position_encoding(),
+        ln_vision=model.get_ln_vision(),
+    )
+
+    # infernece
+    with torch.inference_mode():
+        # get image feature
+        image_features = vision_module(images).flatten(0, 1) # [num_frames * num_query_token, 4096]
+        # concat with text features 
+        vision_token_indice = torch.where(input_ids == MM_TOKEN_INDEX)[0][0]
+        pre_text_token = model.get_model().embed_tokens(input_ids[:vision_token_indice])
+        post_text_token = model.get_model().embed_tokens(input_ids[vision_token_indice+1:])
+        inputs_embeds = torch.cat([pre_text_token, image_features, post_text_token]).unsqueeze(0) # [1, num_token, 4096]
+        # llm.generate
+        output_ids = model.generate_from_base_class(
+            inputs_embeds=inputs_embeds,
+            do_sample=True,
+            temperature=0.01,
+            top_p=None,
+            num_beams=1,
+            max_new_tokens=1024,
+            pad_token_id=tokenizer.eos_token_id, 
+            use_cache=True,            
+        )
+    output = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]
+    output = output.strip()        
+    print(output)
+
+
+if __name__ == '__main__':
+    model_path = '/mnt/bn/algo-masp-nas-2/xiangchen/model/masp_models/llava-thothv2_mar_release_all_data'
+    frame_folder = '/mnt/bn/yukunfeng-nasdrive/xiangchen/masp_data/20231110_ttp/video/v12044gd0000cl5c6rfog65i2eoqcqig'
+    inference(model_path, frame_folder)

inference_deployment/predict_v09.py

@@ -0,0 +1,151 @@
+from llava.model.multimodal_encoder.processor import Blip2ImageTrainProcessor
+from llava.model import LlavaThothForCausalLM
+from transformers import AutoTokenizer
+from llava.constants import MM_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_VIDEO_PATCH_TOKEN, DEFAULT_VIDEO_TOKEN, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN
+from llava.conversation import conv_templates
+import torch
+from llava.mm_utils import tokenizer_image_token, process_images_v2, KeywordsStoppingCriteria
+import numpy as np
+from PIL import Image
+import os
+
+
+NUM_SEGMENTS = 10
+
+
+def load_model(model_path, device_map):
+    kwargs = {"device_map": device_map}
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    model = LlavaThothForCausalLM.from_pretrained(
+        model_path,
+        low_cpu_mem_usage=True,
+        **kwargs
+    )
+    tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_VIDEO_START_TOKEN, DEFAULT_VIDEO_END_TOKEN], special_tokens=True)
+    model.resize_token_embeddings(len(tokenizer))
+
+    vision_tower = model.get_vision_tower()
+    if not vision_tower.is_loaded:
+        vision_tower.load_model(device_map=device_map)
+
+    image_processor = Blip2ImageTrainProcessor(
+        image_size=model.config.img_size,
+        is_training=False)
+    model.to(torch.float16)
+    return model, tokenizer, image_processor
+
+
+def generate_input_ids(tokenizer):
+    conv = conv_templates['thoth'].copy()
+    qs = "Describe the following video in detail."
+    qs = DEFAULT_VIDEO_START_TOKEN + DEFAULT_VIDEO_TOKEN + DEFAULT_VIDEO_END_TOKEN + '\n' + qs
+    conv.append_message(conv.roles[0], qs)
+    conv.append_message(conv.roles[1], None)
+    prompt = conv.get_prompt()
+    input_ids = tokenizer_image_token(prompt, tokenizer, MM_TOKEN_INDEX, return_tensors='pt').unsqueeze(0)
+    return input_ids, conv
+
+
+def generate_images(frame_folder, image_processor, model_cfg):
+    images = load_frames(frame_folder)
+    if len(images) > NUM_SEGMENTS:
+        images = uniform_sample(images, NUM_SEGMENTS)
+    return process_images_v2(images, image_processor, model_cfg)
+ 
+
+def uniform_sample(frames, num_segments):
+    indices = np.linspace(start=0, stop=len(frames) - 1, num=num_segments).astype(int)
+    frames = [frames[ind] for ind in indices]
+    return frames
+
+def load_frames(frames_dir):
+    results = []
+    image_files = [(int(os.path.splitext(img)[0]), img) for img in os.listdir(frames_dir) if img.endswith('jpg')]
+    image_files = sorted(image_files, key=lambda img: img[0])
+
+    for frame_name in image_files:
+        image_path = f"{frames_dir}/{frame_name[1]}"
+        image = Image.open(image_path).convert('RGB')
+        results.append(image)
+    return results
+
+
+class MASPVisionWrapper(torch.nn.Module):
+
+    def __init__(self, vision_tower, qformer, projector, query_tokens, frame_position_encoding, ln_vision):
+        super().__init__()
+        self.vision_tower = vision_tower
+        self.qformer = qformer
+        self.projector = projector
+        self.query_tokens = query_tokens
+        self.ln_vision = ln_vision
+        self.frame_position_encoding = frame_position_encoding
+
+    def forward(self, images): 
+        # images: [num_frames, patches, 3, image_size, image_size]
+        image_features = self.vision_tower(images.flatten(0, 1))
+        image_features = self.ln_vision(image_features)
+        attn_mask = torch.ones(image_features.size()[:-1], dtype=torch.long).to(image_features.device) # [num_frames * num_patches, 256]
+        query_tokens = self.query_tokens.expand(image_features.shape[0], -1, -1) # [num_frames * num_patches, 32, 768]
+        dtype_ = self.vision_tower.dtype
+        image_features = self.qformer.bert(
+            query_embeds= query_tokens.to(dtype_),
+            encoder_hidden_states=image_features.to(dtype_),
+            encoder_attention_mask=attn_mask,
+            return_dict=True
+        ).last_hidden_state.to(dtype_)
+        frame_ids = torch.arange(images.shape[0], dtype=torch.long, device=image_features.device).unsqueeze(1)
+        frame_ids = frame_ids.repeat(1, images.shape[1]).flatten(0, 1) # [num_frames * num_patches]
+        image_features += self.frame_position_encoding(frame_ids).unsqueeze(-2) #[num_frames, 1, 768]
+        return self.projector(image_features) 
+
+
+def inference(model_path, frame_folder):
+    # prepare
+    model, tokenizer, image_processor = load_model(model_path, device_map={"":0})
+    input_ids, conv = generate_input_ids(tokenizer)
+    images = generate_images(frame_folder, image_processor, model.config).to(model.device).half() # [num_frames, patches, 3, image_size, image_size]
+    vision_module = MASPVisionWrapper(
+        vision_tower=model.get_vision_tower(),
+        qformer=model.get_qformer(),
+        projector=model.get_model().mm_projector,
+        query_tokens=model.get_query_tokens(),
+        frame_position_encoding=model.get_frame_position_encoding(),
+        ln_vision=model.get_ln_vision(),
+    )
+    stop_str = conv.sep if conv.sep2 is None else conv.sep2
+    keywords = [stop_str]
+    stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
+    input_ids = input_ids[0].to(model.device) # [token_len]
+
+    # infernece
+    with torch.inference_mode():
+        # get image feature
+        image_features = vision_module(images).flatten(0, 1) # [num_frames * num_patches * num_query_token, 4096]
+        # concat with text features 
+        vision_token_indice = torch.where(input_ids == MM_TOKEN_INDEX)[0][0]
+        pre_text_token = model.get_model().embed_tokens(input_ids[:vision_token_indice])
+        post_text_token = model.get_model().embed_tokens(input_ids[vision_token_indice+1:])
+        inputs_embeds = torch.cat([pre_text_token, image_features, post_text_token]).unsqueeze(0) # [1, num_token, 4096]
+        
+        # llm.generate
+        output_ids = model.generate_from_base_class(
+            inputs_embeds=inputs_embeds,
+            do_sample=True,
+            temperature=0.01,
+            top_p=None,
+            num_beams=1,
+            max_new_tokens=1024,
+            pad_token_id=tokenizer.eos_token_id, 
+            use_cache=True,
+            stopping_criteria=[stopping_criteria]       
+        )
+    output = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]
+    output = output.strip()        
+    print(output)
+
+
+if __name__ == '__main__':
+    model_path = '/mnt/bn/algo-masp-nas-2/xiangchen/model/masp_models/llava-thothv2_mar_release_all_data'
+    frame_folder = '/mnt/bn/yukunfeng-nasdrive/xiangchen/masp_data/20231110_ttp/video/v12044gd0000cl5c6rfog65i2eoqcqig'
+    inference(model_path, frame_folder)

inference_deployment/qformer_tian.py

@@ -0,0 +1,1151 @@
+"""
+Adapted from salesforce@LAVIS. Below is the original copyright:
+ * Copyright (c) 2023, salesforce.com, inc.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+ * By Junnan Li
+ * Based on huggingface code base
+ * https://github.com/huggingface/transformers/blob/v4.15.0/src/transformers/models/bert
+"""
+
+import math
+import os
+import warnings
+from dataclasses import dataclass
+from typing import Optional, Tuple, Dict, Any
+
+import torch
+from torch import Tensor, device, dtype, nn
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+import torch.nn.functional as F
+
+from transformers.activations import ACT2FN
+from transformers.file_utils import (
+    ModelOutput,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    NextSentencePredictorOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers.models.bert.configuration_bert import BertConfig
+from llava.model.utils import LayerNorm
+
+logger = logging.get_logger(__name__)
+
+
+class BertEmbeddings(nn.Module):
+    """Construct the embeddings from word and position embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = None
+        self.position_embeddings = None
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        # self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.LayerNorm = LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1))
+        )
+        self.position_embedding_type = getattr(
+            config, "position_embedding_type", "absolute"
+        )
+
+        self.config = config
+
+    def forward(
+        self,
+        input_ids=None,
+        position_ids=None,
+        query_embeds=None,
+        past_key_values_length=0,
+    ):
+        seq_length = 0
+
+        position_ids = self.position_ids[
+            :, past_key_values_length : seq_length + past_key_values_length
+            ].clone()
+
+        embeddings = query_embeds
+
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class BertSelfAttention(nn.Module):
+    def __init__(self, config, is_cross_attention):
+        super().__init__()
+        self.config = config
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
+            config, "embedding_size"
+        ):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads)
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        if is_cross_attention:
+            self.key = nn.Linear(config.encoder_width, self.all_head_size)
+            self.value = nn.Linear(config.encoder_width, self.all_head_size)
+        else:
+            self.key = nn.Linear(config.hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(
+                2 * config.max_position_embeddings - 1, self.attention_head_size
+            )
+        self.save_attention = False
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    def save_attention_map(self, attention_map):
+        self.attention_map = attention_map
+
+    def get_attention_map(self):
+        return self.attention_map
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        mixed_query_layer = self.query(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(-1, 1)
+            position_ids_r = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(
+                distance + self.max_position_embeddings - 1
+            )
+            positional_embedding = positional_embedding.to(
+                dtype=query_layer.dtype
+            )  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                relative_position_scores_key = torch.einsum(
+                    "bhrd,lrd->bhlr", key_layer, positional_embedding
+                )
+                attention_scores = (
+                    attention_scores
+                    + relative_position_scores_query
+                    + relative_position_scores_key
+                )
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores.to(torch.float32)).to(attention_scores.dtype)
+
+        if is_cross_attention and self.save_attention:
+            self.save_attention_map(attention_probs)
+            attention_probs.register_hook(self.save_attn_gradients)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs_dropped = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs_dropped = attention_probs_dropped * head_mask
+
+        context_layer = torch.matmul(attention_probs_dropped, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (
+            (context_layer, attention_probs) if output_attentions else (context_layer,)
+        )
+
+        outputs = outputs + (past_key_value,)
+        return outputs
+
+
+class BertSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        # self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.LayerNorm = LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+        self.self = BertSelfAttention(config, is_cross_attention)
+        self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads,
+            self.self.num_attention_heads,
+            self.self.attention_head_size,
+            self.pruned_heads,
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = (
+            self.self.attention_head_size * self.self.num_attention_heads
+        )
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+
+        outputs = (attention_output,) + self_outputs[
+            1:
+        ]  # add attentions if we output them
+        return outputs
+
+
+class BertIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BertOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        # self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.LayerNorm = LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertLayer(nn.Module):
+    def __init__(self, config, layer_num):
+        super().__init__()
+        self.config = config
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = BertAttention(config)
+        self.layer_num = layer_num
+        if (
+            self.config.add_cross_attention
+            and layer_num % self.config.cross_attention_freq == 0
+        ):
+            self.crossattention = BertAttention(
+                config, is_cross_attention=self.config.add_cross_attention
+            )
+            self.has_cross_attention = True
+        else:
+            self.has_cross_attention = False
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+
+        self.intermediate_query = BertIntermediate(config)
+        self.output_query = BertOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        query_length=0,
+    ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = (
+            past_key_value[:2] if past_key_value is not None else None
+        )
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:-1]
+
+        present_key_value = self_attention_outputs[-1]
+
+        if query_length > 0:
+            query_attention_output = attention_output[:, :query_length, :]
+
+            if self.has_cross_attention:
+                assert (
+                    encoder_hidden_states is not None
+                ), "encoder_hidden_states must be given for cross-attention layers"
+                cross_attention_outputs = self.crossattention(
+                    query_attention_output,
+                    attention_mask,
+                    head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    output_attentions=output_attentions,
+                )
+                query_attention_output = cross_attention_outputs[0]
+                outputs = (
+                    outputs + cross_attention_outputs[1:-1]
+                )  # add cross attentions if we output attention weights
+
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk_query,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                query_attention_output,
+            )
+            if attention_output.shape[1] > query_length:
+                layer_output_text = apply_chunking_to_forward(
+                    self.feed_forward_chunk,
+                    self.chunk_size_feed_forward,
+                    self.seq_len_dim,
+                    attention_output[:, query_length:, :],
+                )
+                layer_output = torch.cat([layer_output, layer_output_text], dim=1)
+        else:
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                attention_output,
+            )
+        outputs = (layer_output,) + outputs
+
+        outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+    def feed_forward_chunk_query(self, attention_output):
+        intermediate_output = self.intermediate_query(attention_output)
+        layer_output = self.output_query(intermediate_output, attention_output)
+        return layer_output
+
+
+class BertEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [BertLayer(config, i) for i in range(config.num_hidden_layers)]
+        )
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+        query_length=0,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (
+            () if output_attentions and self.config.add_cross_attention else None
+        )
+
+        next_decoder_cache = () if use_cache else None
+
+        for i in range(self.config.num_hidden_layers):
+            layer_module = self.layer[i]
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False) and self.training:
+
+                if use_cache:
+                    logger.warn(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(
+                            *inputs, past_key_value, output_attentions, query_length
+                        )
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                    query_length,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class BertPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class BertPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        # self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.LayerNorm = LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class BertLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = BertPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class BertOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class BertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BertConfig
+    base_model_prefix = "bert"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+class BertModel(BertPreTrainedModel):
+    """
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in `Attention is
+    all you need <https://arxiv.org/abs/1706.03762>`__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
+    Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+    argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an
+    input to the forward pass.
+    """
+
+    def __init__(self, config, add_pooling_layer=False):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+
+        self.encoder = BertEncoder(config)
+
+        self.pooler = BertPooler(config) if add_pooling_layer else None
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def get_extended_attention_mask(
+        self,
+        attention_mask: Tensor,
+        input_shape: Tuple[int],
+        device: device,
+        is_decoder: bool,
+        has_query: bool = False,
+    ) -> Tensor:
+        """
+        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
+
+        Arguments:
+            attention_mask (:obj:`torch.Tensor`):
+                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
+            input_shape (:obj:`Tuple[int]`):
+                The shape of the input to the model.
+            device: (:obj:`torch.device`):
+                The device of the input to the model.
+
+        Returns:
+            :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.
+        """
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if is_decoder:
+                batch_size, seq_length = input_shape
+
+                seq_ids = torch.arange(seq_length, device=device)
+                causal_mask = (
+                    seq_ids[None, None, :].repeat(batch_size, seq_length, 1)
+                    <= seq_ids[None, :, None]
+                )
+
+                # add a prefix ones mask to the causal mask
+                # causal and attention masks must have same type with pytorch version < 1.3
+                causal_mask = causal_mask.to(attention_mask.dtype)
+
+                if causal_mask.shape[1] < attention_mask.shape[1]:
+                    prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
+                    if has_query:  # UniLM style attention mask
+                        causal_mask = torch.cat(
+                            [
+                                torch.zeros(
+                                    (batch_size, prefix_seq_len, seq_length),
+                                    device=device,
+                                    dtype=causal_mask.dtype,
+                                ),
+                                causal_mask,
+                            ],
+                            axis=1,
+                        )
+                    causal_mask = torch.cat(
+                        [
+                            torch.ones(
+                                (batch_size, causal_mask.shape[1], prefix_seq_len),
+                                device=device,
+                                dtype=causal_mask.dtype,
+                            ),
+                            causal_mask,
+                        ],
+                        axis=-1,
+                    )
+                extended_attention_mask = (
+                    causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+                )
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError(
+                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
+                    input_shape, attention_mask.shape
+                )
+            )
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(
+            dtype=self.dtype
+        )  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        return extended_attention_mask
+
+    def forward(
+        self,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        """
+        output_attentions = self.config.output_attentions
+
+        output_hidden_states = self.config.output_hidden_states
+
+        return_dict = True
+        
+        use_cache = False
+
+        input_ids = None
+        
+        position_ids = None
+
+        # past_key_values_length
+        past_key_values_length = 0
+
+        query_length = query_embeds.shape[1]
+        
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            query_embeds=query_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+
+        input_shape = embedding_output.size()[:-1]
+        batch_size = input_shape[-2]
+        seq_length = input_shape[-1]
+        device = embedding_output.device
+
+        attention_mask = None
+        attention_mask = torch.ones(
+            ((batch_size, seq_length + past_key_values_length)), device=device
+            )
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        is_decoder = False
+        extended_attention_mask = self.get_extended_attention_mask(
+                attention_mask, input_shape, device, is_decoder
+            )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        print("encoder_hidden_states.size(): ", encoder_hidden_states.size())
+        encoder_batch_size = encoder_hidden_states.size()[-3]
+        encoder_sequence_length = encoder_hidden_states.size()[-2]
+        encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+        print("encoder_hidden_shape: ", encoder_hidden_shape)
+
+        encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = None
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        past_key_values = None
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            query_length=query_length,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = None
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class BertLMHeadModel(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        past_key_values=None,
+        use_cache=True,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        return_logits=False,
+        is_decoder=True,
+        reduction="mean",
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
+            ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are
+            ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]``
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        Returns:
+        Example::
+            >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig
+            >>> import torch
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
+            >>> config = BertConfig.from_pretrained("bert-base-cased")
+            >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config)
+            >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+            >>> outputs = model(**inputs)
+            >>> prediction_logits = outputs.logits
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+        if labels is not None:
+            use_cache = False
+        if past_key_values is not None:
+            query_embeds = None
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            query_embeds=query_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_decoder=is_decoder,
+        )
+
+        sequence_output = outputs[0]
+        if query_embeds is not None:
+            sequence_output = outputs[0][:, query_embeds.shape[1] :, :]
+
+        prediction_scores = self.cls(sequence_output)
+
+        if return_logits:
+            return prediction_scores[:, :-1, :].contiguous()
+
+        lm_loss = None
+        if labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
+            labels = labels[:, 1:].contiguous()
+            loss_fct = CrossEntropyLoss(reduction=reduction, label_smoothing=0.1)
+            lm_loss = loss_fct(
+                shifted_prediction_scores.view(-1, self.config.vocab_size),
+                labels.view(-1),
+            )
+            if reduction == "none":
+                lm_loss = lm_loss.view(prediction_scores.size(0), -1).sum(1)
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=lm_loss,
+            logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, query_embeds, past=None, attention_mask=None, **model_kwargs
+    ):
+        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
+        if attention_mask is None:
+            attention_mask = input_ids.new_ones(input_ids.shape)
+        query_mask = input_ids.new_ones(query_embeds.shape[:-1])
+        attention_mask = torch.cat([query_mask, attention_mask], dim=-1)
+
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
+        return {
+            "input_ids": input_ids,
+            "query_embeds": query_embeds,
+            "attention_mask": attention_mask,
+            "past_key_values": past,
+            "encoder_hidden_states": model_kwargs.get("encoder_hidden_states", None),
+            "encoder_attention_mask": model_kwargs.get("encoder_attention_mask", None),
+            "is_decoder": True,
+        }
+
+    def _reorder_cache(self, past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (
+                tuple(
+                    past_state.index_select(0, beam_idx) for past_state in layer_past
+                ),
+            )
+        return reordered_past
+
+
+class BertForMaskedLM(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        return_logits=False,
+        is_decoder=False,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,
+            config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored
+            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``
+        """
+
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            query_embeds=query_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_decoder=is_decoder,
+        )
+
+        if query_embeds is not None:
+            sequence_output = outputs[0][:, query_embeds.shape[1] :, :]
+        prediction_scores = self.cls(sequence_output)
+
+        if return_logits:
+            return prediction_scores
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            masked_lm_loss = loss_fct(
+                prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return (
+                ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+            )
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

llava/__init__.py

@@ -0,0 +1,9 @@
+from .model import LlavaLlamaForCausalLM, LlavaMistralForCausalLM
+try:
+    from .model import LlavaGemmaForCausalLM
+except:
+    pass
+try:
+    from .model import LlavaThothForCausalLM
+except:
+    pass

llava/configs/action_dataset_ablation/finetune_webvid.yaml

@@ -0,0 +1,11 @@
+datasets:
+
+  lk_image:
+    data_type: image
+
+  lk_video:
+    data_type: frames
+    conv_type: multi
+    fps: 1.0
+    select_datasets: ['webvid10m', 'webvid2m']
+    # select_datasets: ['webvid10m', 'webvid2m', 'activitynet', 'vidal', 'hdvila']

llava/configs/action_dataset_ablation/finetune_webvid_act.yaml

@@ -0,0 +1,11 @@
+datasets:
+
+  lk_image:
+    data_type: image
+
+  lk_video:
+    data_type: frames
+    conv_type: multi
+    fps: 1.0
+    select_datasets: ['webvid10m', 'webvid2m', 'activitynet']
+    # select_datasets: ['webvid10m', 'webvid2m', 'activitynet', 'vidal', 'hdvila']

llava/configs/action_dataset_ablation/finetune_webvid_hdvila.yaml

@@ -0,0 +1,11 @@
+datasets:
+
+  lk_image:
+    data_type: image
+
+  lk_video:
+    data_type: frames
+    conv_type: multi
+    fps: 1.0
+    select_datasets: ['webvid10m', 'webvid2m', 'hdvila']
+    # select_datasets: ['webvid10m', 'webvid2m', 'activitynet', 'vidal', 'hdvila']

llava/configs/action_dataset_ablation/finetune_webvid_vidal.yaml

@@ -0,0 +1,11 @@
+datasets:
+
+  lk_image:
+    data_type: image
+
+  lk_video:
+    data_type: frames
+    conv_type: multi
+    fps: 1.0
+    select_datasets: ['webvid10m', 'webvid2m', 'vidal']
+    # select_datasets: ['webvid10m', 'webvid2m', 'activitynet', 'vidal', 'hdvila']

llava/configs/adso_increasing_ablation/finetune_data_pure_gpt4v.yaml

@@ -0,0 +1,55 @@
+datasets:
+
+#  m3it:
+#    data_type: images
+#    sample_ratio: 4
+#    tasks:
+#      - coco
+#      - coco-goi
+#      - coco-text
+#      - imagenet
+#      - coco-itm
+#      - iqa
+#      - mocheg
+#      - vsr
+#      - refcoco
+#      - science-qa
+#      - vqa-v2
+#      - gqa
+#      - st-vqa
+#      - text-vqa
+#      - okvqa
+#      - a-okvqa
+#
+#  tt_vqa:
+#    data_type: frames
+#    sample_ratio: 1
+
+  ShareGPT4V:
+    data_type: images
+    sample_ratio: 1
+
+  gpt4v_tt_vqa:
+    data_type: frames
+    fps: 0.5
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['caption', 'qas']
+
+  gpt4v_public:
+    data_type: frames
+    fps: 1.0
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']
+    sample_method: sequential
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/data/shared_gpt4v_data/data_130k.json
+
+  gpt4v_internal:
+    data_type: frames
+    fps: 2.0
+    sample_ratio: 1
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']
+
+

llava/configs/adso_increasing_ablation/finetune_gpt4v_adso135k.yaml

@@ -0,0 +1,57 @@
+datasets:
+
+#  m3it:
+#    data_type: images
+#    sample_ratio: 4
+#    tasks:
+#      - coco
+#      - coco-goi
+#      - coco-text
+#      - imagenet
+#      - coco-itm
+#      - iqa
+#      - mocheg
+#      - vsr
+#      - refcoco
+#      - science-qa
+#      - vqa-v2
+#      - gqa
+#      - st-vqa
+#      - text-vqa
+#      - okvqa
+#      - a-okvqa
+#
+  tt_vqa:
+    data_type: frames
+    sample_ratio: 2
+    fps: 2.0
+    conv_type: single  
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/dataset/masp/20240208_meta_data_single_135k_caption_160k_QA.json
+
+  ShareGPT4V:
+    data_type: images
+    sample_ratio: 1
+
+  gpt4v_tt_vqa:
+    data_type: frames
+    fps: 0.5
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['caption', 'qas']
+
+
+  gpt4v_public:
+    data_type: frames
+    fps: 1.0
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']
+    sample_method: sequential
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/data/shared_gpt4v_data/data_130k.json
+
+  gpt4v_internal:
+    data_type: frames
+    fps: 2.0
+    sample_ratio: 1
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']

llava/configs/adso_increasing_ablation/finetune_gpt4v_adso185k.yaml

@@ -0,0 +1,57 @@
+datasets:
+
+#  m3it:
+#    data_type: images
+#    sample_ratio: 4
+#    tasks:
+#      - coco
+#      - coco-goi
+#      - coco-text
+#      - imagenet
+#      - coco-itm
+#      - iqa
+#      - mocheg
+#      - vsr
+#      - refcoco
+#      - science-qa
+#      - vqa-v2
+#      - gqa
+#      - st-vqa
+#      - text-vqa
+#      - okvqa
+#      - a-okvqa
+#
+  tt_vqa:
+    data_type: frames
+    sample_ratio: 3
+    fps: 2.0
+    conv_type: single  
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/dataset/masp/20240220_meta_data_single_190k_caption_160k_QA.json
+
+  ShareGPT4V:
+    data_type: images
+    sample_ratio: 1
+
+  gpt4v_tt_vqa:
+    data_type: frames
+    fps: 0.5
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['caption', 'qas']
+
+
+  gpt4v_public:
+    data_type: frames
+    fps: 1.0
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']
+    sample_method: sequential
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/data/shared_gpt4v_data/data_130k.json
+
+  gpt4v_internal:
+    data_type: frames
+    fps: 2.0
+    sample_ratio: 1
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']

llava/configs/adso_increasing_ablation/finetune_gpt4v_adso185k_baseline.yaml

@@ -0,0 +1,55 @@
+datasets:
+
+#  m3it:
+#    data_type: images
+#    sample_ratio: 4
+#    tasks:
+#      - coco
+#      - coco-goi
+#      - coco-text
+#      - imagenet
+#      - coco-itm
+#      - iqa
+#      - mocheg
+#      - vsr
+#      - refcoco
+#      - science-qa
+#      - vqa-v2
+#      - gqa
+#      - st-vqa
+#      - text-vqa
+#      - okvqa
+#      - a-okvqa
+#
+  tt_vqa:
+    data_type: frames
+    sample_ratio: 3
+    fps: 2.0
+    conv_type: single  
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/dataset/masp/20240220_meta_data_single_190k_caption_160k_QA.json
+
+  ShareGPT4V:
+    data_type: images
+    sample_ratio: 1
+
+  gpt4v_tt_vqa:
+    data_type: frames
+    fps: 0.5
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['caption', 'qas']
+
+
+  lk_video:
+    data_type: frames
+    conv_type: multi
+    fps: 1.0
+    sample_ratio: 6
+  
+  gpt4v_internal:
+    data_type: frames
+    fps: 2.0
+    sample_ratio: 1
+    conv_type: single
+    task_types: ['detail']
+

llava/configs/adso_increasing_ablation/finetune_gpt4v_adso185k_no_qa.yaml

@@ -0,0 +1,57 @@
+datasets:
+
+#  m3it:
+#    data_type: images
+#    sample_ratio: 4
+#    tasks:
+#      - coco
+#      - coco-goi
+#      - coco-text
+#      - imagenet
+#      - coco-itm
+#      - iqa
+#      - mocheg
+#      - vsr
+#      - refcoco
+#      - science-qa
+#      - vqa-v2
+#      - gqa
+#      - st-vqa
+#      - text-vqa
+#      - okvqa
+#      - a-okvqa
+#
+  tt_vqa:
+    data_type: frames
+    sample_ratio: 3
+    fps: 2.0
+    conv_type: single  
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/dataset/masp/20240220_meta_data_single_190k_caption_no_QA.json
+
+  ShareGPT4V:
+    data_type: images
+    sample_ratio: 1
+
+  gpt4v_tt_vqa:
+    data_type: frames
+    fps: 0.5
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['caption', 'qas']
+
+
+  gpt4v_public:
+    data_type: frames
+    fps: 1.0
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']
+    sample_method: sequential
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/data/shared_gpt4v_data/data_130k.json
+
+  gpt4v_internal:
+    data_type: frames
+    fps: 2.0
+    sample_ratio: 1
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']

llava/configs/adso_increasing_ablation/finetune_gpt4v_adso65k.yaml

@@ -0,0 +1,57 @@
+datasets:
+
+#  m3it:
+#    data_type: images
+#    sample_ratio: 4
+#    tasks:
+#      - coco
+#      - coco-goi
+#      - coco-text
+#      - imagenet
+#      - coco-itm
+#      - iqa
+#      - mocheg
+#      - vsr
+#      - refcoco
+#      - science-qa
+#      - vqa-v2
+#      - gqa
+#      - st-vqa
+#      - text-vqa
+#      - okvqa
+#      - a-okvqa
+#
+  tt_vqa:
+    data_type: frames
+    sample_ratio: 2
+    fps: 2.0
+    conv_type: single  
+    train_data_path: /mnt/bn/algo-masp-nas-2/baiyi.by/data/ADSO_Anno_Data/batch_20231128/meta_data_single_60k_caption_170k_QA.json
+
+  ShareGPT4V:
+    data_type: images
+    sample_ratio: 1
+
+  gpt4v_tt_vqa:
+    data_type: frames
+    fps: 0.5
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['caption', 'qas']
+
+
+  gpt4v_public:
+    data_type: frames
+    fps: 1.0
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']
+    sample_method: sequential
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/data/shared_gpt4v_data/data_130k.json
+
+  gpt4v_internal:
+    data_type: frames
+    fps: 2.0
+    sample_ratio: 1
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']

llava/configs/finetune_debug.yaml

@@ -0,0 +1,8 @@
+datasets:
+  gpt4v_public:
+    data_type: frames
+    fps: 1.0
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/data/shared_gpt4v_data/data_130k.json

llava/configs/finetune_gpt4v_adso65k.yaml

@@ -0,0 +1,56 @@
+datasets:
+
+#  m3it:
+#    data_type: images
+#    sample_ratio: 4
+#    tasks:
+#      - coco
+#      - coco-goi
+#      - coco-text
+#      - imagenet
+#      - coco-itm
+#      - iqa
+#      - mocheg
+#      - vsr
+#      - refcoco
+#      - science-qa
+#      - vqa-v2
+#      - gqa
+#      - st-vqa
+#      - text-vqa
+#      - okvqa
+#      - a-okvqa
+#
+  tt_vqa:
+    data_type: frames
+    sample_ratio: 2
+    fps: 2.0
+    conv_type: single  
+    train_data_path: /mnt/bn/algo-masp-nas-2/baiyi.by/data/ADSO_Anno_Data/batch_20231128/meta_data_single_60k_caption_170k_QA.json
+
+  ShareGPT4V:
+    data_type: images
+    sample_ratio: 1
+
+  gpt4v_tt_vqa:
+    data_type: frames
+    fps: 0.5
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['caption', 'qas']
+
+
+  gpt4v_public:
+    data_type: frames
+    fps: 1.0
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/data/shared_gpt4v_data/data_130k.json
+
+  gpt4v_internal:
+    data_type: frames
+    fps: 2.0
+    sample_ratio: 1
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']

llava/configs/gpt4v_increasing_ablation/finetune_gpt4v_public500k.yaml

@@ -0,0 +1,57 @@
+datasets:
+
+#  m3it:
+#    data_type: images
+#    sample_ratio: 4
+#    tasks:
+#      - coco
+#      - coco-goi
+#      - coco-text
+#      - imagenet
+#      - coco-itm
+#      - iqa
+#      - mocheg
+#      - vsr
+#      - refcoco
+#      - science-qa
+#      - vqa-v2
+#      - gqa
+#      - st-vqa
+#      - text-vqa
+#      - okvqa
+#      - a-okvqa
+#
+  # tt_vqa:
+  #   data_type: frames
+  #   sample_ratio: 2
+  #   fps: 2.0
+  #   conv_type: single  
+  #   train_data_path: /mnt/bn/algo-masp-nas-2/baiyi.by/data/ADSO_Anno_Data/batch_20231128/meta_data_single_60k_caption_170k_QA.json
+
+  ShareGPT4V:
+    data_type: images
+    sample_ratio: 1
+
+
+  gpt4v_tt_vqa:
+    data_type: frames
+    fps: 0.5
+    sample_ratio: 6
+    conv_type: single
+    task_types: ['caption', 'qas']
+
+
+  gpt4v_public:
+    data_type: frames
+    fps: 1.0
+    sample_ratio: 10
+    conv_type: single
+    task_types: ['summary', 'detail']
+    train_data_path: /mnt/bn/algo-masp-nas-2/xiangchen/data/shared_gpt4v_data/data_500k_filtered.json
+
+  gpt4v_internal:
+    data_type: frames
+    fps: 2.0
+    sample_ratio: 1
+    conv_type: single
+    task_types: ['summary', 'detail', 'qa_pairs']