curl -LsSf https://astral.sh/uv/install.sh | sh
cd text-to-lora
uv self update
uv venv --python 3.10 --seed
uv sync
uv pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.6.3/flash_attn-2.6.3+cu123torch2.3cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
uv pip install src/fishfarm
source .venv/bin/activate
pip install --upgrade transformers>=4.53.3
deactivate
tmux new -s lora_training
cd text-to-lora
source .venv/bin/activateaccelerate launch \
--num_processes=4 \
--gpu_ids=0,1,2,3 \
--num_machines=1 \
--mixed_precision=no \
scripts/train_custom_sft.py \
configs/hyper_lora_decontam_lol_tasks.yaml \
--model_dir=Qwen/Qwen3-32B \
--emb_model=Alibaba-NLP/gte-large-en-v1.5 \
--warmup_frac=0.2 \
--lr=2.5e-5 \
--n_tasks_per_batch=8 \
--n_points_per_task=1 \
--grad_accum_steps=1 \
--epochs=100 \
--n_descs_per_ds=128 \
--n_train_ds=479 \
--exp_setup=hyper_lora \
--encoder_type=linear \
--l2_reg_generated_w=1e-3 \
--label_smoothing=0.1 \
--neftune_noise_alpha=5 \
--weight_decay=1e-2 \
--hypernet_latent_size=512 \
--head_in_size=4096 \
--val_batch_size=16
Ctrl+B, D でデタッチ
tmux attach -t lora_training
tmux,よくわからんね〜〜〜〜