Reasoning-SFT

This repository is a customized version of NVIDIA Megatron-LM, extended to support Supervised Fine-Tuning (SFT) of reasoning models. Reasoning-SFT applies prompt masking to train exclusively on the response. It was used to train the hybrid reasoning model Thinkless-1.5B-Warmup. This code is also suitable for standard SFT.

📄 Paper Link	ArXiv
💻 Thinkless GitHub	VainF/Thinkless
🤖 RL Model	Thinkless-1.5B-RL-DeepScaleR
🐣 Warmup Model	Thinkless-1.5B-Warmup
📊 Data for Warmup	Hybrid-OpenThoughts2-1M-1.5B
📊 Data for RL	agentica-org/DeepScaleR-Preview-Dataset

Setup

We recommend using a Docker container to run this code, as installing Transformer Engine and Megatron-LM might be a bit complex.

# In your user account
cd Megatron-SFT
pip install -r requirements.txt # install the transformers in your user account
docker run --gpus all --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 -v "$PWD":"$PWD" -v $HOME:$HOME -w "$PWD" -it --rm nvcr.io/nvidia/pytorch:24.12-py3 

Running the above command will mount both the current directory and your home directory into the Docker container. To mount additional directories, simply add -v /path/to/dir:/path/to/dir to the command.

Once inside the Docker container, install all necessary packages:

# In the docker
pip install -r requirements.txt

Example: Hybrid Reasoning via SFT (DeepSeek-R1-Distill-Qwen-1.5B)

In this example, we show how to fine-tune the deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B to enable hybrid reasoning (Warm-up).

Important

Since Docker creates files with root permissions, we download and preprocess the models and data using your user account. This ensures you can easily modify the files later using your preferred editor, such as VSCode.

0. GPU Resource

The default config for 1.5B LLM with Tensor Parallel=1 and Pipeline Parallel=1 (TP1PP1) requires ~70 GB of memory per GPU. If you're using GPUs with less memory, consider using TP2PP1 or TP1PP2 to distribute the model parameters across multiple GPUs. You can also decrease the sequence length from 16384 to 8192.

1. Prepare the LLM

# In your user account 
python scripts/checkpoints/download_hf_models.py --model-card deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

The huggingface model will be saved in assets/checkpoints.

assets
├── cache
└── checkpoints
    └── deepseek_ai_DeepSeek_R1_Distill_Qwen_1.5B
        ├── config.json
        ├── generation_config.json
        ├── model.safetensors
        ├── special_tokens_map.json
        ├── tokenizer_config.json
        └── tokenizer.json

Then, we modify the tokenizer files to replace the <|quad_start|> token with a control token <short>.

#assets/checkpoints/deepseek_ai_DeepSeek_R1_Distill_Qwen_1.5B/tokenizer_config.json
"151650": {
      "content": "<|short|>", # originally "<|quad_start|>"
      "lstrip": false,
      "normalized": false,
      "rstrip": false,
      "single_word": false,
      "special": false # originally true
    },

#assets/checkpoints/deepseek_ai_DeepSeek_R1_Distill_Qwen_1.5B/tokenizer.json
    {
      "id": 151650,
      "content": "<short>", # originally "<|quad_start|>"
      "single_word": false,
      "lstrip": false,
      "rstrip": false,
      "normalized": false,
      "special": false # originally true
    },

Remove the final <think> in the chat template, and remove the split (content = content.split('</think>')[-1]).

#assets/checkpoints/deepseek_ai_DeepSeek_R1_Distill_Qwen_1.5B/tokenizer_config.json
  "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='') %}{%- for message in messages %}{%- if message['role'] == 'system' %}{% set ns.system_prompt = message['content'] %}{%- endif %}{%- endfor %}{{bos_token}}{{ns.system_prompt}}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is none %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls']%}{%- if not ns.is_first %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- set ns.is_first = true -%}{%- else %}{{'\\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- endfor %}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is not none %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'\\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜>\\n'}}{% endif %}",

2. Convert to Megatron Format

Convert the HF model to Megatron format:

# In the docker
bash scripts/checkpoints/convert_deepseek_r1_to_megatron.sh 1.5B 1 1 

We have three parameters here:

• The model size: 1.5B, 32B etc.
• Tensor Parallel and Pipeline Parallel: 1 1 means no tensor parallel and no pipeline parallel. You can try 2 1 or 1 2 to use tensor parallel or pipeline parallel, respectively.

The above command will create a megatron checkpoint like this:

assets
├── cache
└── checkpoints
    ├── deepseek_ai_DeepSeek_R1_Distill_Qwen_1.5B
    └── deepseek_ai_DeepSeek_R1_Distill_Qwen_1.5B_Megatron_TP1PP1
        ├── iter_0000001
        │   └── mp_rank_00
        │       └── model_optim_rng.pt
        └── latest_checkpointed_iteration.txt

3. Prepare the Hybrid Reasoning Dataset

Download the hybrid reasoning dataset from Huggingface and save it as a JSON file. We assume that the dataset already contains instruction and output fields. For other datasets, you may customize the scripts/data/download_hf_dataset.py.

# In your user account
bash scripts/data/download_hf_dataset.py --dataset-card Vinnnf/Hybrid-OpenThoughts2-1M-1.5B

assets
├── cache
├── checkpoints
└── data
    └── Vinnnf-Hybrid-OpenThoughts2-1M-1.5B
        └── Vinnnf-Hybrid-OpenThoughts2-1M-1.5B.json

Pre-tokenize the dataset:

# In the docker
bash scripts/data/tokenize_dataset.sh assets/checkpoints/deepseek_ai_DeepSeek_R1_Distill_Qwen_1.5B assets/data/Vinnnf-Hybrid-OpenThoughts2-1M-1.5B/Vinnnf-Hybrid-OpenThoughts2-1M-1.5B.json 16384

assets/
├── cache
├── checkpoints
└── data
    └── Vinnnf-Hybrid-OpenThoughts2-1M-1.5B
        ├── Tokenized-Vinnnf-Hybrid-OpenThoughts2-1M-1.5B-deepseek_ai_DeepSeek_R1_Distill_Qwen_1.5B-16384_text_document.bin
        ├── Tokenized-Vinnnf-Hybrid-OpenThoughts2-1M-1.5B-deepseek_ai_DeepSeek_R1_Distill_Qwen_1.5B-16384_text_document.idx
        └── Vinnnf-Hybrid-OpenThoughts2-1M-1.5B.json

The parameters are:

• The path to the tokenizer model (usually the HF model path)
• The path to the dataset JSON file
• The maximum sequence length for training, a value larger than 16384 is recommended.

4. Training

Run the fine-tuning script:

# In the Docker
bash scripts/sft/SFT_Hybrid_R1_1.5B_OpenThoughts_1M.sh train

Auto Resume:

# In the Docker
bash scripts/sft/SFT_Hybrid_R1_1.5B_OpenThoughts_1M.sh resume

5. Export to Huggingface Format

# In the Docker
bash scripts/checkpoints/merge_and_export.sh PATH_TO_YOUR_CKPT assets/checkpoints/deepseek_ai_DeepSeek_R1_Distill_Qwen_1.5B assets/checkpoints/export/Hybrid_R1_1.5B

6. Training Loss for Reference

Acknowledgement

This implementation is also heavily based on alibaba/Pai-Megatron-Patch.