search

Blealtan / RWKV-LM-LoRA

RWKV is a RNN with transformer-level LLM performance. It can be directly trained like a GPT (parallelizable). So it's combining the best of RNN and transformer - great performance, fast inference, saves VRAM, fast training, "infinite" ctx_len, and free sentence embedding.
Apache License 2.0
407 stars 41 forks source link

readme

LoRA fork of RWKV-LM

A RWKV-LM fork, added with LoRA finetuning support. Currently only RWKV-v4neo is supported. The LoRA module is self-implemented to work with the TorchScript JIT. Existing RWKV-v4neo models/checkpoints should work out of the box. Now only LoRA-finetuned weights are checkpointed during training: it provides much smaller checkpoints, but you now need to specify the base model to use it. See args.MODEL_LOAD and args.MODEL_LORA in RWKV-v4neo/chat.py.

To finetune an existing model with LoRA, just work like full finetuning but with the LoRA options, in the directory RWKV-v4neo:

python3 train.py \
  --load_model <pretrained base model> \
  --proj_dir <place to save checkpoints> \
  --data_file <data for finetune> \
  --data_type <data type for finetune, recommend binidx> \
  --vocab_size 50277 --ctx_len 1024 --epoch_steps 1000 --epoch_count 1000 --epoch_begin 0 --epoch_save 5 --micro_bsz 2 --accumulate_grad_batches 4 \
  --n_layer 24 --n_embd 1024 --pre_ffn 0 --head_qk 0 --lr_init 1e-4 --lr_final 1e-4 --warmup_steps 0 --beta1 0.9 --beta2 0.999 --adam_eps 1e-8 --accelerator gpu --devices 1 --precision bf16 --strategy deepspeed_stage_2 --grad_cp 0 \ # all your familiar options
  --lora --lora_r 8 --lora_alpha 16 --lora_dropout 0.01 \
  --lora_load <lora checkpoint to continue training> \ # optional
  --lora_parts=att,ffn,time,ln # configure which parts to finetune

The r, alpha and dropout options are up to your choice. The att, ffn, time and ln refers to the TimeMix, ChannelMix, time decay/first/mix parameters, and layernorm parameters; DON'T FORGET to add the set of parameters to be finetuned here. I'm still experimenting with different configurations; your experience is also welcomed!

Use json2binidx to convert your data into binidx, which is best suited for this trainer implementation. Once you have the pair of files path/to/foo.bin and path/to/foo.idx, pass --data_file path/to/foo --data_type binidx as arguments. Notice that the .bin and .idx suffix is not there.

To use the finetuned model, use chat.py as usual with the checkpoints in your specified proj_dir, but remember to align the LoRA-corresponded options with what you have specified during training!

args.MODEL_LORA = 'your_lora_checkpoint.pth'
args.lora_r = 8
args.lora_alpha = 32