Metadata-Version: 2.4
Name: nanorlhf
Version: 1.0.0
Summary: nanoRLHF: from-scratch journey into how LLMs and RLHF really work.
Author-email: hyunwoongko <gusdnd852@naver.com>
License:                                  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.
        
Project-URL: Homepage, https://github.com/hyunwoongko/nanorlhf
Project-URL: Repository, https://github.com/hyunwoongko/nanorlhf
Keywords: rlhf,rl,llm,language model,deep learning,ai
Requires-Python: >=3.9
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: torch
Requires-Dist: triton==3.4.0
Requires-Dist: transformers==4.57.3
Requires-Dist: cloudpickle==3.1.1
Requires-Dist: zstandard==0.25.0
Requires-Dist: wandb==0.17.9
Requires-Dist: math-verify==0.8.0
Dynamic: license-file

# nanoRLHF

<p align="center">
  <a href="https://github.com/karpathy/nanoGPT">
    <img src="assets/title.png" alt="title" title="Homage to Karpathy’s nanoGPT README title image." />
  </a>
</p>

This project aims to perform RLHF training from scratch, implementing almost all core components manually except for PyTorch and Triton. Each module is a minimal, educational reimplementation of large-scale systems focusing on clarity and core concepts rather than production readiness. This includes an SFT and RL training pipeline with evaluation, for training a small Qwen3 model on open-source math datasets.

## Motivation
A few years ago, it still felt possible for an individual to meaningfully train and contribute a model, and I was fortunate to do so with [Polyglot-Ko](https://github.com/EleutherAI/polyglot), the first commercially usable open-source Korean LLM, despite not owning a single GPU, thanks to support from the open-source community. But as the field entered an era where large companies train [massive](https://huggingface.co/deepseek-ai/DeepSeek-R1) [models](https://huggingface.co/Qwen/Qwen3-235B-A22B) at unimaginable scale and release them freely, individual efforts began to feel small by comparison. The same shift happened in libraries. [Open sources](https://github.com/volcengine/verl) [maintained](https://github.com/NVIDIA/Megatron-LM) [by full-time corporate teams](https://github.com/langchain-ai/langchain) quickly outpaced what a single person could sustainably build or maintain. I have always loved developing my own open source, but in this reality, facing clear limits in time and capital, I found myself stepping back from it for a while. Eventually, I reframed the question, not how to compete, but how a single person could still create something genuinely useful. Inspired by projects like [Karpathy’s nano series](https://github.com/karpathy/nanoGPT), I returned to building small, clear, educational implementations, focused not on scale or efficiency, but on understanding and teaching. Even without massive resources, I still believe that individuals can create meaningful work that influences and helps others.

## Requirements and Limitations
I worked on this project using a single server with 8 * H200 GPUs.
It should also run well on A100 80GB GPUs, but to fully experiment with all features including 3D parallelism, a server with at least 8 GPUs is required.
Note that this repository only supports single node multi-GPU training, so multi-node distributed training is not supported.

## Installation
In this project, internal APIs from libraries such as Hugging Face Transformers are used in a hacky way, so all dependency versions except PyTorch are strictly pinned.
It is strongly recommended to run the code in an isolated environment such as a Conda virtual environment.

```bash
# 1) create conda environment
conda create -n nanorlhf python=3.10
conda activate nanorlhf

# 2) install PyTorch for your environment from https://pytorch.org/get-started/locally/
# e.g. pip install torch --index-url https://download.pytorch.org/whl/cu126

# 3) install nanoRLHF
git clone https://github.com/hyunwoongko/nanoRLHF
cd nanoRLHF
pip install -e .
```

## Modules
I recommend approaching this repository as a learn-by-building: first install the library, then study the modules, running each module’s example as you go. Once you’ve worked through all modules, finish by running the full RLHF training pipeline end-to-end.

### 1) `nanosets`: Arrow-like Zero-copy Dataset Library
- Implementation: [nanosets](https://github.com/hyunwoongko/nanoRLHF/tree/main/nanorlhf/nanosets)
- Example: [available](https://github.com/hyunwoongko/nanoRLHF/tree/main/examples/nanosets.py)
- References: [arrow](https://github.com/apache/arrow), [datasets](https://github.com/huggingface/datasets)

`nanosets` is a minimal Arrow-like columnar dataset library focusing on zero-copy and clarity.
The goal is to understand how columnar data formats work without copying and how they are implemented from scratch.

### 2) `nanoray`: Distributed Computing Engine
- Implementation: [nanoray](https://github.com/hyunwoongko/nanoRLHF/tree/main/nanorlhf/nanoray)
- Example: [available](https://github.com/hyunwoongko/nanoRLHF/tree/main/examples/nanoray.py)
- References: [ray](https://github.com/ray-project/ray)

`nanoray` is a minimal distributed computing engine inspired by Ray. 
The goal is to understand how distributed execution works and how to build a simple distributed computing framework from scratch.

### 3) `nanotron`: 3D Parallelism Engine
- Implementation: [nanotron](https://github.com/hyunwoongko/nanoRLHF/tree/main/nanorlhf/nanotron)
- Example: [available](https://github.com/hyunwoongko/nanoRLHF/tree/main/examples/nanotron.py)
- References: [Megatron-LM](https://github.com/NVIDIA/Megatron-LM), [oslo](https://github.com/EleutherAI/oslo)

`nanotron` implements 3D parallelism (data, tensor, pipeline) from scratch, focusing on clarity over efficiency. 
The goal is to understand why and how large models are parallelized across many GPUs.

### 4) `kernels`: Set of Triton Kernels
- Implementation: [kernels](https://github.com/hyunwoongko/nanoRLHF/tree/main/nanorlhf/kernels)
- Example: [available](https://github.com/hyunwoongko/nanoRLHF/tree/main/examples/kernels.py)
- References: [flash-attention](https://github.com/Dao-AILab/flash-attention/), [trident](https://github.com/kakaobrain/trident)

`kernels` is a set of custom Triton kernels for LLM training and inference.
The goal is to understand what optimizations matter in practice and how they are implemented at the kernel level.

### 5) `nanovllm`: High Performance Inference Engine
- Implementation: [nanovllm](https://github.com/hyunwoongko/nanoRLHF/tree/main/nanorlhf/nanovllm)
- Example: [available](https://github.com/hyunwoongko/nanoRLHF/tree/main/examples/nanovllm.py)
- References: [vllm](https://github.com/vllm-project/vllm), [nano-vllm](https://github.com/GeeeekExplorer/nano-vllm)

`nanovllm` is a minimal high-performance inference engine for LLMs, built to understand what makes inference fast in practice, including paged attention and efficient batching. 

Thanks to the excellent [nano-vllm](https://github.com/GeeeekExplorer/nano-vllm) project, this module began as a study-driven reimplementation and was then adapted to fit our stack and parallelism design.

### 6) `nanoverl`: RLHF Training Framework
- Implementation: [nanoverl](https://github.com/hyunwoongko/nanoRLHF/tree/main/nanorlhf/nanoverl)
- Example: [available](https://github.com/hyunwoongko/nanoRLHF/tree/main/scripts)
- References: [verl](https://github.com/volcengine/verl), [OpenRLHF](https://github.com/OpenRLHF/OpenRLHF)

`nanoverl` is a minimal RLHF training framework focusing on clarity.
The goal is to understand how RLHF training pipelines are structured, including SFT, RL algorithms and asynchronous RL.

### 7) Full RLHF Training Pipeline
This section is the final step after you have studied all the modules above.

#### Prepare Supervised Fine-tuning Dataset
In the examples included in this project, supervised fine-tuning is performed using [NuminaMath-CoT-Small-Hard-200k](https://huggingface.co/datasets/NotASI/NuminaMath-CoT-Small-Hard-200k).
From the original dataset, 180k samples are used as training data, and 1k samples are used as validation data.
Running the following command will tokenize the dataset and save it as zero-copy `.nano` format (similar with `.arrow` format).

```bash
bash ./scripts/prepare_sft_data.sh
```

#### Supervised Fine-tuning
Supervised fine-tuning is performed using [Qwen3-0.6B-base](https://huggingface.co/Qwen/Qwen3-0.6B-base) model with 3D parallelism by default config.
If you want to modify hyperparameters, please edit `configs/train_sft.yaml` file.
Running the following command will start supervised fine-tuning. Moreover, you can monitor the training process if you have wandb account.

![sft_log](assets/sft_log.png)

```bash
bash ./scripts/train_sft.sh
```

#### Merge Parallelized Checkpoints
After supervised fine-tuning is completed, the parallelized checkpoints are saved in the directory you specified (default is `./checkpoints`).
To use the model for inference or further training, you need to merge the parallelized checkpoints into a single model checkpoint.
The following script will merge the checkpoints and save them in `$YOUR_CHECKPOINT_PATH/merged` directory.

```bash
bash ./scripts/merge_sft_model.sh $STEP
```

#### Evaluate Supervised Fine-tuned Model
After merging the supervised fine-tuned model, you can evaluate it using the following script.
The evaluation is performed using [MATH-500](https://huggingface.co/datasets/HuggingFaceH4/MATH-500) dataset (500 samples from MATH dataset), and [Math-Verify](https://github.com/huggingface/Math-Verify) is used to parse and verify the model's generated output.

| step  | MATH-500 score |
|------:|----------------:|
| 500   | 40.8            |
| 1000  | 39.4            |
| 1500  | 41.2            |
| 2000  | 43.4            |
| 2109  | 41.8            |

```bash
bash ./scripts/eval_sft_model.sh $STEP
```

#### Prepare Reinforcement Learning Dataset
Reinforcement learning is performed using [DeepMath-103K](https://huggingface.co/datasets/zwhe99/DeepMath-103K) dataset.
I removed samples that have one of 'yes', 'no', 'true' or 'false' as the answer, so about 84k samples are used for training.
And [MATH-500](https://huggingface.co/datasets/HuggingFaceH4/MATH-500) dataset is used for validation.
Running the following command will tokenize the dataset and save it as zero-copy `.nano` format.

```bash
bash ./scripts/prepare_rl_data.sh
```

#### Reinforcement Learning
Reinforcement learning is performed using PPO algorithm with the SFT model at 2000 steps as the initial policy.
To improve training efficiency, [One-step off-policy asynchronous RL](https://github.com/volcengine/verl/tree/main/recipe/one_step_off_policy) is applied.
If you want to modify hyperparameters, please edit `configs/train_rl.yaml` file.
Running the following command will start reinforcement learning. Moreover, you can monitor the training process if you have wandb account.

![one_step](assets/one_step.png)

![rl_log](assets/rl_log.png)

```bash
bash ./scripts/train_rl.sh
```

#### Merge Parallelized Checkpoints
After reinforcement learning is completed, the parallelized checkpoints are saved in the directory you specified (default is `./checkpoints`).
To use the model for inference or further training, you need to merge the parallelized checkpoints into a single model checkpoint.
The following script will merge the checkpoints and save them in `$YOUR_CHECKPOINT_PATH/merged` directory.

```bash
bash ./scripts/merge_rl_model.sh $STEP
```

#### Evaluate Reinforcement Learning Model
After merging the reinforcement learning model, you can evaluate it using the following script.
The evaluation is performed same as the supervised fine-tuned model using [MATH-500](https://huggingface.co/datasets/HuggingFaceH4/MATH-500) dataset.
Qwen3-0.6B (non-thinking) model is also evaluated as a reference.

| step                      | MATH-500 score |
|---------------------------|----------------:|
| 50                        | 40.8            |
| 100                       | 43.8            |
| 150                       | 46.6            |
| 200                       | 43.8            |
| Qwen3-0.6B (non-thinking) | 49.8            |

```bash
bash ./scripts/eval_rl_model.sh $STEP
```

```bash
bash ./scripts/eval_ref_model.sh
```

## License
This project is licensed under the Apache 2.0 License.

```text
Copyright 2025 Hyunwoong Ko

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.
```
