This repo contains the code for the paper "Towards Economical Inference: Enabling DeepSeek's Multi-Head Latent Attention in Any Transformer-based LLMs".

• [2025.06.13] Release the refactored code and add support for the Qwen model.
• [2025.03.12] Released the inference code implemented using PyTorch (support for FlashMLA inference requires additional development time).
• [2025.03.04] The four MLA checkpoints ($d_{kv}$=8/16/32/128) derived from SmolLM-135M/360M/1B7 are publicly available.
• [2025.03.03] The four MLA checkpoints ($d_{kv}$=16/32/64/256) derived from Llama-2-7B are publicly available.
• [2025.02.21] The paper of MHA2MLA is publicly available: https://arxiv.org/abs/2502.14837
• [2025.02.19] Released the first version of the MHA2MLA code, providing usage code for Llama fine-tuning and evaluating.

• Provide the code for incorporating the projection matrix and inference.
• Thanks to DeepSeek for open-sourcing the FlashMLA inference framework. It’s theoretically possible to save more GPU memory usage using this framework. Let’s see how economical MHA2MLA + FlashMLA (+ KV quanto) can be!
• Release the code of MHA2MLA based on HuggingFace Transformers

First download the datasets.

• smollm-corpus(fineweb-edu-dedup, cosmopedia-v2, python-edu): https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus
• open-web-math: https://huggingface.co/datasets/open-web-math/open-web-math
• stackoverflow: https://huggingface.co/datasets/bigcode/stackoverflow-clean

Secondly, process the datasets according to https://github.com/huggingface/nanotron/blob/main/docs/nanoset.md.

Install pytorch and other packages.

conda create -n mha2mla python=3.11
pip install torch==2.4.0 torchvision==0.19.0
pip install -r requirements.txt

First, prepare configuration files referencing 135M_4GPU.yaml.

For information on the configuration of mha2mla, you can refer to the arguments.py file.

Then, use the following command for fine-tuning:

torchrun --nproc_per_node 4 \
    ./src/mha2mla/run_train.py \
    --cfg_file ./cfgs/SmolLM1-135M-4GPU.yml

If you want to use the partial-RoPE version 2-norm, you should get the qk_tensor first. Using the following command, you can get the qk_tensor:

torchrun --nproc_per_node 1 \
    ./src/mha2mla/2_norm.py \
    --config_file ./cfgs/SmolLM1-135M-8GPU.yaml \

For evaluation, you can use the following command:

accelerate launch --multi_gpu --num_processes=4 \
    ./eval/eval.py \
    accelerate \
    --model_args "pretrained=${model_name_or_path},revision=main,dtype=bfloat16,max_length=2048" \
    --override_batch_size 48 \
    --custom_tasks "./eval/tasks.py" \
    --tasks "./eval/smollm1_base.txt" \
    --output_dir "./eval_results/"

For the baseline evaluation, you can use the following command:

torchrun --nproc_per_node=4 \
    ./eval/longbench.py \
    --model_path ${model_name_or_path} \
    --tokenizer_path ${model_name_or_path} \
    --longbench True \
    --lb_max_tokens 2048 \
    --lb_batch_size 16 \
    --output_dir ./longbench/bf16 \
    --dtype "bfloat16"

If you want to use the quantized KV cache, you can use the following command:

torchrun --nproc_per_node=4 \
    ./eval/longbench.py \
    --model_path ${model_name_or_path} \
    --tokenizer_path ${model_name_or_path} \
    --longbench True \
    --lb_max_tokens 2048 \
    --lb_batch_size 16 \
    --output_dir ./longbench/${model_name_or_path}_hqq_int4 \
    --dtype "bfloat16" \
    --cache_implementation "quantized" \
    --backend "HQQ" \
    --nbits 4 \
    --residual_length 128

@misc{ji2025economicalinferenceenablingdeepseeks,
      title={Towards Economical Inference: Enabling DeepSeek's Multi-Head Latent Attention in Any Transformer-based LLMs}, 
      author={Tao Ji and Bin Guo and Yuanbin Wu and Qipeng Guo and Lixing Shen and Zhan Chen and Xipeng Qiu and Qi Zhang and Tao Gui},
      year={2025},
      eprint={2502.14837},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.14837}, 
}