### model
model_name_or_path: deepseek-ai/DeepSeek-V2-Lite
trust_remote_code: true

### method
stage: sft
do_train: true
finetuning_type: lora
lora_rank: 8
lora_target: all

### dataset
dataset: identity, alpaca_en_demo
template: deepseek
cutoff_len: 2048
max_samples: 100000
overwrite_cache: true
preprocessing_num_workers: 16
dataloader_num_workers: 4

### output
output_dir: saves/KT_FT_deepseekV2
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
save_only_model: false
report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
resume_from_checkpoint: null

### ktransformers
use_kt: true
# Pair with fsdp2_kt_bf16.yaml for original BF16 checkpoints.
# For pre-converted expert weights, uncomment kt_weight_path and use fsdp2_kt_int8.yaml or fsdp2_kt_int4.yaml.
# kt_weight_path: /path/to/DeepSeek-V2-Lite-AMXINT8
