[ English | 中文 ]

RWKV-PEFT is the official implementation for efficient parameter fine-tuning of RWKV5/6 models, supporting various advanced fine-tuning methods across multiple hardware platforms.

git clone https://github.com/JL-er/RWKV-PEFT.git
cd RWKV-PEFT
pip install -r requirements.txt

streamlit run web/app.py

• Hardware Requirements
• Quick Start
• Main Features
• Detailed Configuration
• GPU Support
• Citation

The following shows memory usage when using an RTX 4090 (24GB VRAM) + 64GB RAM (with parameters: --strategy deepspeed_stage_1 --ctx_len 1024 --micro_bsz 1 --lora_r 64):

Note:

• OOM when batch size is 8 ** Requires 19.5GB VRAM when batch size is 8

1. Install dependencies:

pip install -r requirements.txt

2. Run example script:

sh scripts/run_lora.sh

Note: Please refer to the RWKV official tutorial for detailed data preparation

3. Start with web GUI:

streamlit run web/app.py

• Multiple Fine-tuning Methods: Supports LoRA, PISSA, Bone, State Tuning, etc.
• Quantized Training: Supports INT8/NF4 quantization for significant VRAM reduction
• Flexible Data Loading: Supports various data sampling strategies
• Memory Optimization: Multiple DeepSpeed strategies available
• Loss Masking: Supports loss masking for QA dialogue and padding
• Infinite Context Training: Supports infctx training mode, utilizing RWKV's constant memory usage advantage to train with "infinite" context under limited resources
• Multi-Hardware Support: RWKV-PEFT officially supports NVIDIA, AMD, Moore Threads, Musa, Iluvatar CoreX, and other hardware platforms. Ascend NPU implementation will be available later. Note: Currently we only support issues for NVIDIA hardware
• RWKV-FLA Efficient Training: rwkv-fla is a Triton-based linear attention operator that can run efficiently on hardware without CUDA support

--peft bone --bone_config $lora_config

--train_parts ["time", "ln"]

• Available parts: emb, head, time, ln
• Default training: time, ln (small parameter ratio)

--quant int8/nf4

--train_type infctx --chunk_ctx 512 --ctx_len 2048

• ctx_len: Target training length
• chunk_ctx: Slice length, must be smaller than ctx_len

--dataload pad

• get: Default random sampling (RWKV-LM style)
• pad: Fixed-length padding sampling
• only: Single data sampling (only supports bsz=1)

--strategy deepspeed_stage_1

Available strategies:

• deepspeed_stage_1: Preferred option
• deepspeed_stage_2/3: For large models or full fine-tuning
• deepspeed_stage_2_offload
• deepspeed_stage_3_offload

By default, RWKV-PEFT uses custom CUDA kernels for wkv computation. However, you can use --fla to enable the Triton kernel:

--fla

• NVIDIA: CUDA
• Intel, Moore Threads, Musa, Iluvatar CoreX: FLA, which means you need to pass --fla
• Ascend: CANN (soon)

If you find this project helpful, please cite our work:

@misc{kang2024boneblockaffinetransformation,
      title={Bone: Block Affine Transformation as Parameter Efficient Fine-tuning Methods for Large Language Models},
      author={Jiale Kang},
      year={2024},
      eprint={2409.15371},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.15371}
}