init

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

@@ -1,10 +1,438 @@
----
-title: Starvector 1b Im2svg
-emoji: 😻
-colorFrom: pink
-colorTo: indigo
-sdk: docker
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+<div align="center">
+  <h1>💫 StarVector: Generating Scalable Vector Graphics Code from Images and Text</h1>
+  <img src="assets/starvector-xyz.png" alt="starvector" style="width: 800px; display: block; margin-left: auto; margin-right: auto;"/>
+
+<a href="https://arxiv.org/abs/2312.11556" target="_blank">
+    <img alt="arXiv" src="https://img.shields.io/badge/arXiv-StarVector-red?logo=arxiv" height="25" />
+</a>
+<a href="https://starvector.github.io/" target="_blank">
+    <img alt="Website" src="https://img.shields.io/badge/🌎_Website-starvector.github.io-blue.svg" height="25" />
+</a>
+<a href="https://huggingface.co/starvector/starvector-1b-im2svg" target="_blank">
+    <img alt="HF Models: StarVector" src="https://img.shields.io/badge/%F0%9F%A4%97%20_Model-StarVector--1B-ffc107?color=ffc107&logoColor=white" height="25" />
+</a>
+<a href="https://huggingface.co/starvector/starvector-8b-im2svg" target="_blank">
+    <img alt="HF Models: StarVector" src="https://img.shields.io/badge/%F0%9F%A4%97%20_Model-StarVector--8B-ffc107?color=ffc107&logoColor=white" height="25" />
+</a>
+<a href="https://huggingface.co/datasets/starvector/svg-stack" target="_blank">
+    <img alt="HF Dataset: SVG-Stack" src="https://img.shields.io/badge/%F0%9F%A4%97%20_Data-SVG--Stack-ffc107?color=ffc107&logoColor=white" height="25" />
+</a>
+<a href="https://huggingface.co/collections/starvector/starvector-svg-datasets-svg-bench-67811204a76475be4dd66d09" target="_blank">
+    <img alt="HF Dataset: SVG-Bench" src="https://img.shields.io/badge/%F0%9F%A4%97%20_Benchmark-SVG--Bench-ffc107?color=ffc107&logoColor=white" height="25" />
+</a>
+
+<div style="font-family: charter;">
+    <a href="https://joanrod.github.io" target="_blank">Juan A. Rodriguez</a>,
+    <a href="https://abhaypuri.github.io/portfolio/" target="_blank">Abhay Puri</a>,
+    <a href="https://shubhamagarwal92.github.io/" target="_blank">Shubham Agarwal</a>,
+    <a href="https://scholar.google.ca/citations?user=8vRS7F0AAAAJ&hl=en" target="_blank">Issam H. Laradji</a>,
+    <a href="https://scholar.google.es/citations?user=IwBx73wAAAAJ&hl=ca" target="_blank">Pau Rodriguez</a>,
+    <a href="https://scholar.google.es/citations?user=1jHvtfsAAAAJ&hl=ca" target="_blank">David Vazquez</a>,
+    <a href="https://scholar.google.com/citations?user=1ScWJOoAAAAJ&hl=en" target="_blank">Chris Pal</a>,
+    <a href="https://scholar.google.com/citations?user=aVfyPAoAAAAJ&hl=en" target="_blank">Marco Pedersoli</a>
+</div>
+
+</div>
+
+## 🔥 News
+- March 2025: **StarVector Accepted at CVPR 2025**,
+  - StarVector has been accepted at CVPR 2025! [[Link](https://arxiv.org/abs/2312.11556)]
+  - Check out our website for more information [[Link](https://starvector.github.io/)]
+  - StarVector models are now available on Hugging Face Model Hub! [[Link](https://huggingface.co/starvector/starvector-1b-im2svg)] [[Link](https://huggingface.co/starvector/starvector-8b-im2svg)]
+  - SVGBench and SVG-Stack datasets are now available on Hugging Face Datasets Hub! [[Link](https://huggingface.co/datasets/starvector/svg-bench)] [[Link](https://huggingface.co/datasets/starvector/svg-stack)]
+  
+## 🚀 Introduction
+StarVector is a multimodal vision-language model for Scalable Vector Graphics (SVG) generation. It can be used to perform image2SVG and text2SVG generation. We pose image generation as a code generation task, using the power of multimodal VLMs
+
+<div align="center">
+  <img src="assets/starvector-teaser.png" alt="starvector" style="width: 900px; display: block; margin-left: auto; margin-right: auto;" />
+</div>
+
+> **Abstract**: Scalable Vector Graphics (SVGs) are vital for modern image rendering due to their scalability and versatility. Previous SVG generation methods have focused on curve-based vectorization, lacking semantic understanding, often producing artifacts, and struggling with SVG primitives beyond \textit{path} curves. To address these issues, we introduce StarVector, a multimodal large language model for SVG generation. It performs image vectorization by understanding image semantics and using SVG primitives for compact, precise outputs. Unlike traditional methods, StarVector works directly in the SVG code space, leveraging visual understanding to apply accurate SVG primitives. To train StarVector, we create SVG-Stack, a diverse dataset of 2M samples that enables generalization across vectorization tasks and precise use of primitives like ellipses, polygons, and text. We address challenges in SVG evaluation, showing that pixel-based metrics like MSE fail to capture the unique qualities of vector graphics. We introduce SVG-Bench, a benchmark across 10 datasets, and 3 tasks: Image-to-SVG, Text-to-SVG generation, and diagram generation. Using this setup, StarVector achieves state-of-the-art performance, producing more compact and semantically rich SVGs.
+
+### Multimodal Architecture
+
+StarVector uses a multimodal architecture to process images and text. When performing Image-to-SVG (or image vectorization), the image is projected into visual tokens, and SVG code is generated. When performing Text-to-SVG, the model only recieves the text instruction (no image is provided), and a novel SVG is created. The LLM is based of StarCoder, which we leverage to transfer coding skills to SVG generation.
+
+<div align="center">
+  <img src="assets/starvector-arch.png" alt="starvector" style="width: 700px; display: block; margin-left: auto; margin-right: auto;" />
+</div>
+
+## 📖 Table of Contents
+- [💿 Installation](#installation)
+- [🏎️ Quick Start - Image2SVG Generation](#quick-start---image2svg-generation)
+- [🎨 Models](#models)
+- [📊 Datasets](#datasets---svg-bench)
+- [🏋️‍♂️ Training](#training)
+- [🏆 Evaluation on SVG-Bench](#validation-on-svg-benchmarks-svg-bench)
+- [🧩 Demo](#starvector-demo)
+- [📚 Citation](#citation)
+- [📝 License](#license)
+
+
+## Installation
+
+1. Clone this repository and navigate to star-vector folder
+```bash
+git clone https://github.com/joanrod/star-vector.git
+cd star-vector
+```
+
+2. Install Package
+```Shell
+conda create -n starvector python=3.11.3 -y
+conda activate starvector
+pip install --upgrade pip  # enable PEP 660 support
+pip install -e .
+```
+
+3. Install additional packages for training
+```
+pip install -e ".[train]"
+```
+
+### Upgrade to latest code base
+
+```Shell
+git pull
+pip install -e .
+```
+
+## Quick Start - Image2SVG Generation
+
+```Python
+from PIL import Image
+from starvector.model.starvector_arch import StarVectorForCausalLM
+from starvector.data.util import process_and_rasterize_svg
+
+model_name = "starvector/starvector-8b-im2svg"
+
+starvector = StarVectorForCausalLM.from_pretrained(model_name)
+
+starvector.cuda()
+starvector.eval()
+
+image_pil = Image.open('assets/examples/sample-0.png')
+image = starvector.process_images([image_pil])[0].cuda()
+batch = {"image": image}
+
+raw_svg = starvector.generate_im2svg(batch, max_length=1000)[0]
+svg, raster_image = process_and_rasterize_svg(raw_svg)
+```
+
+### Use it from HuggingFace AutoModel
+
+```Python
+from PIL import Image
+from transformers import AutoModelForCausalLM, AutoTokenizer, AutoProcessor
+from starvector.data.util import process_and_rasterize_svg
+import torch
+
+model_name = "starvector/starvector-8b-im2svg"
+
+starvector = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, trust_remote_code=True)
+processor = starvector.model.processor
+tokenizer = starvector.model.svg_transformer.tokenizer
+
+starvector.cuda()
+starvector.eval()
+
+image_pil = Image.open('assets/examples/sample-18.png')
+
+image = processor(image_pil, return_tensors="pt")['pixel_values'].cuda()
+if not image.shape[0] == 1:
+    image = image.squeeze(0)
+batch = {"image": image}
+
+raw_svg = starvector.generate_im2svg(batch, max_length=4000)[0]
+svg, raster_image = process_and_rasterize_svg(raw_svg)
+```
+
+
+## Models
+
+We provide [Hugging Face 🤗 model checkpoints](https://huggingface.co/collections/starvector/starvector-models-6783b22c7bd4b43d13cb5289) for image2SVG vectorization, for 💫 StarVector-8B and 💫 StarVector-1B. These are the results on SVG-Bench, using the DinoScore metric.
+
+| Method        | SVG-Stack | SVG-Fonts | SVG-Icons | SVG-Emoji | SVG-Diagrams |
+|---------------|-----------|-----------|-----------|-----------|--------------|
+| AutoTrace    | 0.942     | 0.954     | 0.946     | 0.975     | 0.874        |
+| Potrace      | 0.898     | 0.967     | 0.972     | 0.882     | 0.875        |
+| VTracer      | 0.954     | 0.964     | 0.940     | 0.981     | 0.882        |
+| Im2Vec        | 0.692     | 0.733     | 0.754     | 0.732     | -            |
+| LIVE          | 0.934     | 0.956     | 0.959     | 0.969     | 0.870        |
+| DiffVG        | 0.810     | 0.821     | 0.952     | 0.814     | 0.822        |
+| GPT-4-V       | 0.852     | 0.842     | 0.848     | 0.850     | -            |
+| 💫 StarVector-1B (🤗 [Link](https://huggingface.co/starvector/starvector-1b-im2svg)) | 0.926     | 0.978     | 0.975     | 0.929     | 0.943        |
+| 💫 StarVector-8B (🤗 [Link](https://huggingface.co/starvector/starvector-8b-im2svg)) | **0.966** | **0.982** | **0.984** | **0.981** | **0.959**    |
+
+*Note*: StarVector models will not work for natural images or illustrations, as they have not been trained on those images. They excel in vectorizing icons, logotypes, technical diagrams, graphs, and charts.
+
+## Datasets - SVG-Bench
+SVG-Bench is a benchmark for evaluating SVG generation models. It contains 10 datasets, and 3 tasks: Image-to-SVG, Text-to-SVG, and Diagram-to-SVG.
+
+See our [Huggingface 🤗 Dataset Collection](https://huggingface.co/collections/starvector/starvector-svg-datasets-67811204a76475be4dd66d09)  
+
+| Dataset         |  Train  | Val   | Test | Token Length     | SVG Primitives | Annotation     |
+|-----------------|--------|-------|------|------------------|----------------|----------------|
+| SVG-Stack (🤗 [Link](https://huggingface.co/datasets/starvector/svg-stack)) | 2.1M   | 108k  | 5.7k | 1,822 ± 1,808    | All            | [Captions](https://huggingface.co/datasets/starvector/text2svg-stack)        |
+| SVG-Stack_sim (🤗 [Link](https://huggingface.co/datasets/starvector/svg-stack-simple)) | 601k   | 30.1k | 1.5k | 2k ± 918         | Vector path    | -        |
+| SVG-Diagrams (🤗 [Link](https://huggingface.co/datasets/starvector/svg-diagrams)) | -      | -     | 472  | 3,486 ± 1,918    | All            | -        |
+| SVG-Fonts (🤗 [Link](https://huggingface.co/datasets/starvector/svg-fonts)) | 1.8M   | 91.5k | 4.8k | 2,121 ± 1,868    | Vector path    | Font letter      |
+| SVG-Fonts_sim (🤗 [Link](https://huggingface.co/datasets/starvector/svg-fonts-simple)) | 1.4M   | 71.7k | 3.7k | 1,722 ± 723      | Vector path    | Font letter      |
+| SVG-Emoji (🤗 [Link](https://huggingface.co/datasets/starvector/svg-emoji)) | 8.7k   | 667   | 668  | 2,551 ± 1,805    | All            | -          |
+| SVG-Emoji_sim (🤗 [Link](https://huggingface.co/datasets/starvector/svg-emoji-simple)) | 580    | 57    | 96   | 2,448 ± 1,026    | Vector Path    | -          |
+| SVG-Icons (🤗 [Link](https://huggingface.co/datasets/starvector/svg-icons)) | 80.4k  | 6.2k  | 2.4k | 2,449 ± 1,543    | Vector path    | -              |
+| SVG-Icons_sim (🤗 [Link](https://huggingface.co/datasets/starvector/svg-icons-simple)) | 80,435 | 2,836 | 1,277| 2,005 ± 824      | Vector path    | -              |
+| SVG-FIGR (🤗 [Link](https://huggingface.co/datasets/starvector/FIGR-SVG)) | 270k   | 27k   | 3k   | 5,342 ± 2,345    | Vector path    | Class, Caption | 
+
+
+>We offer a summary of statistics about the datasets used in our training and evaluation experiments. This datasets are included in SVG-Bench. The subscript _sim_ stands for the simplified version of the dataset, as required by some baselines.
+
+## Training
+
+### Confirm dependencies are installed
+
+```bash
+pip install -e ".[train]"
+```
+
+### Set environment variables
+We recommend setting the following environment variables:
+
+```bash
+  export HF_HOME=<path to the folder where you want to store the models>
+  export HF_TOKEN=<your huggingface token>
+  export WANDB_API_KEY=<your wandb token>
+  export OUTPUT_DIR=<path/to/output>
+```
+
+cd the root of the repository.
+
+```Shell
+cd star-vector
+```
+
+### Image2SVG Pretraining (Stage 1)
+
+We have different training approaches for StarVector-1B and StarVector-8B. StarVector-1B can be trained using Deepspeed, while StarVector-8B requires FSDP.
+
+#### StarVector-1B Training
+
+You can use the following command to train StarVector-1B on SVG-Stack for the Image2SVG vectorization task, using Deepspeed and Accelerate
+
+```bash
+# StarVector-1B
+accelerate launch --config_file configs/accelerate/deepspeed-8-gpu.yaml starvector/train/train.py config=configs/models/starvector-1b/im2svg-stack.yaml
+```
+
+#### StarVector-8B Training
+
+You can use the following command to train StarVector-8B on SVG-Stack for the Image2SVG vectorization task, using FSDP and Accelerate. We provide the torchrun command to support multi-nodes and multi-GPUs.
+
+```bash
+# StarVector-8B
+torchrun \
+  --nproc-per-node=8 \
+  --nnodes=1 \
+  starvector/train/train.py \
+  config=configs/models/starvector-8b/im2svg-stack.yaml
+```
+
+
+### Finetuning StarVector (Stage 2)
+
+After pretraining StarVector on image vectorization, we finetune it on additional SVG tasks like Text2SVG, and SVG-Bench datasets.
+
+#### Text2SVG Finetuning
+
+```bash
+# StarVector-1B
+accelerate launch --config_file config/accelerate/deepspeed-8-gpu.yaml starvector/train/train.py config=configs/models/starvector-1b/text2svg-stack.yaml
+
+# StarVector-8B
+torchrun \
+  --nproc-per-node=8 \
+  --nnodes=1 \
+  starvector/train/train.py \
+  config=configs/models/starvector-8b/text2svg-stack.yaml
+```
+
+#### SVG-Bench Finetuning
+
+```bash
+# StarVector-1B
+accelerate launch --config_file config/accelerate/deepspeed-8-gpu.yaml starvector/train/train.py config=configs/models/starvector-1b/im2svg-{fonts,icons,emoji}.yaml
+
+# StarVector-8B
+torchrun \
+  --nproc-per-node=8 \
+  --nnodes=1 \
+  starvector/train/train.py \
+  config=configs/models/starvector-8b/im2svg-{fonts,icons,emoji}.yaml
+```
+
+We also provide shell scripts in `scripts/train/*` 
+
+## Validation on SVG Benchmarks (⭐ SVG-Bench)
+
+We validate StarVector on ⭐ SVG-Bench Benchmark. We provide the SVGValidator class that allows you to run StarVector using **1) the HuggingFace generation backend** or **2) the VLLM backend**. The later is substantially faster thanks to the use of Paged Attention. 
+
+### HuggingFace Generation Backend
+Let's start with the evaluation for StarVector-1B and StarVector-8B on SVG-Stack, using the HuggingFace generation backend (StarVectorHFAPIValidator). To override the input arguments, you can add cli args following the yaml file structure.
+
+```bash
+# StarVector-1B on SVG-Stack, using the HuggingFace backend 
+python starvector/validation/validate.py \
+config=configs/generation/hf/starvector-1b/im2svg.yaml \
+dataset.name=starvector/svg-stack
+
+# StarVector-8B on SVG-Stack, using the vanilla HuggingFace generation API
+python starvector/validation/validate.py \
+config=configs/generation/hf/starvector-8b/im2svg.yaml \
+dataset.name=starvector/svg-stack
+```
+
+### vLLM Backend
+
+For using the vLLM backend (StarVectorVLLMAPIValidator), first install our StarVector fork of VLLM, [here](https://github.com/starvector/vllm).
+
+```bash
+git clone https://github.com/starvector/vllm.git
+cd vllm
+pip install -e .
+```
+
+Then, launch the using the vllm config file (it uses StarVectorVLLMValidator):
+
+```bash
+# StarVector-1B
+python starvector/validation/validate.py \
+config=configs/generation/vllm/starvector-1b/im2svg.yaml \
+dataset.name=starvector/svg-stack
+
+# StarVector-8B
+python starvector/validation/validate.py \
+config=configs/generation/vllm/starvector-8b/im2svg.yaml \
+dataset.name=starvector/svg-stack
+```
+
+#### Generate using Temperature Sweep
+Temperature sweep is an evaluation technique where we:
+1. Generate multiple SVG candidates using different temperature values (controlling randomness in generation)
+2. Evaluate each candidate using the DinoScore metric
+3. Select the best performing SVG as the final output
+
+This approach improves result quality by exploring multiple generation possibilities, though it requires more computation time.
+
+
+```bash
+# StarVector-1B (vLLM)
+python starvector/validation/run_validator.py \
+config=configs/generation/vllm/starvector-1b/im2svg.yaml \
+dataset.name=svg-stack \
+generation_params.generation_sweep=True \
+generation_params.num_generations_different_temp=5 \
+generation_params.min_temperature=0.0 \
+generation_params.max_temperature=0.5
+
+# StarVector-8B (vLLM)
+python starvector/validation/run_validator.py \
+config=configs/generation/vllm/starvector-8b/im2svg.yaml \
+dataset.name=svg-stack \
+generation_params.generation_sweep=True \
+generation_params.num_generations_different_temp=10 \
+generation_params.min_temperature=0.0 \
+generation_params.max_temperature=0.5
+    
+```
+
+We provide evaluation scripts in `scripts/eval/*`
+
+
+## StarVector Demo
+
+The demo provides two options for converting images to SVG code:
+1. HuggingFace generation functionality
+2. VLLM (recommended) - offers faster generation speed
+
+### Option 1: HuggingFace Generation with Gradio Web UI
+
+We provide a Gradio web UI for you to play with our model.
+
+#### Launch a controller
+```Shell
+python -m starvector.serve.controller --host 0.0.0.0 --port 10000
+```
+
+#### Launch a gradio web server.
+```Shell
+python -m starvector.serve.gradio_web_server --controller http://localhost:10000 --model-list-mode reload --port 7000
+```
+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
+
+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 starvector.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path joanrodai/starvector-1.4b
+```
+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
+vllm serve starvector/starvector-8b-im2svg --chat-template configs/chat-template.jinja --trust-remote-code --port 8001 --max-model-len 16000
+
+python -m starvector.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>
+```
+
+#### Option 2: Launch VLLM
+
+0. Remember to clone the starvector/vllm fork (it has modifications for starvector).
+
+```Shell
+git clone https://github.com/starvector/vllm.git
+cd vllm
+pip install -e .
+```
+
+1. Call this to launch the VLLM endpoint
+
+
+```Shell
+vllm serve starvector/starvector-1b-im2svg --chat-template configs/chat-template.jinja --trust-remote-code --port 8000 --max-model-len 8192
+```
+
+2. Create the demo for VLLM
+
+```Shell
+python -m starvector.serve.vllm_api_gradio.controller --host 0.0.0.0 --port 10000
+python -m starvector.serve.vllm_api_gradio.gradio_web_server --controller http://localhost:10000 --model-list-mode reload --port 7000
+python -m starvector.serve.vllm_api_gradio.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-name starvector/starvector-1b-im2svg --vllm-base-url http://localhost:8000
+```
+
+3. Add more models by serving them with VLLM and calling a new model worker
+
+```Shell
+vllm serve starvector/starvector-8b-im2svg --chat-template configs/chat-template.jinja --trust-remote-code --port 8001 --max-model-len 16384
+
+python -m starvector.serve.vllm_api_gradio.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40001 --worker http://localhost:40001 --model-name starvector/starvector-8b-im2svg --vllm-base-url http://localhost:8001
+```
+
+## Citation
+```
+@misc{rodriguez2024starvector,
+      title={StarVector: Generating Scalable Vector Graphics Code from Images and Text}, 
+      author={Juan A. Rodriguez and Abhay Puri and Shubham Agarwal and Issam H. Laradji and Pau Rodriguez and Sai Rajeswar and David Vazquez and Christopher Pal and Marco Pedersoli},
+      year={2024},
+      eprint={2312.11556},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2312.11556}, 
+}
+```
+
+## License
+This project is licensed under the Apache License, Version 2.0 - see the [LICENSE](LICENSE) file for details.

configs/accelerate/1-gpu.yaml

@@ -0,0 +1,16 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config: {}
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+dynamo_backend: 'NO'
+fsdp_config: {}
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+megatron_lm_config: {}
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+use_cpu: false

configs/accelerate/2-gpu.yaml

@@ -0,0 +1,16 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config: {}
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+dynamo_backend: 'NO'
+fsdp_config: {}
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+megatron_lm_config: {}
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 2
+rdzv_backend: static
+same_network: true
+use_cpu: false

configs/accelerate/4-gpu.yaml

@@ -0,0 +1,16 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config: {}
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+dynamo_backend: 'NO'
+fsdp_config: {}
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+megatron_lm_config: {}
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 4
+rdzv_backend: static
+same_network: true
+use_cpu: false

configs/accelerate/8-gpu.yaml

@@ -0,0 +1,16 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config: {}
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+dynamo_backend: 'NO'
+fsdp_config: {}
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+megatron_lm_config: {}
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+use_cpu: false

configs/accelerate/deepspeed-1-gpu.yaml

@@ -0,0 +1,21 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config:
+  gradient_clipping: 1.0
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+fsdp_config: {}
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

configs/accelerate/deepspeed-2-gpu.yaml

@@ -0,0 +1,21 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config:
+  gradient_clipping: 1.0
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+fsdp_config: {}
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 2
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

configs/accelerate/deepspeed-4-gpu.yaml

@@ -0,0 +1,21 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config:
+  gradient_clipping: 1.0
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+fsdp_config: {}
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 4
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

configs/accelerate/deepspeed-8-gpu.yaml

@@ -0,0 +1,21 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config:
+  gradient_clipping: 1.0
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+fsdp_config: {}
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

configs/accelerate/deespeed.json

@@ -0,0 +1,29 @@
+{
+    "bf16": {
+        "enabled": false
+    },
+    "zero_optimization": {
+        "stage": 3,
+        "offload_optimizer": {
+            "device": "cpu"
+        },
+        "offload_param": {
+            "device": "cpu"
+        },
+        "overlap_comm": true,
+        "contiguous_gradients": true,
+        "reduce_bucket_size": "auto",
+        "stage3_prefetch_bucket_size": "auto",
+        "stage3_param_persistence_threshold": "auto",
+        "sub_group_size": 1e9,
+        "stage3_max_live_parameters": 1e9,
+        "stage3_max_reuse_distance": 1e9,
+        "stage3_gather_16bit_weights_on_model_save": true
+    },
+    "gradient_accumulation_steps": 4,
+    "gradient_clipping": "auto",
+    "steps_per_print": 2000,
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+}

configs/accelerate/val-deepspeed-1-gpu.yaml

@@ -0,0 +1,21 @@
+compute_environment: LOCAL_MACHINE
+deepspeed_config:
+  gradient_clipping: 1.0
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: false
+  zero_stage: 1
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+fsdp_config: {}
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

configs/chat-template.jinja

@@ -0,0 +1 @@
+{% for message in messages %}{{ message.content }}{% endfor %}

configs/generation/hf/starvector-1b/im2svg.yaml

@@ -0,0 +1,52 @@
+# General configuration
+run:
+  project_name: "starvector-RL-eval"
+  out_dir: "/mnt/starvector/RL/logs/eval"
+  device: cuda
+  report_to: wandb
+  run_id: test-run
+  log_images: false
+
+# Model configuration
+model:
+  name: "starvector/starvector-1b-im2svg"   # Required: Model name for HF-based model
+  from_checkpoint: false
+  generation_engine: "hf"
+  task: im2svg
+  torch_dtype: float16
+  # image_processor: clip # is this needed?
+
+# Dataset configuration
+dataset:
+  dataset_name: starvector/svg-stack-RL  # Required: Name of the dataset to evaluate on
+  config_name: null # in bigodcs set Image2SVG
+  split: test
+  batch_size: 8
+  num_workers: 4
+  im_size: 224
+  num_samples: -1
+
+# vllm https://docs.vllm.ai/en/v0.5.5/dev/sampling_params.html
+# hf https://huggingface.co/docs/transformers/main_classes/text_generation
+generation_params:
+  # Text generation parameters
+  max_length: 7800
+  min_length: 10
+  num_beams: 1
+  temperature: 0.2
+  generation_sweep: false # Controls multi-temperature sampling, rank based sampling
+    # num_generations_different_temp: 1
+    # min_temperature: 0.0
+    # max_temperature: 0.5
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 1.0
+  presence_penalty: 0.0 # only used in vllm 
+  frequency_penalty: 0.0
+  top_p: 0.95
+  do_sample: true # turn this off for greedy decoding
+  use_nucleus_sampling: true
+  logit_bias: 5 # if this is not false, the model will be biased to the svg_end_token_id
+  stream: false
+
+

configs/generation/hf/starvector-1b/text2svg.yaml

@@ -0,0 +1,44 @@
+# General configuration
+run:
+  project_name: "starvector-RL-eval"
+  out_dir: "eval_results"
+  device: cuda
+  report_to: wandb
+  run_id: test-run
+  log_images: false
+
+# Model configuration
+model:
+  name: "starvector/starvector-1b-text2svg"
+  generation_engine: "hf"
+  task: text2svg
+  torch_dtype: bfloat16
+  image_processor: null
+
+# Dataset configuration
+dataset:
+  name: svg-stack
+  batch_size: 2
+  num_workers: 4
+  num_samples: -1
+
+# vllm https://docs.vllm.ai/en/v0.5.5/dev/sampling_params.html
+# hf https://huggingface.co/docs/transformers/main_classes/text_generation
+generation_params:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  presence_penalty: 0.0 # only used in vllm 
+  frequency_penalty: 0.0
+  top_p: 0.95
+  do_sample: true # turn this off for greedy decoding
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+  logit_bias: 10 # if this is not false, the model will be biased to the svg_end_token_id
+  stream: false

configs/generation/hf/starvector-8b/im2svg.yaml

@@ -0,0 +1,50 @@
+# General configuration
+run:
+  project_name: "starvector-RL-eval"
+  out_dir: "eval_results"
+  device: cuda
+  report_to: wandb
+  run_id: test-run
+  log_images: false
+
+# Model configuration
+model:
+  name: "starvector/starvector-8b-im2svg"
+  from_checkpoint: false
+  generation_engine: "hf"
+  task: im2svg
+  torch_dtype: bfloat16
+  # image_processor: siglip_384
+
+# Dataset configuration
+dataset:
+  dataset_name: starvector/svg-stack-RL
+  config_name: null
+  split: test
+  batch_size: 2
+  num_workers: 4
+  im_size: 384
+  num_samples: -1
+
+# vllm https://docs.vllm.ai/en/v0.5.5/dev/sampling_params.html
+# hf https://huggingface.co/docs/transformers/main_classes/text_generation
+generation_params:
+  max_length: 16000
+  min_length: 10
+  num_beams: 1
+  temperature: 0.7
+  generation_sweep: false # Controls multi-temperature sampling, rank based sampling
+    # num_generations_different_temp: 1
+    # min_temperature: 0.0
+    # max_temperature: 0.5
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  presence_penalty: 0.0 # only used in vllm 
+  frequency_penalty: 0.0
+  top_p: 0.95
+  do_sample: true # turn this off for greedy decoding
+  use_nucleus_sampling: true
+  logit_bias: 5 # if this is not false, the model will be biased to the svg_end_token_id
+  stream: false
+

configs/generation/hf/starvector-8b/text2svg.yaml

@@ -0,0 +1,40 @@
+# General configuration
+run:
+  out_dir: "eval_results"
+  device: cuda
+
+# Model configuration
+model:
+  name: "starvector/starvector-8b-text2svg"
+  generation_engine: "hf" 
+  task: text2svg
+  torch_dtype: bfloat16
+  image_processor: clip
+
+# Dataset configuration
+dataset:
+  name: svg-stack
+  batch_size: 2
+  num_workers: 4
+  num_samples: -1
+
+# vllm https://docs.vllm.ai/en/v0.5.5/dev/sampling_params.html
+# hf https://huggingface.co/docs/transformers/main_classes/text_generation
+generation_params:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  presence_penalty: 0.0 # only used in vllm 
+  frequency_penalty: 0.0
+  top_p: 0.95
+  do_sample: true # turn this off for greedy decoding
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+  logit_bias: 10 # if this is not false, the model will be biased to the svg_end_token_id
+  stream: false

configs/generation/vllm/starvector-1b/im2svg.yaml

@@ -0,0 +1,51 @@
+# General configuration
+run:
+  project_name: "starvector-RL-eval"
+  out_dir: "/mnt/starvector/RL/logs/eval"
+  report_to: wandb
+  run_id: test-eval3
+  log_images: false
+
+# Model configuration
+model:
+  name: "starvector/starvector-1b-im2svg"   # Required: Model name for HF-based model
+  from_checkpoint: false
+  generation_engine: "vllm"
+  task: im2svg
+  torch_dtype: float16
+  # image_processor: clip # is this needed?
+
+# Dataset configuration
+dataset:
+  dataset_name: starvector/svg-stack-RL  # Required: Name of the dataset to evaluate on
+  config_name: null # in bigodcs set Image2SVG
+  split: test
+  batch_size: 8
+  num_workers: 8
+  im_size: 224
+  num_samples: 500
+
+# vllm https://docs.vllm.ai/en/v0.5.5/dev/sampling_params.html
+# hf https://huggingface.co/docs/transformers/main_classes/text_generation
+generation_params:
+  # Text generation parameters
+  max_length: 7933 # 8192 - (visual tokens)
+  num_generations: 1
+  min_length: 10
+  num_beams: 1
+  temperature: 0.5
+  generation_sweep: false # Controls multi-temperature sampling, rank based sampling
+    # num_generations_different_temp: 5
+    # min_temperature: 0.0
+    # max_temperature: 0.5
+  num_captions: 1
+  frequency_penalty: 0.0
+  presence_penalty: 0.0
+  repetition_penalty: 1.0
+  top_p: 0.9
+  logit_bias: 5 # if this is not false, the model will be biased to the svg_end_token_id
+  min_p: 0.0
+  top_k: -1
+  stream: false
+
+

configs/generation/vllm/starvector-1b/text2svg.yaml

@@ -0,0 +1,40 @@
+# General configuration
+run:
+  out_dir: "eval_results"
+  api:
+    key: "EMPTY"
+    base_url: "http://0.0.0.0:40000/v1"
+
+# Model configuration
+model:
+  name: "starvector/starvector-1b-text2svg"
+  generation_engine: "vllm"
+  task: text2svg
+  torch_dtype: bfloat16
+  image_processor: null
+
+# Dataset configuration
+dataset:
+  name: svg-stack
+  batch_size: 2
+  num_workers: 4
+  num_samples: -1
+
+# vllm https://docs.vllm.ai/en/v0.5.5/dev/sampling_params.html
+# hf https://huggingface.co/docs/transformers/main_classes/text_generation
+generation_params:
+  # Text generation parameters
+  max_length: 16384
+  min_length: 10
+  num_beams: 1
+  temperature: 0.6
+  generation_sweep: false # Controls multi-temperature sampling, rank based sampling
+    # num_generations_different_temp: 1
+    # min_temperature: 0.0
+    # max_temperature: 0.5
+  num_captions: 1
+  frequency_penalty: 0.0
+  presence_penalty: 0.0
+  top_p: 0.9
+  logit_bias: False # if this is not false, the model will be biased to the svg_end_token_id
+  stream: false

configs/generation/vllm/starvector-8b/im2svg.yaml

@@ -0,0 +1,45 @@
+# General configuration
+run:
+  out_dir: "eval_results"
+  api:
+    key: "EMPTY"
+    base_url: "http://0.0.0.0:40000/v1"
+
+# Model configuration
+model:
+  name: "starvector/starvector-8b-im2svg"
+  generation_engine: "vllm"
+  task: im2svg
+  torch_dtype: bfloat16
+  image_processor: siglip_384
+
+# Dataset configuration
+dataset:
+  name: svg-stack
+  batch_size: 2
+  num_workers: 4
+  im_size: 384
+  num_samples: -1
+  dpi: 2
+  scale: 300
+
+
+# vllm https://docs.vllm.ai/en/v0.5.5/dev/sampling_params.html
+# hf https://huggingface.co/docs/transformers/main_classes/text_generation
+generation_params:
+  # Text generation parameters
+  max_length: 15806 # 16384 - (visual tokens)
+  min_length: 10
+  num_beams: 1
+  temperature: 0.6
+  generation_sweep: false # Controls multi-temperature sampling, rank based sampling
+    # num_generations_different_temp: 1
+    # min_temperature: 0.0
+    # max_temperature: 0.5
+  num_captions: 1
+  frequency_penalty: 0.0
+  presence_penalty: 0.0
+  top_p: 0.9
+  logit_bias: False # if this is not false, the model will be biased to the svg_end_token_id
+  stream: false
+

configs/generation/vllm/starvector-8b/text2svg.yaml

@@ -0,0 +1,41 @@
+# General configuration
+run:
+  out_dir: "eval_results"
+  api:
+    key: "EMPTY"
+    base_url: "http://0.0.0.0:40000/v1"
+
+# Model configuration
+model:
+  name: "starvector/starvector-8b-text2svg"
+  generation_engine: "vllm"  
+  task: text2svg
+  torch_dtype: bfloat16
+  image_processor: clip
+
+# Dataset configuration
+dataset:
+  name: svg-stack
+  batch_size: 2
+  num_workers: 4
+  num_samples: -1
+
+
+# vllm https://docs.vllm.ai/en/v0.5.5/dev/sampling_params.html
+# hf https://huggingface.co/docs/transformers/main_classes/text_generation
+generation_params:
+  # Text generation parameters
+  max_length: 16384
+  min_length: 10
+  num_beams: 1
+  temperature: 0.6
+  generation_sweep: false # Controls multi-temperature sampling, rank based sampling
+    # num_generations_different_temp: 1
+    # min_temperature: 0.0
+    # max_temperature: 0.5
+  num_captions: 1
+  frequency_penalty: 0.0
+  presence_penalty: 0.0
+  top_p: 0.9
+  logit_bias: False # if this is not false, the model will be biased to the svg_end_token_id
+  stream: false

configs/metrics/im2svg.yaml

@@ -0,0 +1,12 @@
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/metrics/text2svg.yaml

@@ -0,0 +1,12 @@
+metrics:
+  L2: false
+  Masked-L2: false
+  LPIPS: false
+  SSIM: false
+  FID: true
+  FID_clip: true
+  CLIPScore: true
+  CountTokenLength: true
+  ratio_post_processed: true
+  ratio_non_compiling: true
+  DinoScore: false

configs/models/default.yaml

@@ -0,0 +1,93 @@
+project:
+  project: starvector-im2svg
+  use_wandb: false
+  entity: abc
+  copy_code: false
+model:
+  max_length: 8192
+  model_name: null # in case of creating a new model, set this to None (null)
+  starcoder_model_name: bigcode/starcoderbase-1b
+  pretrained: true
+  image_encoder_type: clip
+  use_flash_attn: true
+  adapter_norm: layer_norm
+  init_type: glorot
+  dropout: 0.1
+  task: im2svg
+  transformer_layer_cls: None # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 3
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 4
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  start_generation_at_step: 0
+  train_image_encoder: true
+  train_LLM: true
+  use_gradient_checkpointing: false
+fsdp:
+  enable: false
+data:
+  num_workers: 4
+  train:
+    batch_size: 2
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: train
+      dataset_name: ServiceNow/svg-stack
+      im_size: 224
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+  test:
+    batch_size: 2
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: test
+      dataset_name: ServiceNow/svg-stack
+      im_size: 224
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+generation:
+  max_length: 8192
+  min_length: 10
+  num_beams: 3
+  num_captions: 1
+  num_generations_different_temp: 1.5
+  start_temperature: 0.5
+  repetition_penalty: 1.0
+  length_penalty: 1.0
+  temperature: 1.0
+  top_p: 0.9
+  use_nucleus_sampling: true
+  im_size: 224
+  dpi: 2
+  scale: 300
+  num_samples_to_generate: -1
+  log_wandb_images: true
+  start_generation_at_step: -1
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/models/starvector-1b/im2svg-emoji.yaml

@@ -0,0 +1,87 @@
+project:
+  project: starvector-1b-im2svg
+  use_wandb: false
+  entity: abc
+  copy_code: false
+model:  
+  max_length: 8192
+  model_name: null # in case of creating a new model, set this to None (null)
+  starcoder_model_name: bigcode/starcoderbase-1b
+  pretrained: true
+  image_encoder_type: clip
+  use_flash_attn: true
+  adapter_norm: batch_norm
+  init_type: glorot
+  dropout: 0.1
+  task: im2svg
+  transformer_layer_cls: null # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 3
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 10
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  train_image_encoder: true
+  train_LLM: true
+  use_gradient_checkpointing: false
+fsdp:
+  enable: false
+data:
+  num_workers: 16
+  train:
+    batch_size: 2
+    target: starvector.data.emojisvg.EmojiSVGDataset
+    params:
+      split: train
+      dataset_name: starvector/svg-emoji
+      im_size: 224
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+  test:
+    batch_size: 8
+    target: starvector.data.emojisvg.EmojiSVGDataset
+    params:
+      split: test
+      dataset_name: starvector/svg-emoji
+      im_size: 224
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+generation:
+  max_length: 8192
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 224
+  dpi: 2
+  scale: 300
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/models/starvector-1b/im2svg-fonts.yaml

@@ -0,0 +1,87 @@
+project:
+  project: starvector-1b-im2svg
+  use_wandb: false
+  entity: abc
+  copy_code: false
+model:  
+  max_length: 8192
+  model_name: null # in case of creating a new model, set this to None (null)
+  starcoder_model_name: bigcode/starcoderbase-1b
+  pretrained: true
+  image_encoder_type: clip
+  use_flash_attn: true
+  adapter_norm: batch_norm
+  init_type: glorot
+  dropout: 0.1
+  task: im2svg
+  transformer_layer_cls: null # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 3
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 4
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  train_image_encoder: true
+  train_LLM: true
+  use_gradient_checkpointing: false
+fsdp:
+  enable: false
+data:
+  num_workers: 16
+  train:
+    batch_size: 4
+    target: starvector.data.fontsvg.FontSVGDataset
+    params:
+      split: train
+      dataset_name: starvector/svg-fonts
+      im_size: 224
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+  test:
+    batch_size: 8
+    target: starvector.data.fontsvg.FontSVGDataset
+    params:
+      split: test
+      dataset_name: starvector/svg-fonts
+      im_size: 224
+      num_samples: 1000
+      transforms: false
+      select_dataset_name: false
+generation:
+  max_length: 8192
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 224
+  dpi: 2
+  scale: 300
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/models/starvector-1b/im2svg-icons.yaml

@@ -0,0 +1,87 @@
+project:
+  project: starvector-1b-im2svg
+  use_wandb: false
+  entity: abc
+  copy_code: false
+model:  
+  max_length: 8192
+  model_name: null # in case of creating a new model, set this to None (null)
+  starcoder_model_name: bigcode/starcoderbase-1b
+  pretrained: true
+  image_encoder_type: clip
+  use_flash_attn: true
+  adapter_norm: batch_norm
+  init_type: glorot
+  dropout: 0.1
+  task: im2svg
+  transformer_layer_cls: null # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 3
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 10
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  train_image_encoder: true
+  train_LLM: true
+  use_gradient_checkpointing: false
+fsdp:
+  enable: false
+data:
+  num_workers: 16
+  train:
+    batch_size: 4
+    target: starvector.data.iconsvg.SVGIconsDataset
+    params:
+      split: train
+      dataset_name: starvector/svg-icons
+      im_size: 224
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+  test:
+    batch_size: 8
+    target: starvector.data.iconsvg.SVGIconsDataset
+    params:
+      split: test
+      dataset_name: starvector/svg-icons
+      im_size: 224
+      num_samples: 1000
+      transforms: false
+      select_dataset_name: false
+generation:
+  max_length: 8192
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 224
+  dpi: 2
+  scale: 300
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/models/starvector-1b/im2svg-stack.yaml

@@ -0,0 +1,87 @@
+project:
+  project: starvector-1b-im2svg
+  use_wandb: false
+  entity: abc
+  copy_code: false
+model:  
+  max_length: 8192
+  model_name: null # in case of creating a new model, set this to None (null)
+  starcoder_model_name: bigcode/starcoderbase-1b
+  pretrained: true
+  image_encoder_type: clip
+  use_flash_attn: true
+  adapter_norm: batch_norm
+  init_type: glorot
+  dropout: 0.1
+  task: im2svg
+  transformer_layer_cls: null # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 3
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 4
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  train_image_encoder: true
+  train_LLM: true
+  use_gradient_checkpointing: false
+fsdp:
+  enable: false
+data:
+  num_workers: 4
+  train:
+    batch_size: 2
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: train
+      dataset_name: starvector/svg-stack
+      im_size: 224
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+  test:
+    batch_size: 2
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: test
+      dataset_name: starvector/svg-stack
+      im_size: 224
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+generation:
+  max_length: 8192
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 224
+  dpi: 2
+  scale: 300
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/models/starvector-1b/text2svg-figr.yaml

@@ -0,0 +1,94 @@
+project:
+  project: starvector-1b-text2svg
+  use_wandb: false
+  entity: joanrod
+  copy_code: false
+model:
+  max_length: 8192
+  model_name: starvector/starvector-1b-im2svg
+  starcoder_model_name: bigcode/starcoderbase-1b
+  pretrained: true
+  image_encoder_type: clip
+  use_flash_attn: true
+  adapter_norm: layer_norm
+  init_type: normal
+  dropout: 0.1
+  task: text2svg
+  transformer_layer_cls: Starcoder2DecoderLayer # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 5
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 2
+  lr: 2e-5
+  gradient_accumulation_steps: 8
+  lr_scheduler: cosine
+  lr_warmup_steps: 100
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  use_gradient_checkpointing: false
+  train_image_encoder: false
+  train_LLM: true
+fsdp:
+  enable: false # TODO: set this reasonably, i.e., false only if you want to use DDP or have PyTorch < 2.1
+  cpu_offload: false
+  sharding_strategy: hsdp
+  backward_prefetch: BACKWARD_PRE
+  use_orig_params: true
+  sync_module_states: true
+  forward_prefetch: false
+  cpu_ram_efficient_loading: true
+data:
+  num_workers: 16
+  train:
+    batch_size: 4
+    target: starvector.data.figrsvg.FigrSVGDataset
+    params:
+      split: train
+      dataset_name: starvector/FIGR-SVG
+      im_size: 224
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+  test:
+    batch_size: 8
+    target: starvector.data.figrsvg.FigrSVGDataset
+    params:
+      split: test
+      dataset_name: starvector/FIGR-SVG
+      im_size: 224 
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+generation:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+metrics:
+  L2: false
+  Masked-L2: false
+  LPIPS: false
+  SSIM: false
+  FID: false
+  FID_clip: false
+  CLIPScore: true
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: false

configs/models/starvector-1b/text2svg-stack.yaml

@@ -0,0 +1,96 @@
+project:
+  project: starvector-1b-text2svg
+  use_wandb: false
+  entity: joanrod
+  copy_code: false
+model:
+  max_length: 8192
+  model_name: starvector/starvector-1b-im2svg
+  starcoder_model_name: bigcode/starcoderbase-1b
+  pretrained: true
+  image_encoder_type: clip
+  use_flash_attn: true
+  adapter_norm: layer_norm
+  init_type: normal
+  dropout: 0.1
+  task: text2svg
+  transformer_layer_cls: Starcoder2DecoderLayer # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 5
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 2
+  lr: 2e-5
+  gradient_accumulation_steps: 8
+  lr_scheduler: cosine
+  lr_warmup_steps: 100
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  use_gradient_checkpointing: false
+  train_image_encoder: false
+  train_LLM: true
+fsdp:
+  enable: false # TODO: set this reasonably, i.e., false only if you want to use DDP or have PyTorch < 2.1
+  cpu_offload: false
+  sharding_strategy: hsdp
+  backward_prefetch: BACKWARD_PRE
+  use_orig_params: true
+  sync_module_states: true
+  forward_prefetch: false
+  cpu_ram_efficient_loading: true
+data:
+  num_workers: 16
+  train:
+    batch_size: 1
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: train
+      dataset_name: starvector/text2svg-stack
+      im_size: 224
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+      image_processor: siglip_384
+  test:
+    batch_size: 4
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: test
+      dataset_name: starvector/text2svg-stack
+      im_size: 224
+      num_samples: 64
+      transforms: false
+      select_dataset_name: false
+      image_processor: siglip_384
+generation:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+metrics:
+  L2: false
+  Masked-L2: false
+  LPIPS: false
+  SSIM: false
+  FID: false
+  FID_clip: false
+  CLIPScore: true
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: false

configs/models/starvector-8b/im2svg-emoji.yaml

@@ -0,0 +1,94 @@
+project:
+  project: starvector-8b-im2svg
+  use_wandb: false
+  entity: joanrod
+  copy_code: false
+model:
+  max_length: 16000
+  model_name: starvector/starvector-8b-im2svg
+  starcoder_model_name: bigcode/starcoder2-7b
+  pretrained: true
+  image_encoder_type: siglip_384
+  use_flash_attn: true
+  adapter_norm: layer_norm
+  init_type: normal
+  dropout: 0.1
+  task: im2svg
+  transformer_layer_cls: Starcoder2DecoderLayer # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 5
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 4
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  use_gradient_checkpointing: true
+  train_image_encoder: true
+  train_LLM: true
+fsdp:
+  enable: true # TODO: set this reasonably, i.e., false only if you want to use DDP or have PyTorch < 2.1
+  cpu_offload: false
+  sharding_strategy: hsdp
+  backward_prefetch: BACKWARD_PRE
+  use_orig_params: true
+  sync_module_states: true
+  forward_prefetch: false
+  cpu_ram_efficient_loading: true
+data:
+  num_workers: 16
+  train:
+    batch_size: 1
+    target: starvector.data.emojisvg.EmojiSVGDataset
+    params:
+      split: train
+      dataset_name: starvector/svg-emoji
+      im_size: 384
+      num_samples: -1
+      transforms: false
+      image_processor: siglip_384
+  test:
+    batch_size: 1
+    target: starvector.data.emojisvg.EmojiSVGDataset
+    params:
+      split: test
+      dataset_name: starvector/svg-emoji
+      im_size: 384
+      num_samples: 128
+      transforms: false
+      image_processor: siglip_384
+generation:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/models/starvector-8b/im2svg-fonts-simple.yaml

@@ -0,0 +1,94 @@
+project:
+  project: starvector-8b-im2svg
+  use_wandb: false
+  entity: joanrod
+  copy_code: false
+model:
+  max_length: 16000
+  model_name: starvector/starvector-8b-im2svg
+  starcoder_model_name: bigcode/starcoder2-7b
+  pretrained: true
+  image_encoder_type: siglip_384
+  use_flash_attn: true
+  adapter_norm: layer_norm
+  init_type: normal
+  dropout: 0.1
+  task: im2svg
+  transformer_layer_cls: Starcoder2DecoderLayer # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 5
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 4
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  use_gradient_checkpointing: true
+  train_image_encoder: true
+  train_LLM: true
+fsdp:
+  enable: true # TODO: set this reasonably, i.e., false only if you want to use DDP or have PyTorch < 2.1
+  cpu_offload: false
+  sharding_strategy: hsdp
+  backward_prefetch: BACKWARD_PRE
+  use_orig_params: true
+  sync_module_states: true
+  forward_prefetch: false
+  cpu_ram_efficient_loading: true
+data:
+  num_workers: 16
+  train:
+    batch_size: 1
+    target: starvector.data.fontsvg.FontSVGDataset
+    params:
+      split: train
+      dataset_name: starvector/svg-fonts-simple
+      im_size: 384
+      num_samples: -1
+      transforms: false
+      image_processor: siglip_384
+  test:
+    batch_size: 4
+    target: starvector.data.fontsvg.FontSVGDataset
+    params:
+      split: test
+      dataset_name: starvector/svg-fonts-simple
+      im_size: 384
+      num_samples: 128
+      transforms: false
+      image_processor: siglip_384
+generation:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/models/starvector-8b/im2svg-fonts.yaml

@@ -0,0 +1,94 @@
+project:
+  project: starvector-8b-im2svg
+  use_wandb: false
+  entity: joanrod
+  copy_code: false
+model:
+  max_length: 16000
+  model_name: starvector/starvector-8b-im2svg
+  starcoder_model_name: bigcode/starcoder2-7b
+  pretrained: true
+  image_encoder_type: siglip_384
+  use_flash_attn: true
+  adapter_norm: layer_norm
+  init_type: normal
+  dropout: 0.1
+  task: im2svg
+  transformer_layer_cls: Starcoder2DecoderLayer # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 5
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 4
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  use_gradient_checkpointing: true
+  train_image_encoder: true
+  train_LLM: true
+fsdp:
+  enable: true # TODO: set this reasonably, i.e., false only if you want to use DDP or have PyTorch < 2.1
+  cpu_offload: false
+  sharding_strategy: hsdp
+  backward_prefetch: BACKWARD_PRE
+  use_orig_params: true
+  sync_module_states: true
+  forward_prefetch: false
+  cpu_ram_efficient_loading: true
+data:
+  num_workers: 16
+  train:
+    batch_size: 1
+    target: starvector.data.fontsvg.FontSVGDataset
+    params:
+      split: train
+      dataset_name: starvector/svg-fonts
+      im_size: 384
+      num_samples: -1
+      transforms: false
+      image_processor: siglip_384
+  test:
+    batch_size: 4
+    target: starvector.data.fontsvg.FontSVGDataset
+    params:
+      split: test
+      dataset_name: starvector/svg-fonts
+      im_size: 384
+      num_samples: 128
+      transforms: false
+      image_processor: siglip_384
+generation:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/models/starvector-8b/im2svg-icons.yaml

@@ -0,0 +1,94 @@
+project:
+  project: starvector-8b-im2svg
+  use_wandb: false
+  entity: joanrod
+  copy_code: false
+model:
+  max_length: 16000
+  model_name: starvector/starvector-8b-im2svg
+  starcoder_model_name: bigcode/starcoder2-7b
+  pretrained: true
+  image_encoder_type: siglip_384
+  use_flash_attn: true
+  adapter_norm: layer_norm
+  init_type: normal
+  dropout: 0.1
+  task: im2svg
+  transformer_layer_cls: Starcoder2DecoderLayer # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 5
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 4
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  use_gradient_checkpointing: true
+  train_image_encoder: true
+  train_LLM: true
+fsdp:
+  enable: true # TODO: set this reasonably, i.e., false only if you want to use DDP or have PyTorch < 2.1
+  cpu_offload: false
+  sharding_strategy: hsdp
+  backward_prefetch: BACKWARD_PRE
+  use_orig_params: true
+  sync_module_states: true
+  forward_prefetch: false
+  cpu_ram_efficient_loading: true
+data:
+  num_workers: 16
+  train:
+    batch_size: 1
+    target: starvector.data.iconsvg.SVGIconsDataset
+    params:
+      split: train
+      dataset_name: starvector/svg-icons
+      im_size: 384
+      num_samples: -1
+      transforms: false
+      image_processor: siglip_384
+  test:
+    batch_size: 1
+    target: starvector.data.iconsvg.SVGIconsDataset
+    params:
+      split: test
+      dataset_name: starvector/svg-icons
+      im_size: 384
+      num_samples: 128
+      transforms: false
+      image_processor: siglip_384
+generation:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/models/starvector-8b/im2svg-stack.yaml

@@ -0,0 +1,96 @@
+project:
+  project: starvector-8b-im2svg
+  use_wandb: false
+  entity: joanrod
+  copy_code: false
+model:
+  max_length: 16000
+  model_name: null
+  starcoder_model_name: bigcode/starcoder2-7b
+  pretrained: true
+  image_encoder_type: siglip_384
+  use_flash_attn: true
+  adapter_norm: layer_norm
+  init_type: normal
+  dropout: 0.1
+  task: im2svg
+  transformer_layer_cls: Starcoder2DecoderLayer # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 5
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 4
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  use_gradient_checkpointing: true
+  train_image_encoder: true
+  train_LLM: true
+fsdp:
+  enable: true # TODO: set this reasonably, i.e., false only if you want to use DDP or have PyTorch < 2.1
+  cpu_offload: false
+  sharding_strategy: hsdp
+  backward_prefetch: BACKWARD_PRE
+  use_orig_params: true
+  sync_module_states: true
+  forward_prefetch: false
+  cpu_ram_efficient_loading: true
+data:
+  num_workers: 16
+  train:
+    batch_size: 1
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: train
+      dataset_name: starvector/svg-stack
+      im_size: 384
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+      image_processor: siglip_384
+  test:
+    batch_size: 2
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: test
+      dataset_name: starvector/svg-stack
+      im_size: 384
+      num_samples: 64
+      transforms: false
+      select_dataset_name: false
+      image_processor: siglip_384
+generation:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+metrics:
+  L2: true
+  Masked-L2: false
+  LPIPS: true
+  SSIM: true
+  FID: false
+  FID_clip: false
+  CLIPScore: false
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: true

configs/models/starvector-8b/text2svg-figr.yaml

@@ -0,0 +1,96 @@
+project:
+  project: starvector-8b-text2svg
+  use_wandb: false
+  entity: joanrod
+  copy_code: false
+model:
+  max_length: 16000
+  model_name: starvector/starvector-8b-im2svg
+  starcoder_model_name: bigcode/starcoder2-7b
+  pretrained: true
+  image_encoder_type: siglip_384
+  use_flash_attn: true
+  adapter_norm: layer_norm
+  init_type: normal
+  dropout: 0.1
+  task: text2svg
+  transformer_layer_cls: Starcoder2DecoderLayer # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 5
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 10
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  use_gradient_checkpointing: true
+  train_image_encoder: true
+  train_LLM: true
+fsdp:
+  enable: true # TODO: set this reasonably, i.e., false only if you want to use DDP or have PyTorch < 2.1
+  cpu_offload: false
+  sharding_strategy: hsdp
+  backward_prefetch: BACKWARD_PRE
+  use_orig_params: true
+  sync_module_states: true
+  forward_prefetch: false
+  cpu_ram_efficient_loading: true
+data:
+  num_workers: 16
+  train:
+    batch_size: 4
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: train
+      dataset_name: starvector/FIGR-SVG
+      im_size: 384
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+      image_processor: siglip_384
+  test:
+    batch_size: 4
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: test
+      dataset_name: starvector/FIGR-SVG
+      im_size: 384
+      num_samples: 64
+      transforms: false
+      select_dataset_name: false
+      image_processor: siglip_384
+generation:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+metrics:
+  L2: false
+  Masked-L2: false
+  LPIPS: false
+  SSIM: false
+  FID: false
+  FID_clip: false
+  CLIPScore: true
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: false

configs/models/starvector-8b/text2svg-stack.yaml

@@ -0,0 +1,96 @@
+project:
+  project: starvector-8b-text2svg
+  use_wandb: false
+  entity: joanrod
+  copy_code: false
+model:
+  max_length: 16000
+  model_name: starvector/starvector-8b-im2svg
+  starcoder_model_name: bigcode/starcoder2-7b
+  pretrained: true
+  image_encoder_type: siglip_384
+  use_flash_attn: true
+  adapter_norm: layer_norm
+  init_type: normal
+  dropout: 0.1
+  task: text2svg
+  transformer_layer_cls: Starcoder2DecoderLayer # fsdp specific
+  use_cache: false
+training:
+  save_model_epochs: 1
+  checkpointing_steps: 500
+  checkpoints_total_limit: 5
+  model_precision: bf16
+  resume_from_checkpoint: false
+  continue_training: false
+  n_epochs: 4
+  lr: 0.00001
+  gradient_accumulation_steps: 4
+  lr_scheduler: cosine
+  lr_warmup_steps: 10
+  adam_beta1: 0.95
+  adam_beta2: 0.999
+  adam_weight_decay: 1.0e-06
+  adam_epsilon: 1e-08
+  optimizer: adamw
+  use_gradient_checkpointing: true
+  train_image_encoder: true
+  train_LLM: true
+fsdp:
+  enable: true # TODO: set this reasonably, i.e., false only if you want to use DDP or have PyTorch < 2.1
+  cpu_offload: false
+  sharding_strategy: hsdp
+  backward_prefetch: BACKWARD_PRE
+  use_orig_params: true
+  sync_module_states: true
+  forward_prefetch: false
+  cpu_ram_efficient_loading: true
+data:
+  num_workers: 16
+  train:
+    batch_size: 4
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: train
+      dataset_name: starvector/text2svg-stack
+      im_size: 384
+      num_samples: -1
+      transforms: false
+      select_dataset_name: false
+      image_processor: siglip_384
+  test:
+    batch_size: 4
+    target: starvector.data.stacksvg.SVGStackDataset
+    params:
+      split: test
+      dataset_name: starvector/text2svg-stack
+      im_size: 384
+      num_samples: 64
+      transforms: false
+      select_dataset_name: false
+      image_processor: siglip_384
+generation:
+  max_length: 10000
+  min_length: 10
+  num_beams: 3
+  temperature: 1.0
+  num_captions: 1
+  repetition_penalty: 1.0
+  length_penalty: 0.5
+  top_p: 0.95
+  use_nucleus_sampling: true
+  im_size: 384
+  dpi: 2
+  scale: 300
+metrics:
+  L2: false
+  Masked-L2: false
+  LPIPS: false
+  SSIM: false
+  FID: false
+  FID_clip: false
+  CLIPScore: true
+  CountTokenLength: true
+  ratio_post_processed: false
+  ratio_non_compiling: false
+  DinoScore: false

controller.log

@@ -0,0 +1,31 @@
+2025-03-23 15:00:44 | INFO | controller | args: Namespace(host='0.0.0.0', port=10000, dispatch_method='shortest_queue')
+2025-03-23 15:00:44 | INFO | controller | Init controller
+2025-03-23 15:00:45 | ERROR | stderr | INFO:     Started server process [48368]
+2025-03-23 15:00:45 | ERROR | stderr | INFO:     Waiting for application startup.
+2025-03-23 15:00:45 | ERROR | stderr | INFO:     Application startup complete.
+2025-03-23 15:00:45 | ERROR | stderr | INFO:     Uvicorn running on http://0.0.0.0:10000 (Press CTRL+C to quit)
+2025-03-23 15:01:04 | INFO | controller | Register a new worker: http://localhost:40000
+2025-03-23 15:01:04 | INFO | controller | Register done: http://localhost:40000, {'model_names': ['/home/agent_h/data/starvector-1b-im2svg'], 'speed': 1, 'queue_length': 0}
+2025-03-23 15:01:04 | INFO | stdout | INFO:     127.0.0.1:51486 - "POST /register_worker HTTP/1.1" 200 OK
+2025-03-23 15:01:19 | INFO | controller | Receive heart beat. http://localhost:40000
+2025-03-23 15:01:19 | INFO | stdout | INFO:     127.0.0.1:51523 - "POST /receive_heart_beat HTTP/1.1" 200 OK
+2025-03-23 15:01:34 | INFO | controller | Receive heart beat. http://localhost:40000
+2025-03-23 15:01:34 | INFO | stdout | INFO:     127.0.0.1:51562 - "POST /receive_heart_beat HTTP/1.1" 200 OK
+2025-03-23 15:01:49 | INFO | controller | Receive heart beat. http://localhost:40000
+2025-03-23 15:01:49 | INFO | stdout | INFO:     127.0.0.1:51607 - "POST /receive_heart_beat HTTP/1.1" 200 OK
+2025-03-23 15:02:04 | INFO | controller | Receive heart beat. http://localhost:40000
+2025-03-23 15:02:04 | INFO | stdout | INFO:     127.0.0.1:51648 - "POST /receive_heart_beat HTTP/1.1" 200 OK
+2025-03-23 15:02:19 | INFO | controller | Receive heart beat. http://localhost:40000
+2025-03-23 15:02:19 | INFO | stdout | INFO:     127.0.0.1:51683 - "POST /receive_heart_beat HTTP/1.1" 200 OK
+2025-03-23 15:02:34 | INFO | controller | Receive heart beat. http://localhost:40000
+2025-03-23 15:02:34 | INFO | stdout | INFO:     127.0.0.1:51729 - "POST /receive_heart_beat HTTP/1.1" 200 OK
+2025-03-23 15:02:55 | ERROR | stderr | INFO:     Shutting down
+2025-03-23 15:02:55 | ERROR | stderr | INFO:     Waiting for application shutdown.
+2025-03-23 15:02:55 | ERROR | stderr | INFO:     Application shutdown complete.
+2025-03-23 15:02:55 | ERROR | stderr | INFO:     Finished server process [48368]
+2025-03-23 15:04:32 | INFO | controller | args: Namespace(host='0.0.0.0', port=10000, dispatch_method='shortest_queue')
+2025-03-23 15:04:32 | INFO | controller | Init controller
+2025-03-23 15:04:32 | ERROR | stderr | INFO:     Started server process [50695]
+2025-03-23 15:04:32 | ERROR | stderr | INFO:     Waiting for application startup.
+2025-03-23 15:04:32 | ERROR | stderr | INFO:     Application startup complete.
+2025-03-23 15:04:32 | ERROR | stderr | INFO:     Uvicorn running on http://0.0.0.0:10000 (Press CTRL+C to quit)

model_worker_ad9563.log

@@ -0,0 +1,17 @@
+2025-03-23 15:01:04 | INFO | model_worker | args: Namespace(host='0.0.0.0', port=40000, worker_address='http://localhost:40000', controller_address='http://localhost:10000', model_name='/home/agent_h/data/starvector-1b-im2svg', multi_modal=False, limit_model_concurrency=5, stream_interval=1, no_register=False, openai_api_key='EMPTY', vllm_base_url='http://localhost:8000')
+2025-03-23 15:01:04 | INFO | model_worker | Loading the model /home/agent_h/data/starvector-1b-im2svg on worker ad9563 ...
+2025-03-23 15:01:04 | INFO | model_worker | Register to controller
+2025-03-23 15:01:04 | ERROR | stderr | INFO:     Started server process [48407]
+2025-03-23 15:01:04 | ERROR | stderr | INFO:     Waiting for application startup.
+2025-03-23 15:01:04 | ERROR | stderr | INFO:     Application startup complete.
+2025-03-23 15:01:04 | ERROR | stderr | INFO:     Uvicorn running on http://0.0.0.0:40000 (Press CTRL+C to quit)
+2025-03-23 15:01:19 | INFO | model_worker | Send heart beat. Models: ['/home/agent_h/data/starvector-1b-im2svg']. Semaphore: None. global_counter: 0
+2025-03-23 15:01:34 | INFO | model_worker | Send heart beat. Models: ['/home/agent_h/data/starvector-1b-im2svg']. Semaphore: None. global_counter: 0
+2025-03-23 15:01:49 | INFO | model_worker | Send heart beat. Models: ['/home/agent_h/data/starvector-1b-im2svg']. Semaphore: None. global_counter: 0
+2025-03-23 15:02:04 | INFO | model_worker | Send heart beat. Models: ['/home/agent_h/data/starvector-1b-im2svg']. Semaphore: None. global_counter: 0
+2025-03-23 15:02:19 | INFO | model_worker | Send heart beat. Models: ['/home/agent_h/data/starvector-1b-im2svg']. Semaphore: None. global_counter: 0
+2025-03-23 15:02:34 | INFO | model_worker | Send heart beat. Models: ['/home/agent_h/data/starvector-1b-im2svg']. Semaphore: None. global_counter: 0
+2025-03-23 15:02:45 | ERROR | stderr | INFO:     Shutting down
+2025-03-23 15:02:45 | ERROR | stderr | INFO:     Waiting for application shutdown.
+2025-03-23 15:02:45 | ERROR | stderr | INFO:     Application shutdown complete.
+2025-03-23 15:02:45 | ERROR | stderr | INFO:     Finished server process [48407]

models/modeling_starvector.py

@@ -0,0 +1,224 @@
+"""
+This module defines a self-contained StarVector model with support for remote code loading.
+"""
+
+import os
+import torch
+import torch.nn as nn
+from transformers import PreTrainedModel, PretrainedConfig
+from typing import Optional, Union, List
+from abc import ABC, abstractmethod
+
+# Import components - these will be included in the HF repo
+from .starvector.image_encoder import ImageEncoder
+from .starvector.adapter import Adapter
+
+# === Model Configuration ===
+
+class StarVectorConfig(PretrainedConfig):
+    model_type = "starvector"
+
+    def __init__(
+        self,
+        starcoder_model_name: str = "bigcode/starcoderbase-1b",
+        image_encoder_type: str = "clip",
+        adapter_norm: str = "layer_norm",
+        image_size: int = 224,
+        max_length: int = 8192,
+        max_length_train: int = 8192,
+        use_flash_attn: bool = True,
+        use_cache: bool = True,
+        num_attention_heads: int = 16,
+        num_hidden_layers: int = 24,
+        vocab_size: int = 49152,
+        hidden_size: int = 2048,
+        num_kv_heads: int = 4,
+        torch_dtype: str = "bfloat16",
+        **kwargs,
+    ):
+        # Initialize the parent config first
+        super().__init__(**kwargs)
+        self.starcoder_model_name = starcoder_model_name
+        self.image_encoder_type = image_encoder_type
+        self.adapter_norm = adapter_norm
+        self.image_size = image_size
+        self.max_length = max_length
+        self.max_length_train = max_length_train
+        self.use_flash_attn = use_flash_attn
+        self.use_cache = use_cache
+        self.num_attention_heads = num_attention_heads
+        self.num_hidden_layers = num_hidden_layers
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_kv_heads = num_kv_heads
+        self.torch_dtype = torch_dtype
+
+# === Base Model Classes ===
+
+class StarVectorBase(nn.Module, ABC):
+    def __init__(self, config, **kwargs):
+        super().__init__()
+        self.task = kwargs.get('task', 'im2svg')
+        self.model_precision = kwargs.get('model_precision', config.torch_dtype)
+        
+        # Instantiate the SVG transformer using the abstract method.
+        self.svg_transformer = self._get_svg_transformer(config, **kwargs)
+        
+        if self.use_image_encoder():
+            self.image_encoder = ImageEncoder(config, **kwargs)
+            self.image_projection = self.get_adapter(config, **kwargs).to(dtype=self.model_precision)
+        else:
+            self.query_length = 0
+            
+        self.max_length = config.max_length_train - getattr(self, "query_length", 0) - 4  
+        self.train_image_encoder = kwargs.get('train_image_encoder', False)
+        self.train_LLM = kwargs.get('train_LLM', False)
+        
+        self._freeze_parameters(self.train_image_encoder, self.train_LLM)
+
+    @abstractmethod
+    def _get_svg_transformer(self, config, **kwargs):
+        """Get SVG transformer model - implementation differs between versions"""
+        pass
+
+    def _freeze_parameters(self, train_image_encoder, train_LLM):
+        if self.use_image_encoder():
+            for _, param in self.image_encoder.named_parameters():
+                param.requires_grad = train_image_encoder
+            for _, param in self.image_projection.named_parameters():
+                param.requires_grad = train_image_encoder
+        for _, param in self.svg_transformer.transformer.named_parameters():
+            param.requires_grad = train_LLM
+
+    def use_image_encoder(self):
+        return self.task == 'im2svg'
+
+    def get_adapter(self, config, **kwargs):
+        # Determine hidden size and query length based on the image encoder type.
+        if config.image_encoder_type == 'clip':
+            hidden_size = self.image_encoder.num_features
+            self.query_length = 257
+        elif config.image_encoder_type == 'vqgan':
+            hidden_size = 256
+            self.query_length = 196
+        else:
+            hidden_size = 256  # default fallback
+            self.query_length = 200
+        llm_hidden_size = config.hidden_size  # assuming the transformer hidden size
+        return Adapter(hidden_size, llm_hidden_size, adapter_norm=config.adapter_norm, query_length=self.query_length, dropout_prob=kwargs.get('dropout', 0.1))
+
+    def forward(self, batch):
+        # Simplified forward pass where we assume batch has an "image" key.
+        image = batch["image"]
+        if self.use_image_encoder():
+            embedded_image = self.image_encoder(image)
+            conditioning_embeds = self.image_projection(embedded_image)
+            # For demo purposes, we generate dummy input embeddings (replace with your logic)
+            inputs_embeds = self.svg_transformer.transformer.wte(
+                torch.randint(0, self.svg_transformer.transformer.wte.num_embeddings, (image.size(0), self.max_length))
+            )
+        else:
+            inputs_embeds = self.svg_transformer.transformer.wte(
+                torch.randint(0, self.svg_transformer.transformer.wte.num_embeddings, (image.size(0), self.max_length))
+            )
+        return inputs_embeds  # Dummy return
+
+    def generate_im2svg(self, batch, **kwargs):
+        # Prepare generation inputs (dummy implementation)
+        image = batch["image"]
+        if self.use_image_encoder():
+            embedded_image = self.image_encoder(image)
+            conditioning_embeds = self.image_projection(embedded_image)
+        else:
+            conditioning_embeds = torch.zeros((image.size(0), 10, 1), device=image.device)
+        generation_output = self.svg_transformer.transformer.generate(inputs_embeds=conditioning_embeds, max_length=kwargs.get('max_length', 30))
+        raw_svg = self.svg_transformer.tokenizer.batch_decode(generation_output, skip_special_tokens=True)
+        return raw_svg
+
+    @abstractmethod
+    def _get_embeddings(self, input_ids):
+        """Get embeddings from input ids - implementation differs between v1 and v2"""
+        pass
+
+    @abstractmethod
+    def _get_svg_text(self, svg_list):
+        """Get SVG text with appropriate end tokens - implementation differs between v1 and v2"""
+        pass
+
+# V1 implementation: Delegates transformer creation to the external LLM file.
+class StarVectorStarCoder(StarVectorBase):
+    def __init__(self, config, **kwargs):
+        super().__init__(config, **kwargs)
+    
+    def _get_svg_transformer(self, config, **kwargs):
+        from starvector.model.llm.starcoder import StarCoderModel  # V1: use StarCoderModel from external file
+        return StarCoderModel(config, **kwargs)
+
+    def _get_embeddings(self, input_ids):
+        """V1-specific embedding method"""
+        # This follows the implementation in starvector/model/models/starvector_v1.py.
+        return self.svg_transformer.transformer.transformer.wte(input_ids)
+
+    def _get_svg_text(self, svg_list):
+        """V1-specific SVG text preparation"""
+        return [t + self.svg_transformer.tokenizer.eos_token for t in svg_list]
+
+# V2 implementation: Delegates transformer creation to the external V2 LLM file.
+class StarVectorStarCoder2(StarVectorBase):
+    def __init__(self, config, **kwargs):
+        super().__init__(config, **kwargs)
+    
+    def _get_svg_transformer(self, config, **kwargs):
+        from starvector.model.llm.starcoder2 import StarCoderModel  # V2: use external StarCoderModel from starcoder2.py
+        return StarCoderModel(config, **kwargs)
+
+    def _get_embeddings(self, input_ids):
+        """V2-specific embedding method"""
+        return self.svg_transformer.transformer.model.embed_tokens(input_ids)
+
+    def _get_svg_text(self, svg_list):
+        """V2-specific SVG text preparation"""
+        return [t + self.svg_transformer.svg_end_token + self.svg_transformer.tokenizer.eos_token for t in svg_list]
+
+    def _get_im2svg_specific_kwargs(self, kwargs):
+        """V2-specific generation kwargs"""
+        return {
+            'eos_token_id': self.svg_transformer.svg_end_token_id,
+        }
+
+    def _get_text2svg_specific_kwargs(self, kwargs):
+        """V2-specific text2svg generation kwargs"""
+        return {
+            'eos_token_id': self.svg_transformer.tokenizer.eos_token_id,
+        }
+
+# === Main Model Class for Hugging Face ===
+
+class StarVectorForCausalLM(PreTrainedModel):
+    config_class = StarVectorConfig
+    _no_split_modules = []
+    
+    def __init__(self, config, **kwargs):
+        super().__init__(config)
+        # Choose V2 if the model name indicates starcoder2; otherwise use V1.
+        if "starcoder2" in config.starcoder_model_name.lower():
+            self.model = StarVectorStarCoder2(config=config, **kwargs)
+        else:
+            self.model = StarVectorStarCoder(config=config, **kwargs)
+
+    def forward(self, batch):
+        return self.model(batch)
+
+    def generate_im2svg(self, batch, **kwargs):
+        return self.model.generate_im2svg(batch, **kwargs)
+    
+    def generate_im2text(self, batch, **kwargs):
+        return self.model.generate_im2text(batch, **kwargs)
+
+    def process_images(self, images):
+        return self.model.image_encoder.process_images(images)
+
+# === Registration for Autonomous Loading ===
+
+StarVectorConfig.register_for_auto_class()
+StarVectorForCausalLM.register_for_auto_class("AutoModelForCausalLM") 

models/starvector/adapter.py

@@ -0,0 +1,15 @@
+import torch.nn as nn
+
+class Adapter(nn.Module):
+    def __init__(self, in_features, out_features, adapter_norm="layer_norm", query_length=1, dropout_prob=0.1):
+        super().__init__()
+        self.fc = nn.Linear(in_features, out_features)
+        self.norm = nn.LayerNorm(out_features) if adapter_norm == "layer_norm" else None
+        self.dropout = nn.Dropout(dropout_prob)
+        self.query_length = query_length
+
+    def forward(self, x):
+        out = self.fc(x)
+        if self.norm is not None:
+            out = self.norm(out)
+        return self.dropout(out)