TruthRL/train_grpo.sh at main · facebookresearch/TruthRL · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
set -x

export RAY_DEDUP_LOGS=0
export OPENAI_API_BASE=http://localhost:8000/v1
export OPENAI_API_KEY="token-abc123"

export WANDB_PROJECT="TruthRL"

DATA_DIR=<path_to_data_dir> # refer to HF data repo: weizhepei/TruthRL-CRAG

N_GPUS=8
ROLLOUT_TP_SIZE=2

MODEL_NAME=meta-llama/Llama-3.1-8B-Instruct
LR=1e-6
KL_LOSS_COEF=0.001
BSZ=64

python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=$DATA_DIR/train.parquet \
    data.val_files=$DATA_DIR/test.parquet \
    data.train_batch_size=$BSZ \
    data.max_prompt_length=16384 \
    data.max_response_length=2048 \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    actor_rollout_ref.model.path=$MODEL_NAME \
    actor_rollout_ref.actor.optim.lr=$LR \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=$BSZ \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=$KL_LOSS_COEF \
    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
    actor_rollout_ref.actor.entropy_coeff=0 \
    actor_rollout_ref.model.enable_gradient_checkpointing=True \
    actor_rollout_ref.actor.fsdp_config.param_offload=False \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=$ROLLOUT_TP_SIZE \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.8 \
    actor_rollout_ref.rollout.n=8 \
    actor_rollout_ref.rollout.max_num_batched_tokens=131072 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.logger='["console","wandb"]' \
    trainer.project_name=$WANDB_PROJECT \
    trainer.experiment_name='TruthRL-'$MODEL_NAME'_bsz_'$BSZ'_lr_'$LR'_kl_loss_coef_'$KL_LOSS_COEF'' \
    trainer.n_gpus_per_node=$N_GPUS \
    trainer.nnodes=1 \
    trainer.save_freq=10 \
    trainer.test_freq=5 \
    trainer.resume_mode=auto \
    trainer.total_epochs=1 $@